Data Mining As A Financial Auditing Tool M.Sc. Thesis in Accounting Swedish School of Economics and Business Administration 2002 The Swedish School of Economics and Business Administration Department: Accounting Type of Document: Thesis Title: Data Mining As A Financial Auditing Tool Author: Supatcharee Sirikulvadhana Abstract In recent years, the volume and complexity of accounting transactions in major organizations have increased dramatically. To audit such organizations, auditors frequently must deal with voluminous data with rather complicated data structure. Consequently, auditors no longer can rely only on reporting or summarizing tools in the audit process. Rather, additional tools such as data mining techniques that can automatically extract information from a large amount of data might be very useful. Although adopting data mining techniques in the audit processes is a relatively new field, data mining has been shown to be cost effective in many business applications related to auditing such as fraud detection, forensics accounting and security evaluation. The objective of this thesis is to determine if data mining tools can directly improve audit performance. The selected test area was the sample selection step of the test of control process. The research data was based on accounting transactions provided by AVH PricewaterhouseCoopers Oy. Various samples were extracted from the test data set using data mining software and generalized audit software and the results evaluated. IBM’s DB2 Intelligent Miner for Data Version 6 was selected to represent the data mining software and ACL for Windows Workbook Version 5 was chosen for generalized audit software. Based on the results of the test and the opinions solicited from experienced auditors, the conclusion is that, within the scope of this research, the results of data mining software are more interesting than the results of generalized audit software. However, there is no evidence that the data mining technique brings out material matters or present significant enhancement over the generalized audit software. Further study in a different audit area or with a more complete data set might yield a different conclusion. Search Words: Data Mining, Artificial Intelligent, Auditing, Computerized Audit Assisted Tools, Generalized Audit Software Table of Contents 1. Introduction 1 1.1. Background 1 1.2. Research Objective 2 1.3. Thesis Structure 2 2. Auditing 4 2.1. Objective and Structure 4 2.2. What Is Auditing? 4 2.3. Audit Engagement Processes 5 2.3.1. Client Acceptance or Client Continuance 5 2.3.2. Planning 6 2.3.2.1. Team Mobilization 6 2.3.2.2. Client’s Information Gathering 7 2.3.2.3. Risk Assessment 7 2.3.2.4. Audit Program Preparation 9 2.3.3. Execution and Documentation 10 2.3.4. Completion 11 2.4. Audit Approaches 12 2.4.1. Tests of Controls 12 2.4.2. Substantive Tests 13 2.4.2.1. Analytical Procedures 13 2.4.2.2. Detailed Tests of Transactions 13 2.4.2.3. Detailed Tests of Balances 14 2.5. Summary 3. Computer Assisted Auditing Tools 14 17 3.1. Objective and Structure 17 3.2. Why Computer Assisted Auditing Tools? 17 3.3. Generalized Audit Software 18 3.4. Other Computerized Tools and Techniques 22 3.5. Summary 23 4. Data mining 24 4.1. Objective and Structure 24 4.2. What Is Data Mining? 24 4.3. Data Mining process 25 4.3.1. Business Understanding 26 4.3.2. Data Understanding 27 4.3.3. Data Preparation 27 4.3.4. Modeling 27 4.3.5. Evaluation 28 4.3.6. Deployment 28 4.4. Data Mining Tools and Techniques 29 4.4.1. Database Algorithms 29 4.4.2. Statistical Algorithms 30 4.4.3. Artificial Intelligence 30 4.4.4. Visualization 30 4.5. Methods of Data Mining Algorithms 32 4.5.1. Data Description 32 4.5.2. Dependency Analysis 33 4.5.3. Classification and Prediction 33 4.5.4. Cluster Analysis 34 4.5.5. Outlier Analysis 34 4.5.6. Evolution Analysis 35 4.6. Examples of Data Mining Algorithms 36 4.6.1. Apriori Algorithms 36 4.6.2. Decision Trees 37 4.6.3. Neural Networks 39 4.7. Summary 5. Integration of Data Mining and Auditing 40 43 5.1. Objective and Structure 43 5.2. Why Integrate Data Mining with Auditing? 43 5.3. Comparison between Currently Used Generalized Auditing Software and Data Mining Packages 44 5.3.1. Characteristics of Generalized Audit Software 45 5.3.2. Characteristics of Data Mining Packages 46 5.4. Possible Areas of Integration 48 5.5. Examples of Tests 58 5.6. Summary 66 6. Research Methodology 68 6.1. Objective and Structure 68 6.2. Research Period 68 6.3. Data Available 68 6.4. Research Methods 69 6.5. Software Selection 70 6.5.1. Data Mining Software 70 6.5.2. Generalized Audit Software 71 6.6. Analysis Methods 71 6.7. Summary 72 7. The Research 73 7.1. Objective and Structure 73 7.2. Hypothesis 73 7.3. Research Processes 73 7.3.1. Business Understanding 73 7.3.2. Data Understanding 74 7.3.3. Data Preparation 75 7.3.3.1. Data Transformation 75 7.3.3.2. Attribute Selection 76 7.3.3.3. Choice of Tests 80 7.3.4. Software Deployment 82 7.3.4.1. IBM’s DB2 Intelligent Miner for Data 82 7.3.4.2. ACL 91 7.4. Result Interpretations 94 7.4.1. IBM’s DB2 Intelligent Miner for Data 94 7.4.2. ACL 95 7.5. Summary 99 8. Conclusion 101 8.1. Objective and Structure 101 8.2. Research Perspective 101 8.3. Implications of the Results 102 8.4. Restrictions and Constraints 103 8.4.1. Data Limitation 103 8.4.1.1. Incomplete Data 103 8.4.1.2. Missing Information 103 8.4.1.3. Limited Understanding 104 8.4.2. Limited Knowledge of Software Packages 104 8.4.3. Time Constraint 105 8.5. Suggestions for Further Researches 105 8.6. Summary 105 List of Figures 105 List of Tables 105 References 105 a) Books and Journals 105 b) Web Pages 105 Appendix A: List of Columns of Data Available 109 Appendix B Results of IBM’s Intelligent Miner for Data 105 a) Preliminary Neural Clustering (with Six Attributes) 105 b) Demographic Clustering: First Run 105 c) Demographic Clustering: Second Run 105 d) Neural Clustering: First Run 105 e) Neural Clustering: Second Run 105 f) Neural Clustering: Third Run 105 g) Tree Classification: First Run 105 h) Tree Classification: Second Run 105 i) Tree Classification: Third Run 105 Appendix C: Sample Selection Result of ACL 105 -1- 1. Introduction 1.1. Background Auditing is a relatively archaic field and the auditors are frequently viewed as stuffily fussy people. That is no longer true. In recent years, auditors have recognized the dramatic increase in the transaction volume and complexity of their clients’ accounting and non-accounting records. Consequently, computerized tools such as general-purpose and generalized audit software (GAS) have increasingly been used to supplement the traditional manual audit process. The emergence of enterprise resource planning (ERP) system, with the concept of integrating all operating functions together in order to increase the profitability of an organization as a whole, makes accounting system no longer a simple debit-and-credit system. Instead, it is the central registrar of all operating activities. Though it can be argued which is, or which is not, accounting transaction, still, it contains valuable information. It is auditors’ responsibility to audit sufficient amount of transactions recorded in the client’s databases in order to gain enough evidence on which an audit opinion may be based and to ensure that there is no risk left unaddressed. The amount and complexity of the accounting transactions have increased tremendously due to the innovation of electronic commerce, online payment and other high-technology devices. Electronic records have become more common; therefore, online auditing is increasingly challenging let alone manual access. Despite those complicated accounting transactions can now be presented in the more comprehensive format using today’s improved generalized audit software (GAS), they still require auditors to make assumptions, perform analysis and interpret the results. The GAS or other computerized tools currently used only allows auditors to examine a company’s data in certain predefined formats by running varied query commands but not to extract any information from that data especially when such information is unknown and hidden. Auditors need something more than presentation tools to enhance their investigation of fact, or simply, material matters. On the other side, data mining techniques have improved with the advancement of database technology. In the past two decades, database has become commonplace in -2- business. However, the database itself does not directly benefit the company; in order to reap the benefit of database, the abundance of data has to be turned into useful information. Thus, Data mining tools that facilitate data extraction and data analysis have received greater attention. There seems to be opportunities for auditing and data mining to converge. Auditing needs a mean to uncover unusual transaction patterns and data mining can fulfill that need. This thesis attempts to explore the opportunities of using data mining as a tool to improve audit performance. The effectiveness of various data mining tools in reaching that goal will also be evaluated. 1.2. Research Objective The research objective of this thesis is to preliminarily evaluate the usefulness of data mining techniques in supporting auditing by applying selected techniques with available data sets. However, it is worth nothing that the data sets available are still in question whether it could be induced as generalization. According to the data available, the focus of this research is sample selection step of the test of control process. The relationship patterns discovered by data mining techniques will be used as a basis of sample selection and the sample selected will be compared with the sample drawn by generalized audit software. 1.3. Thesis Structure The remainder of this thesis is structured as follows: Chapter 2 is a brief introduction to auditing. It introduces some essential auditing terms as a basic background. The audit objectives, audit engagement processes and audit approaches are also described here. Chapter 3 discusses some computer assisted auditing tools and techniques currently used in assisting auditors in their audit work. The main focus will be on the generalized audit software (GAS), particularly in Audit Command Language (ACL) -the most popular software in recent years. Chapter 4 provides an introduction to data mining. Data mining process, tools and techniques are reviewed. Also, the discussions will attempt to explore the concept, -3- methods and appropriate techniques of each type of data mining patterns in greater detail. Additionally, some examples of the most frequently used data mining algorithms will be demonstrated as well. Chapter 5 explores many areas where data mining techniques may be utilized to support the auditors’ performance. It also compares GAS packages and data mining packages from the auditing profession’s perspective. The characteristics of these techniques and their roles as a substitution of manual processes are also briefly discussed. For each of those areas, audit steps, potential mining methods, and required data sets are identified. Chapter 6 describes the selected research methodology, the reasons for selection, and relevant material to be used. The research method and the analysis technique of the results are identified as well. Chapter 7 illustrates the actual study. The hypothesis, relevant facts of the research processes and the study results are presented. Finally, the interpretation of study results will be attempted. Finally, chapter 8 provides a summary of the entire study. The assumptions, restrictions and constraints of the research will be reviewed, followed by suggestions for further research. -4- 2. Auditing 2.1. Objective and Structure The objective of this chapter is to introduce the background information on auditing. In section 2.2, definitions of essential terms as well as main objectives and tasks of auditing profession are covered. Four principal audit procedures are discussed in section 2.3. Audit approaches including test of controls and substantive tests are discussed in greater details in section 2.4. Finally, section 2.5 provides a brief summary of auditing perspective. Notice that dominant content covered in this chapter are based on the notable textbook “Auditing: An Integrated Approach” (Arens & Loebbecke, 2000) and my own experiences. 2.2. What Is Auditing? Auditing is the accumulation and evaluation of evidence about information to determine and report on the degree of correspondence between the information and established criteria (Arens & Loebbecke, 2000, 16). Normally, independent auditors, also known as certified public accountants (CPAs), conduct audit work to ascertain whether the overall financial statements of a company are, in all material respects, in conformity with the generally accepted accounting principles (GAAP). Financial statements include Balance Sheets, Profit and Loss Statements, Statements of Cash Flow and Statements of Retained Earning. Generally speaking, what auditors do is to apply relevant audit procedures, in accordance with GAAP, in the examination of the underlying records of a business, in order to provide a basis for issuing a report as an attestation of that company’s financial statements. Such written report is called auditor’s opinion or auditor’s report. Auditor’s report expresses the opinion of an independent expert regarding the degree of reliability upon of the information presented in the financial statements. In other words, auditor’s report assures the financial statements users, which normally are external parities such as shareholders, investors, creditors and financial institutions, of the reliability of financial statements, which are prepared by the management of the company. -5- Due to the time and cost constraints, auditors cannot examine every detail records behind the financial statements. The concept of materiality and fairly stated financial statements were introduced to solve this problem. Materiality is the magnitude of an omission or misstatement of information that misleads the financial statement users. The materiality standard applied to each account balance is varied and is depended on auditors’ judgement. It is the responsibility of the auditors to ensure that all material misstatements are indicated in the auditors’ opinion. In business practice, it is more common to find an auditor as a staff of an auditing firm. Generally, several CPAs join together to practice as partners of the auditing firm, offering auditing and other related services including auditing and other reviews to interested parties. The partners normally hire professional staffs and form an audit team to assist them in the audit engagement. In this thesis, auditors, auditing firm and audit team are synonyms. 2.3. Audit Engagement Processes The audit engagement processes of each auditing firm may be different. However, they generally involve the four major steps: client acceptance or client continuance, planning, execution and documentation, and completion. 2.3.1. Client Acceptance or Client Continuance Client acceptance, or client continuance in case of a continued engagement, is a process through which the auditing firm decides whether or not the firm should be engaged by this client. Major considerations are: - Assessment of engagement risks: Each client presents different level of risk to the firm. The important risk that an auditing firm must evaluate carefully in accepting an audit client are: accepting a company with a bad reputation or questionable ethics that involves in illegal business activities or material misrepresentation of business and accounting records. Some auditing firms have basic requirements of favorable clients. On the other hand, some have a list of criteria to identify the unfavorable ones. Unfavorable clients, for example, are in dubious businesses or have too complex a financial structure. -6- - Relationship conflicts: Independence is a key requirement of the audit profession, of equal importance is the auditor’s objectivity and integrity. These factors help to ensure a quality audit and to earn people’s trust in the audit report. - Requirements of the clients: The requirements include, for example, the qualification of the auditor, time constraint, extra reports and estimated budget. - Sufficient competent personnel available - Cost-Benefit Analysis: It is to compare the potential costs of the engagement with the audit fee offered from the client. The major portion of the cost of audit engagement is professional staff charge. If the client is accepted, a written confirmation, generally on an annual basis, of the terms of engagement is established between the client and the firm. 2.3.2. Planning The objective of the planning step is to develop an audit plan. It includes team mobilization, client’s information gathering, risk assessment and audit program preparation. 2.3.2.1. Team Mobilization This step is to form the engagement team and to communicate among team members. First, key team members have to be identified. Team members include engagement partner or partners who will sign the audit report, staff auditors who will conduct most of the necessary audit work and any specialists that are deemed necessary for the engagement. The mobilization meeting, or pre-planning meeting, should be conducted to communicate all engagement matters including client requirements and deliverables, level of involvement, tentative roles and responsibilities of each team member and other relevant substances. The meeting should also cover the determination of the most efficient and effective process of information gathering. In case of client continuance, a review of the prior year audit to assess scope for improving efficiency or effectiveness should be identified. -7- 2.3.2.2. Client’s Information Gathering In order to perform this step, the most important thing is the cooperation between the client and the audit team. A meeting is arranged to update the client’s needs and expectations as well as management’s perception of their business and the control environment. Next, the audit team members need to perform the preliminary analytical procedures which could involve the following tasks: - Obtaining background information: It includes the understanding of client’s business and industry, the business objectives, legal obligations and related risks. - Understanding system structures: System structures include the system and computer environments, operating procedures and the controls embedded in those procedures. - Control assessment: Based upon information about controls identified from the meeting with the client and the understanding of system structures and processes, all internal controls are updated, assessed and documented. The subjects include control environment, general computerized (or system) controls, monitoring controls and application controls. More details about internal control, such as definitions, nature, purpose and means of achieving effective internal control, can be found in “Internal Control – Integrated Framework” (COSO, 1992). Audit team members’ knowledge, expertise and experiences are considered as the most valuable tools in performing this step. 2.3.2.3. Risk Assessment Risk, in this case, is some level of uncertainty in performing audit work. Risks identified in the first two steps are gathered and assessed. The level of risks assessed in this step is directly lead to the audit strategy to be used. In short, the level of task is based on the level of risks. Therefore, the auditor must be careful not to understate or overstate the level of these risks. -8- Level of risks is different from one auditing area to another. In planning the extent of audit evidences of each auditing area, auditors primarily use an audit risk model such as the one shown below: Planned Detection Risk = Acceptable Audit Risk Inherent Risk * Control Risk - Planned detection risk: Planned detection risk is the highest level of misstatement risk that the audit evidence cannot detect in each audit area. The auditors need to accumulate audit evidences until the level of misstatement risk is reduced to planned detection risk level. For example, if the planned detection risk is 0.05, then audit testing needs to be expanded until audit evidence obtained supports the assessment that there is only five percent misstatement risk left. - Acceptable audit risk: Audit risk is the probability that auditor will unintentionally render inappropriate opinion on client’s financial statements. Acceptable audit risk, therefore, is a measure of how willing the auditor is to accept that the financial statements may be materially misstated after the audit is completed (Arens & Loebbecke, 2000, 261). - Inherent risk: Inherent risk is the probability that there are material misstatements in financial statements. There are many risk factors that affect inherent risk including errors, fraud, business risk, industry risk, and change risk. The first two are preventable and detectable but others are not. Auditors have to ensure that all risks are taken into account when considering the probability of inherent risk. - Control risk: Control risk is the probability that a client’s control system cannot prevent or detect errors. Normally, after defining inherent risks, controls that are able to detect or prevent such risks are identified. Then, auditors will assess whether the client’s system has such controls and, if it has, how much they can rely on those controls. The more reliable controls, the lower the control risk. In other words, control risk represents auditor’s reliance on client’s control structure. It is the responsibility of the auditors to ensure that no risk factors of each audit area are left unaddressed and the evidence obtained is sufficient to reduce all risks to an acceptable audit risk level. More information about audit risk can be -9- found in Statement of Auditing Standard (SAS) No. 47: Audit Risk and Materiality in Conducting an Audit (AICPA, 1983). 2.3.2.4. Audit Program Preparation The purpose of this step is to determine the most appropriate audit strategy and tasks for each audit objective within each audit area based on client’s background information about related audit risks and controls identified from the previous steps. Firstly, the audit objectives, both transaction-related and balancerelated, of each audit area have to be identified. These two types of objectives share one thing in common -- that they must be met before auditors can conclude that the information presented in the financial statements are fairly stated. The difference is that while transaction-related audit objectives are to ensure the correctness of the total transactions for any given class, balance-related audit objectives are to ensure the correctness of any given account balance. A primary purpose of audit strategy and task is to ensure that those objectives are materially met. Such objectives include the following. Transaction-Related and Balance-Related Audit Objectives - Existence or occurrence: To ensure that all balances in the balance sheet have really existed and the transactions in the income statement have really occurred. - Completeness: To ensure that all balances and transactions are included in the financial statements. - Accuracy: To ensure that the balances and transactions are recorded accurately. - Classification: To ensure that all transactions are classified in the suitable categories. - Cut-off (timing): To ensure that the transactions are recorded in the proper period. - 10 - Others Balance-Related Audit Objectives - Valuation: To ensure that the balances and transactions are stated at the appropriate value. - Right and obligation: To ensure that the assets are belonged to and the liabilities are the obligation of the company. - Presentation and disclosure: To ensure that the presentation of the financial statements does not mislead the users and the disclosures are enough for users to understand the financial statements clearly. After addressing audit objectives, it is time to develop an overall audit plan. The audit plan should cover audit strategy of each area and all details related to the engagement including the client’s needs and expectations, reporting requirements, timetable. Then, the planning at the detail level has to be performed. This detailed plan is known as a tailored audit program. It should cover tasks identification and schedule, types of tests to be used, materiality thresholds, acceptable audit risk and person responsible. Notice that related risks and controls of each area are taken into account for prescribing audit strategy and tasks. The finalized general plan should be communicated to the client in order to agree upon significant matters such as deliverables and timetable. Both overall audit plan and detailed audit programs need to be clarified to the team as well. 2.3.3. Execution and Documentation In short, this step is to perform the audit examinations by following the audit program. It includes audit tests execution, which will be described in more detail in the next subsection, and documentation. Documentation includes summarizing the results of audit tests, level of satisfaction, matters found during the tests and recommendations. If there is an involvement of specialists, the process performed and the outcome have to be documented as well. Communication practices are considered as the most important skill to perform this step. Not only with the client or the staff working for the client, it is also - 11 - crucial to communicate among the team. Normally, it is a responsibility of the more senior auditor to coach the less senior ones. Techniques used are briefing, coaching, discussing, and reviewing. A meeting with client in order to discuss the issues found during the execution process and the recommendations of those findings can be arranged either formally or informally. It is a good idea to inform and resolve those issues with the responsible client personnel such as the accounting manager before the completion step and leave only the critical matters to the top management. 2.3.4. Completion This step is similar to the final step of every other kind of projects. The results of aforementioned steps are summarized, recorded, assessed and reported. Normally, the assistant auditors report their work results to the senior, or in-charge, auditors. The auditor-in-charge should perform the final review to ensure that all necessary tasks are performed and that the audit evidence gathered for each audit area is sufficient. Also, the critical matters left from the execution process have to be resolved. The resolution of those matters might be either solved by client’s management (adjusting their financial statements or adequately disclosing them in their financial statement) or by auditors (disclosing them in the auditor’s opinion). The last field work for auditors is review of subsequent events. Subsequent events are events occurred subsequent to the balance sheet date but before the auditor’s report date that require recognition in the financial statements. Based on accumulated audit evidences and audit findings, the auditor’s opinion can be issued. Types of auditor’s opinion are unqualified, unqualified with explanatory paragraph or modified wording, qualified, adverse and disclaimer. After everything is done, it is time to arrange the clearance meeting with the client. Generally, auditors are required to report results and all conditions to the audit committee or senior management. Although not required, auditors often make suggestions to management to improve their business performance through the Management Letter. On the other hand, auditors can get feedback from the client according to their needs and expectations as well. - 12 - Also, auditors should consider evaluating their own performances in order to improve their efficiency and effectiveness. The evaluation includes summarizing client’s comments, bottom-up evaluation (more senior auditors evaluate the work of assistant auditors) and top-down evaluation (get feedback from field work auditors). 2.4. Audit Approaches In order to determine whether financial statements are fairly stated, auditors have to perform audit tests to obtain competent evidence. The audit approaches used in each audit area as well as the level of test depended on auditors’ professional judgement. Generally, audit approaches fall into one of these two categories: 2.4.1. Tests of Controls There are as many control objectives as many textbooks about system security nowadays. However, generally, control objectives can be categorized into four broad categories -- validity, completeness, accuracy and restricted access. With these objectives in mind, auditors can distinguish control activities from the normal operating ones. When assessing controls during planning phase, auditors are able to identify the level of control reliance -- the level of controls that help reducing risks. The effectiveness of such controls during the period can be assessed by performing testing of controls. However, only key controls will be tested and the level of tests depends solely on the control reliance level. The higher control reliance is, the more tests are performed. The scope of tests should be sufficiently thorough to allow the auditor to draw a conclusion as to whether controls have operated effectively in a consistent manner and by the proper authorized person. In other words, the level of test should be adequate enough to bring assurance of the relevant control objectives. The assurance evidence can be obtained from observation, inquiry, inspection of supporting documents, re-performance or the combination of these. - 13 - 2.4.2. Substantive Tests Substantive test is an approach designed to test for monetary misstatements or irregularities directly affecting the correctness of the financial statement balances. Normally, the level of tests depends on the level of assurance from the tests of controls. When the tests of controls could not be performed either because there is no or low control reliance or because the amount and extensiveness of the evidence obtained is not sufficient, substantive tests are performed. Substantive tests include analytical procedures, detailed tests of transactions as well as detailed tests of balances. Details of each test are as follows: 2.4.2.1. Analytical Procedures The objective of this approach is to ensure that overall audit results, account balances or other data presented in the financial statements are stated reasonably. Statement of Auditing Standard (SAS) No. 56 also requires auditors to use analytical procedures during planning and final reporting phases of audit engagement (AICPA, 1988). Analytical procedures can be performed in many different ways. Generally, the most accepted one is to develop the expectation of each account balance and the acceptable variation or threshold. Then, this threshold is compared with the actual figure. Further investigation is required only when the difference between actual and expectation balances falls out of the acceptable variation range prescribed. Further investigation includes extending analytical procedures, detail examination of supporting documents, conducting additional inquiries and performing other substantive tests. Notice that the reliabilities of data, the predictive method and the size of the balance or transactions can strongly affect the reliability of assurance. Moreover, this type of test requires significant professional judgement and experience. 2.4.2.2. Detailed Tests of Transactions The purpose of detailed tests of transactions (also known as substantive testing of transactions) is to ensure that the transaction-related audit objectives are met in each accounting transaction. The confidence on transactions will - 14 - lead to the confidence on the account total in the general ledger. Testing techniques include examination of relevant documents and re-performance. The extent of tests remains a matter of professional judgement. It can be varied from a sufficient amount of samples to all transactions depending on the level of assurance that auditors want to obtain. Generally, samples are drawn either from the items with particular characteristics or randomly sampled or a combination of both. Examples of the particular characteristics are size (materiality consideration) and unusualness (risk consideration). This approach is time-consuming. Therefore, it is a good idea to reduce the sampling size by considering whether analytical procedures or tests of controls can be performed to obtain assurance in relation to the items not tested. 2.4.2.3. Detailed Tests of Balances Detailed tests of balances (also called substantive tests of balances) focuses on the ending balances of each general ledger account. They are performed after the balance sheet date to gather sufficient competent evidence as a reasonable basis for expressing an opinion on fair presentation of financial statements (Rezaee, Elam & Sharbatoghlie, 2001, 155). The extent of tests depends on the results of tests of control, analytical procedures and detailed tests of transactions relating to each account. Like detailed tests of transactions, the sample size can be varied and remains a matter of professional judgement. Techniques to be applied for this kind of tests include account reconciliation, third party confirmation, observation of the items comprising an account balance and agreement of account details to supporting documents. 2.5. Summary Auditing is the accumulation and evaluation of evidence about information to determine and report on the degree of correspondence between the information and established criteria. As seen in figure 2.1, the main audit engagement processes are client acceptance, planning, execution and completion. Client Acceptance - 15 - Gather Information Evaluate client Mobilize Planning Gather information in details Perform preliminary analytical procedures Assess risk and control Set materiality Execution & Documentation Develop audit plan and detailed audit program Perform Tests of Controls trol Low Con ance i High Rel Perform Substantive Tests - Detailed Tests of Transactions - Analytical Procedures - Detailed Tests of Balances Analytical Review - Develop expectations - Compare expectations with actual figures - Further investigate for major differences - Evaluate Results Document testing results Gather audit evidence and audit findings Completion Tests of Controls - Identify controls - Assess control reliance - Select samples - Test controls - Further investigate for unusual items - Evaluate Results Review subsequent events Evaluate overall results Detailed Tests - Select samples - Test samples - Further investigate for unusual items - Evaluate results Issue auditor’s report Arrange clearance meeting with client Evaluate team performance Figure 2.1: Summary of audit engagement processes Planning includes mobilization, information gathering, risk assessment and audit program preparation. Two basic types of audit approaches the auditors can use during execution phase are tests of controls and substantive tests. Substantive tests include analytical procedures, detailed tests of transactions and detailed tests of - 16 - balances. The extent of test is based on the professional judgement of auditors. However, materiality, control reliance and risks are also major concerns. The final output of audit work is auditor’s report. The type of audit report -unqualified, unqualified with explanatory paragraph or modified wording, qualified, adverse or disclaimer -- depends on the combination of evidences obtained from the field works and the audit findings. At the end of each working period, the accumulated evidence and performance evaluation should be reviewed to assess scope for improving efficiency or effectiveness for the next auditing period. It is accepted that auditing business is not a profitable area of auditing firms. Instead, the value-added services, also known as assurance services, such as consulting and legal service are more profitable. The reason is that while cost of all services are relatively the same, clients are willing to pay a limited amount for auditing service comparing to other services. However, auditing has to be trustworthy and standardized and all above-mentioned auditing tasks are, more or less, time-consuming and require professional staff involvement. Thus, the main cost of auditing engagement is the salary of professional staffs and it is considerably high. This cost pressure is a major problem the auditing profession is facing nowadays. To improve profitability of auditing business, the efficient utilization of professional staff seems to be the only practical method. The question is how. Some computerized tools and techniques are introduced into auditing profession in order to assist and enhance auditing tasks. However, the level of automation is still questionable. As long as they still require professional staff involvement, auditing cost is unavoidable high. - 17 - 3. Current Auditing Computerized Tools 3.1. Objective and Structure The objective of this chapter is to provide information about technological tools and techniques currently used by auditors. Section 3.2 discusses why computer assisted auditing tools (CAATs) are more than requisite in auditing profession at present. In section 3.3, general audit software (GAS) is reviewed in detail. The topic focuses on the most popular software, Audit Command Language (ACL). Other computerized tools and techniques are briefly identified in section 3.4. Finally, a brief summary of some currently used CAATs is provided in section 3.5. Before proceeding, it is worth noting that this chapter was mainly based on two textbooks and one journal, which are “Accounting Information Systems” (Bonar & Hopwood, 2001), “Core Concept of Accounting Information System” (Moscove, Simkin & Bagranoff, 2000) and “Audit Tools” (Needleman, 2001). 3.2. Why Computer Assisted Auditing Tools? It is accepted that advances in technology have affected the audit process. With the ever increasing system complexity, especially the computer-based accounting information systems, including enterprise resource planning (ERP), and the vast amount of transactions, it is impractical for auditors to conduct the overall audit manually. It is even more impossible in an e-commerce intensive environment because all accounting data auditors need to access are computerized. In the past ten years, auditors frequently outsource technical assistance in some auditing areas from information system (IS) auditor, also called electronic data processing (EDP) auditor. However, when the computer-based accounting information systems become commonplace, such technical skill is even more important. The rate of growth of the information system practices within the big audit firms (known as “the Big Five”) was estimated at between 40 to 100 percent during 1990 and 2005 (Bagranoff & Vendrzyk, 2000, 35). Nowadays, the term “auditing with the computer” is extensively used. It describes the employment of the technologies by auditors to perform some audit work - 18 - that otherwise would be done manually or outsource. Such technologies are extensively referred to as computer assisted auditing tools (CAATs) and they are now play an important role in audit work. In auditing with the computer, auditors employ CAATs with other auditing techniques to perform their work. As its name suggests, CAAT is a tool to assist auditors in performing their work faster, better, and at lower cost. As CAATs become more common, this technical skill is as important to auditing profession as auditing knowledge, experience and professional judgement. There are a variety of software available to assist the auditors. Some are general-purpose software and some are specially designed that are customized to be used to support the entire audit engagement processes. Many auditors consider simple general ledger, automated working paper software or even spreadsheet as audit software. In this thesis, however, the term audit software refers to software that allows the auditors to perform overall auditing process that generally known as the generalized audit software. 3.3. Generalized Audit Software Generalized audit software (GAS) is an automated package originally developed in-house by professional auditing firms. It facilitates auditor in performing necessary tasks during most audit procedures but mostly in the execution and documentation phase. Basic features of a GAS are data manipulation (including importing, querying and sorting), mathematical computation, cross-footing, stratifying, summarizing and file merging. It also involves extracting data according to specification, statistical sampling for detailed tests, generating confirmations, identifying exceptions and unusual transactions and generating reports. In short, they provide auditors the ability to access, manipulate, manage, analyze and report data in a variety of formats. Some packages also provide the more special features such as risk assessment, high-risk transaction and unusual items continuous monitoring, fraud detection, key performance indicators tracking and standardized audit program generation. With the standardized audit program, these packages help the users to adopt some of the profession's best practices. - 19 - Most auditing firms, nowadays, have either developed their own GASs or purchased some commercially available ones. Among a number of the commercial packages, the most popular one is the Audit Command Language (ACL). ACL is widely accepted as the leading software for data-access, analysis and reporting. Some in-house GAS systems of those large auditing firms even allow their systems to interface with ACL for data extraction and analysis. Figure 3.1: ACL software screenshot (version 5.0 Workbook) ACL software (figure 3.1) is developed by ACL Services Ltd. (www.acl.com). It allows auditors to connect personal laptops to the client’s system and then download client’s data into their laptops for further processing. It is capable of working on large data set that makes testing at hundred-percent coverage possible. Moreover, it provides a comprehensive audit trail by allowing auditors to view their files, steps and results at any time. The popularity of the ACL is resulted from its convenience, its flexibility and its reliability. Table 3.1 illustrates the features of ACL and how are they used in each step of audit process. - 20 - Audit Processes ACL Features Planning - Risk assessment - “Statistics” menu - “Evaluation” menu Execution and Documentation Tests of Controls - Sample selection - “Sampling” menu with the ability to specify sampling size and selection criteria - “Filter” menu - Controls Testing - “Analyze” menu including Count, Total, Statistics, Age, Duplicate, Verify and Search - Expression builder - Results evaluation - Evaluation menu Analytical Review - Expectations development - “Statistics” menu - Expected versus actual figures - “Merge” command comparison - “Analyze” menu including Statistics, Age, Verify and Search - Expression builder - Results evaluation - Evaluation menu Table 3.1: ACL features used in assisting each step of audit processes - 21 - Audit Processes ACL Features Detailed Tests - Sample selection - “Sampling” menu with the ability to specify sampling size and selection criteria - “Filter” menu - Sample testing - “Analyze” menu including Count, Total, Statistics, Age, Duplicate, Verify and Search - Expression builder - Results evaluation Documentation - Evaluation menu - Document note - Automatic command log - File history Completion - Lesson learned record - “Document Notes” menu - “Reports” menu Other Possibilities - Fraud detection - “Analyze” menu including Count, Total, Statistics, Age, Duplicate, Verify and Search - Expression builder - “Filter” menu Table 3.1: ACL features used in assisting each step of audit processes (Continued) With ACL’s capacity and speed, auditors can shorten the audit cycle with more thorough investigation. There are three beneficial features that make ACL a promising tool for auditors. First, the interactive capability allows auditors to test, investigate, analyze and get the results at the appropriate time. Second, the audit trail capability - 22 - records history of the files, commands used by auditors and the results of such commands. This includes command log files that are, in a way, considered as record of work done. Finally, the reporting capability produces various kinds of report including both predefined and customized ones. However, there are some shortcomings. The most critical one is that, like other GAS, it is not able to deal with files that have complex data structure. Although ACL’s Open Data Base Connectivity (ODBC) interface is introduced to reduce this problem, some intricate files still require flattening. Thus, it presents control and security problems. 3.4. Other Computerized Tools and Techniques As mentioned above, there are many other computerized tools other than audit software that are capable of assisting some part of the audit processes. Those tools include the following: - Planning tools: project management software, personal information manager, and audit best practice database, etc. - Analysis tools: database management software, and artificial intelligence. - Calculation tools: spreadsheet software, database management software, and automated working paper software, etc. - Sample selection tools: spreadsheet software. - Data manipulation tools: database management software. - Documents preparation tools: word processing software and automated working paper software. In stead of using these tools as a substitution of GAS, auditors can incorporate some of these tools with GAS to improve the efficiency of the audit process. Planning tools is a good example. Together with the computerized tools, computerized auditing technique that used to be performed by the EDP auditors has now become part of an auditor’s repertoire. At least, financial auditors are required to understand what technique to use, - 23 - how to apply those techniques, and how to interpret the result to support their audit findings. Such techniques should be employed appropriately to accomplish the audit objectives. Some examples are as follows: - Test data: test how the system detect invalid data, - Integrated test facility: observe how fictitious transactions are processed, - Parallel simulation: simulate the original transactions and compare the results, - System testing: test controls of the client’s accounting system, and - Continuous auditing: embed audit program into client’s system. 3.5. Summary In these days, technology impacts the ways auditors perform their work. To conduct the audit, auditors can no longer rely solely on their traditional auditing techniques. Instead, they have to combine such knowledge and experience with technical skills. In short, the boundary between the financial auditor and the information system auditor has becomes blurred. Therefore, it is important for the auditors to keep pace with the technological development so that they can decide what tools and techniques to be used and how to use them effectively. Computer assisted auditing tools (CAATs) are used to compliment the manual audit procedures. There are many CAATs available in the market. The challenge to the auditors is to choose the most appropriate ones for their work. Both the generalized audit software (GAS), that integrates overall audit functions, and other similar software are available to support their work. However, GAS packages tend to be more widely used due to its low cost, high capabilities and high reliability. - 24 - 4. Data mining 4.1. Objective and Structure The objective of this chapter is to describe the basic concept of data mining. Section 4.2 provides some background on data mining and explains its basic element. Section 4.3 describes data mining processes in greater detail. Data mining tools and techniques are discussed in section 4.4 and methods of data mining algorithms are discussed in section 4.5. Examples of most frequently used data mining algorithms are provided in section 4.6. Finally, the brief summary of data mining is reviewed in section 4.7. Notice that the major contents in this chapter are based on “CRISP-DM 1.0 Step-by-Step Data Mining Guide” (CRISP-DM, 2000), “Data Mining: Concepts and Techniques” (Han & Kamber, 2000) and “Principles of Data Mining” (Hand, Heikki & Smyth 2001). 4.2. What Is Data Mining? Data mining is a set of computer-assisted techniques designed to automatically mine large volumes of integrated data for new, hidden or unexpected information, or patterns. Data mining is sometimes known as knowledge discovery in databases (KDD). In recent years, database technology has advanced in stride. Vast amounts of data have been stored in the databases and business people have realized the wealth of information hidden in those data sets. Data mining then become the focus of attention as it promises to turn those raw data into valuable information that businesses can use to increase their profitability. Data mining can be used in different kinds of databases (e.g. relational database, transactional database, object-oriented database and data warehouse) or other kinds of information repositories (e.g. spatial database, time-series database, text or multimedia database, legacy database and the World Wide Web) (Han, 2000, 33). Therefore, data to be mined can be numerical data, textual data or even graphics and audio. - 25 - The capability to deal with voluminous data sets does not mean data mining requires huge amount of data as input. In fact, the quality of data to be mined is more important. Aside from being a good representative of the whole population, the data sets should contain the least amount of noise -- errors that might affect mining results. There are many data mining goals have been recognized; these goals may be grouped into two categories -- verification and discovery. Both of the goals share one thing in common -- the final products of mining process are discovered patterns that may be used to predict the future trends. In the verification category, data mining is being used to confirm or disapprove identified hypotheses or to explain events or conditions observed. However, the limitation is that such hypotheses, events or conditions are restricted by the knowledge and understanding of the analyst. This category is also called top-down approach. Another category, the discovery, is also known as bottom-up approach. This approach is simply the automated exploration of hitherto unknown patterns. Since data mining is not limited by the inadequacy of the human brain and it does not require a stated objective, inordinate patterns might be recognized. However, analysts are still required to interpret the mining results to determine if they are interesting. In recent years, data mining has been studied extensively especially on supporting customer relationship management (CRM) and fraud detection. Moreover, many areas have begun to realize the usefulness of data mining. Those areas include biomedicine, DNA analysis, financial industry and e-commerce. However, there are also some criticisms on data mining shortcomings such as its complexity, the required technical expertise, the lower degree of automation, its lack of user friendliness, the lack of flexibility and presentation limitations. Data mining software developers are now trying to mitigate those criticisms by deploying an interactive developing approach. It is expected that with the advancement in this new approach, data mining will continue to improve and attract more attention from other application areas as well. 4.3. Data Mining Process According to CRISP-DM, a consortium that attempted to standardize data mining process, data mining methodology is described in terms of a hierarchical process that includes four levels as shown in Figure 4.1. The first level is data mining phases, - 26 - or processes of how to deploy data mining to solve business problems. Each phase consists of several generic tasks or, in other words, all possible data mining situations. The next level contains specialized tasks or actions to be taken in order to carry out in certain situations. To make it unambiguous, the generic tasks of the second phase have to be enumerated in greater details. The questions of how, when, where and by whom have to be answered in order to develop a detailed execution plan. Finally, the fourth level, process instances, is a record of the actions, decisions and results of an Processes / Phases Generic Tasks Special Tasks Process Instances actual data mining engagement or, in short, the final output of each phase. Figure 4.1: Four level breakdown of the CRISP-DM data mining methodology (CRISP-DM, 2000, 9) The top level, data mining process, consists of six phases which are business understanding, data understanding, data preparation, modeling, evaluation and deployment. Details of each phase are better described as follows. 4.3.1. Business Understanding The first step is to map business issues to data mining problems. Generic tasks of this step include business objective determination, situation assessment, data mining feasibility evaluation and project plan preparation. At the end of the phase, project plan will be produced as a guideline to the whole project. Such plan should include business background, business objectives and deliverables, data mining goals and requirements, resources and capabilities availability and demand, assumptions and constraints identification as well as risks and contingencies assessment. - 27 - This project plan should be dynamic. This means that at the end of each phase or at each prescribed review point, the plan should be reviewed and updated in order to keep up with the situation of the project. 4.3.2. Data Understanding The objective of this phase is to gain insight into the data set to be mined. It includes capturing and understanding the data. The nature of data should be reviewed in order to identify appropriate techniques to be used and the expected patterns. Generic tasks of this phase include data organization, data collection, data description, data analysis, data exploration and data quality verification. At the end of the phase, the results of all above-mentioned tasks have to be reported. 4.3.3. Data Preparation As mentioned above, one of the major concerns in using data mining technique is the quality of data. The objective of this phase is to ensure that data sets are ready to be mined. The process includes data selection (deciding on which data is relevant), data cleaning (removing all, or most, incompleteness, noises and inconsistency), data scrubbing (cleaning data by abrasive action), data integration (combining data from multiple sources into standardized format), data transformation (converting standardized data into ready-to-be-mined and standardized format) and data reduction (removing redundancies and merging data into aggregated format). The end product of this phase includes the prepared data sets and the reports describing the whole processes. The characteristics of data sets could be different from the prescribed ones. Therefore, the review of project plan has to be performed. 4.3.4. Modeling Though, the terms “models” and “patterns” are used interchangeably, there are some differences between them. A model is a global summary of data sets that can describe the population from which the data were drawn while a pattern describes a structure relating to relatively small local part of the data (Hand, Heikki & Smyth, 2001, 165). To make it simplistic, a model can be viewed as a set of patterns. - 28 - In this phase, a set of data mining techniques is applied to the preprocessed data set. The objective is to build a model that most satisfactorily describes the global data set. Steps include data mining technique selection, model design, model construction, model testing, model validation and model assessment. Notice that, typically, several techniques can be used in parallel to the same data mining problem. The model can be focused on either the most promising technique or using many techniques simultaneously. However, the latter technique requires cross-validated capabilities and evaluation criteria. 4.3.5. Evaluation After applying data mining techniques in a model with data sets, the result of the model will be interpreted. However, it does not mean data mining engagement is over once the results are obtained. Such results have to be evaluated in conjunction with business objectives and context. If the results are satisfactory, the engagement can move on to the next phase. Otherwise, another iteration or moving back to the previous phase has to be done. The expertise of analysts is required in this phase. Besides the result of the model, some evaluation criteria should be taken into account. Such criteria include benefits the business would get from the model, accuracy and speed of the model, the actual costs, degree of automation, and scalability. Generic tasks of this phase include evaluating mining result, reviewing processes and determining the next steps. At the end of the phase, the satisfactory model is approved and the list of further actions is identified. 4.3.6. Deployment Data mining results are deployed into business process in this phase. This phase begins with deployment plan preparation. Besides, the plan for monitoring and maintenance has to be developed. Finally, the success of data mining engagement should be evaluated including area to be improved and explored. Another important thing is that the possibility of failure has to be accepted. No matter how well the model is designed and tested, it is just a model that - 29 - was built from a set of sample data sets. Therefore, the ability to adapt to business change and prompt management decision to correct it are required. Moreover, the performance of the model needs to be evaluated on a regular basis. The sequence of those phases is not rigid so moving back and forth between phases is allowed. Besides, the relationship could exist between any phases. At each review point, the next step has to be specified -- a step that can be either forward or backward. The lesson learned during and at the end of each phase should be documented as a guideline for the next phase. Besides, the documentation of all phases as well as the result of deployment should be documented for the next engagement. Details should include results of each phase, matters arising, problem solving options and method selected. Besides CRISP-DM guideline, there are other textbooks dedicating for integrating data mining into business problems. For the sake of simplicity, I would not go into too much detail than mentioned above. However, more information may be found in “Building Data Mining Applications for CRM” (Berson, Smith & Kurt, 2000) and “Data Mining Cookbook” (Rud, 2001). 4.4. Data Mining Tools and Techniques Data mining is developed from many fields including database technology, artificial intelligence, traditional statistics, high-performance computing, computer graphics and data visualization. Hence, there are abundance of data mining tools and techniques available. However, those tools and techniques can be classified into four broad categories, which are database algorithms, statistical algorithms, artificial intelligence and visualization. Details of each category are as follows: 4.4.1. Database algorithms Although data mining does not require large volume of data as input, it is more practical to deploy data mining techniques on large data sets. Data mining is most useful with the information that human brains could not capture. Therefore, it can be said that the objective of data mining is to mine databases for useful information. - 30 - Thus, many database algorithms can be employed in order to assist mining processes especially in the data understanding and preparation phase. The examples of those algorithms are data generalization, data normalization, missing data detection and correction, data aggregation, data transformation, attribute-oriented induction, and fractal and online analytical processing (OLAP). 4.4.2. Statistical algorithms The distinction between statistics and data mining is indistinct as almost all data mining techniques are derived from statistics field. It means statistics can be used in almost all data mining processes including data selection, problem solving, result presentation and result evaluation. Statistical techniques that can be deployed in data mining processes include mean, median, variance, standard deviation, probability, confident interval, correlation coefficient, non-linear regression, chi-square, Bayesian theorem and Fourier transforms. 4.4.3. Artificial Intelligence Artificial intelligence (AI) is the scientific field seeking for the way to locate intelligent behavior in a machine. It can be said that artificial intelligence techniques are the most widely used in mining process. Some statisticians even think of data mining tool as an artificial statistical intelligence. Capability of learning is the greatest benefit of artificial intelligence that is most appreciated in the data mining field. Artificial intelligence techniques used in data mining processes include neural network, pattern recognition, rule discovery, machine learning, case-based reasoning, intelligent agents, decision tree induction, fuzzy logic, genetic algorithm, brute force algorithm and expert system. 4.4.4. Visualization Visualization techniques are commonly used to visualize multidimensional data sets in various formats for analysis purpose. It can be viewed as higher presentation techniques that allow users to explore complex multi-dimensional data in a simpler way. Generally, it requires the integration of human effort to analyze and assess the results from its interactive displays. Techniques include audio, tabular, - 31 - scatter-plot matrices, clustered and stacked chart, 3-D charts, hierarchical projection, graph-based techniques and dynamic presentation. To separate data mining from data warehouse, online analytical processing (OLAP) or statistics is intricate. One thing to be sure of is that data mining is not any of them. The difference between data warehouse and data mining is quite clear. Though there are some textbooks about data warehouse that devoted a few pages to data mining topic, it does not mean that they took data mining as a part of data warehousing. Instead, they all agreed that while data warehouse is a place to store data, data mining is a tool to distil the value of such data. The examples of those textbooks are “Data Management” (McFadden, Hoffer & Prescott, 1999) and “Database Systems : A Practical Approach to Design, Implementation, and Management” (Connolly, Begg & Strachan, 1999). One might argue that the value of data could be realized by using OLAP as claimed in many data warehouse textbooks. OLAP, however, can be thought of as another presentation tool that reform and recompile the same set of data in order to help users find such value easier. It requires human interference in both stating presenting requirements as well as interpreting the results. On the other hand, data mining uses automated techniques to do those jobs. As mentioned above, the differentiation between data mining and statistics is much more complicated. It is accepted that the algorithms underlying data mining tools and techniques are, more or less, derived from statistics. In general, however, statistical tools are not designed for dealing with enormous amount of data but data mining tools are. Moreover, the target users of statistical tools are statisticians while data mining is designed for business people. This simply means that data mining tools are enhancement of statistical tools that blend many statistical algorithms together and possess a capability of handling more data in an automated manner as well as a userfriendly interface. The choice of an appropriate technique and timing depend on the nature of the data to be analyzed, the size of data sets and the type of methods to be mined. A range of techniques can be applied to the problems either alone or in combination. However, when deploying sophisticated blend of data mining techniques, there are at least two - 32 - requirements that need to be met -- the ability to cross validate results and the measurement criteria. 4.5. Methods of Data Mining Algorithms Though nowadays data mining software packages are claimed to be more automated, they still require some directions from users. Expected method of data mining algorithm is one of those requirements. Therefore, in employing data mining tools, users should have a basic knowledge of these methods. The types of data mining methods can be categorized differently. However, in general, they fall into six broad categories which are data description, dependency analysis, classification and prediction, cluster analysis, classification and prediction, cluster analysis, outlier analysis and evolution analysis. Details of each method are as follows: 4.5.1. Data Description The objective of data description is to provide an overall description of data, either in itself or in each class or concept, typically in summarized, concise and precise form. There are two main approaches in obtaining data description -- data characterization and data discrimination. Data characterization is summarizing general characteristics of data and data discrimination, also called data comparison, is comparing characters of data between contrasting groups or classes. Normally, these two approaches are used in aggregated manner. Though data description is one among many types of data mining algorithm methods, usually it is not the real finding target. Often the data description is analyst’s first requirement, as it helps to gain insight into the nature of the data and to find potential hypotheses, or the last one, in order to present data mining results. The example of using data description as a presentation tool is the description of the characteristics of each cluster that could not be identified by neural network algorithm. Appropriate data mining techniques for this method are attribute-oriented induction, data generalization and aggregation, relevance analysis, distance analysis, rule induction and conceptual clustering. - 33 - 4.5.2. Dependency Analysis The purpose of dependency analysis, also called association analysis, is to search for the most significant relationship across large number of variables or attributes. Sometimes, association is viewed as one type of dependencies where affinities of data items are described (e.g., describing data items or events that frequently occur together or in sequence). This type of methods is very common in marketing research field. The most prevalent one is market-basket analysis. It analyzes what products customers always buy together and presents in “[Support, Confident]” association rules. The support measurement states the percentage of events occurring together comparing to the whole population. The confident measurement affirms the percentage of the occurrence of the following events comparing to the leading one. For example, the association rule in figure 4.2 means milk and bread were bought together at 6% of all transactions under analysis and 75% of customers who bought milk also bought bread. Milk => bread [support = 6%, confident = 75%] Figure 4.2: Example of association rule Some techniques for dependency analysis are nonlinear regression, rule induction, statistic sampling, data normalization, Apriori algorithm, Bayesian networks and data visualization. 4.5.3. Classification and Prediction Classification is the process of finding models, also known as classifiers, or functions that map records into one of several discrete prescribed classes. It is mostly used for predictive purpose. Typically, the model construction begins with two types of data sets -training and testing. The training data sets, with prescribed class labels, are fed into the model so that the model is able to find parameters or characters that distinguish one class from the other. This step is called learning process. Then, the testing data sets, without pre-classified labels, are fed into the model. The model will, ideally, automatically assign the precise class labels for those testing items. If the results of - 34 - testing are unsatisfactory, then more training iterations are required. On the other hand, if the results are satisfactory, the model can be used to predict the classes of target items whose class labels are unknown. This method is most effective when the underlying reasons of labeling are subtle. The advantage of this method is that the pre-classified labels can be used as the performance measurement of the model. It gives the confidence to the model developer of how well the model performs. Appropriate techniques include neural network, relevance analysis, discriminant analysis, rule induction, decision tree, case-based reasoning, genetic algorithms, linear and non-linear regression, and Bayesian classification. 4.5.4. Cluster analysis Cluster analysis addresses segmentation problems. The objective of this analysis is to separate data with similar characteristics from the dissimilar ones. The difference between clustering and classification is that while clustering does not require pre-identified class labels, classification does. That is why classification is also called supervised learning while clustering is called unsupervised learning. As mentioned above, sometimes it is more convenient to analyze data in the aggregated form and allow breaking down into details if needed. For data management purpose, cluster analysis is frequently the first required task of the mining process. Then, the most interesting cluster can be focused for further investigation. Besides, description techniques may be integrated in order to identify the character providing best clustering. Examples of appropriate techniques for cluster analysis are neural networks, data partitioning, discriminant analysis and data visualization. 4.5.5. Outlier Analysis Some data items that are distinctly dissimilar to others, or outliers, can be viewed as noises or errors which ordinarily need to be drained before inputting data sets into data mining model. However, such noises can be useful in some cases, where unusual items or exceptions are major concerns. Examples are fraud detection, unusual usage patterns and remarkable response patterns. - 35 - The challenge is to distinguish the outliers from the errors. When performing data understanding phase, data cleaning and scrubbing is required. This step includes finding erroneous data and trying to fix them. Thus, the possibility to detect interesting differentiation might be diminished. On the other hand, if the incorrect data remained in the data sets, the accuracy of the model would be compromised. Appropriate techniques for outlier analysis include data cube, discriminant analysis, rule induction, deviation analysis and non-linear regression. 4.5.6. Evolution Analysis This method is the newest one. The creation of evolution analysis is to support the promising capability of data warehouses which is data or event collection over a period of time. Now that business people came to realize the value of trend capture that can be applied to the time-related data in the data warehouse, it attracts increasing attention in this method. Objective of evolution analysis is to determine the most significant changes in data sets over time. In other words, it is other types of algorithm methods (i.e., data description, dependency analysis, classification or clustering) plus timerelated and sequence-related characteristics. Therefore, tools or techniques available for this type of methods include all possible tools and techniques of other types as well as time-related and sequential data analysis tools. The examples of evolution analysis are sequential pattern discovery and time-dependent analysis. Sequential pattern discovery detects patterns between events such that the presence of one set of items is followed by another (Connolly, 1999, 965). Time-dependent analysis determines the relationship between events that correlate in a definite of time. Different types of methods can be mined in parallel to discover hidden or unexpected patterns, but not all patterns found are interesting. A pattern is interesting if it is easily understood, valid, potentially useful and novel (Han & Kamber, 2000, 27). Therefore, analysts are still needed in order to evaluate whether the mining results are interesting. - 36 - To distinguish interesting patterns, users of data mining tools have to solve at least three problems. First, the correctness of patterns has to be measured. For example, the measurement of dependency analysis is “[Confident, Support]” value. It is easier for the methods that have historical or training data sets to compare the correctness of the patterns with the real ones; i.e., classification and prediction method. For those methods that training data sets are not available, then the professional judgement of the users of data mining tools is required. Second, the optimization model of patterns found has to be created. For example, the significance of “Confident” versus “Support” has to be formulated. To put it in simpler terms, it is how to tell which is better between higher “Confident” with lower “Support” or lower “Confident” with higher “Support”. Finally, the right point to stop finding patterns has to be specified. This is probably the most challenging problem. This leads to two other problems -- how to tell the current optimized pattern is the most satisfactory one and how to know it can be used as a generalized pattern on other data sets. In short, while trying to optimize the patterns, the over-fitting problem has to be taken into account as well. 4.6. Examples of Data Mining Algorithms As mentioned above, there are plenty of algorithms used to mine the data. Due to the limited of space, this section is focused on the most frequently used and widespread recognized algorithms that can be indisputable thought of as data mining algorithms; neither pure statistical, nor database algorithms. The examples include Apriori algorithms, decision trees and neural networks. Details of each algorithms are as follows: 4.6.1. Apriori Algorithms Apriori algorithm is the most frequently used in the dependency analysis method. It attempts to discover frequent item sets using candidate generation for Boolean association rules. Boolean association rule is a rule that concerns associations between the presence or absence of items (Han & Kamber, 2000, 229). The steps of Apriori algorithms are as follows: (a) The analysis data is first partitioned according to the item sets. - 37 - (b) The support count of each item set (1-itemsets), also called Candidate, is performed. (c) The item sets that could not satisfy the required minimum support count are pruned. Thus creating the frequent 1-itemsets (a list of item sets that have at least minimum support count). (d) Item sets are joined together (2-itemsets) to create the second-level candidates. (e) The support count of each candidate is accumulated. (f) After pruning unsatisfactory item sets according to minimum support count, the frequent 2-itemsets is created. (g) The iteration of (d), (e) and (f) are executed until no more frequent kitemsets can be found or, in other words, the next frequent k-itemsets contains empty frequent. (h) At the terminated level, the Candidate with maximum support count wins. By using Apriori algorithms, the group of item sets that most frequently come together is identified. However, dealing with large amount of transactions means the candidate generation, counting and pruning steps needed to be repeated numerous times. Thus, to make the process more efficient, some techniques such as hashing (reducing the candidate size) and transaction reduction can be used (Han & Kamber, 2000, 237). 4.6.2. Decision Trees Decision tree is a predictive model with tree or hierarchical structure. It is used most in classification and prediction methods. It consists of nodes, which contained classification questions, and branches, or the results of the questions. At the lowest level of the tree -- leave nodes -- the label of each classification is identified. The structure of decision tree is illustrated in figure 4.3. Typically, like other classification and prediction techniques, the decision tree begins with exploratory phase. It requires training data sets with labels to be fed. - 38 - The underlying algorithm will try to find the best-fit criteria to distinguish one class from another. This is also called tree growing. The major concerns are the quality of the classification problems as well as the appropriate number of levels of the tree. Some leaves and branches need to be removed in order to improve the performance of the decision tree. This step is also called tree pruning. On the higher level, the predetermined model can be used as a prediction tool. Before that, the testing data sets should be fed into the model to evaluate the Transaction = 50 x > 35 ? No Yes Transaction = 15 y > 52 ? No Transaction = 9 Group E Transaction = 35 y > 25 ? Yes No Transaction = 6 Group D Yes Transaction = 25 x > 65 ? No Transaction = 15 Group A Transaction = 10 Group C Yes Transaction = 10 Group B model performance. Scalability of the model is the major concern in this phase. Figure 4.3: A decision tree classifying transactions into five groups The fundamental algorithms can be different in each model. Probably the most popular ones are Classification and Regression Trees (CART) and Chi-Square Automatic Interaction Detector (CHAID). For the sake of simplicity, I will not go into the details of these algorithms and only perspectives of them are provided. CART is an algorithm developed by Leo Breiman, Jerome Friedman, Richard Olshen and Charles Stone. The advantage of CART is that it automates the - 39 - pruning process by cross validation and other optimizers. It is capable of handling missing data and it sets the unqualified records apart from the training data sets. CHAID is another decision tree algorithm that uses contingency tables and the chi-square test to create the tree. The disadvantage of CHAID comparing to CART is that it requires more data preparation process. 4.6.3. Neural Networks Nowadays, neural networks, or more correctly the artificial neural networks (ANN), attract the most interest among all data mining algorithms. It is a computer model based on the architecture of the brain. To put it simply, it first detects the pattern from data sets. Then, it predicts the best classifiers. And finally, it learns from the mistakes. It works best in classification and prediction as well as clustering methods. The structure of neural network is shown in figure 4.4. First hidden layer Second hidden layer Input Layer Output layer Figure 4.4: A neural network with two hidden layers As noticed in figure 4.4, neural network is comprised of neurons in input layer, one or more hidden layers and output layer. Each pair of neurons is connected with a weight. In the cases where there are more than one input neurons, the input weights are combined using a combination function such as summation (Berry & - 40 - Linoff, 2000, 122). During training phase, the network learns by adjusting the weights so as to be able to predict the correct output (Han & Kamber, 2000, 303). The most well known neural network learning algorithm is backpropagation. It is the method of updating the weights of the neurons. Unlike other learning algorithms, backpropagation algorithm works, or learns and adjusts the weights, backward which simply means that it predicts the weighted algorithms by propagating the input from the output. Neural networks are widely recognized for its robustness; however, the weakness is its lack of self-explanation capability. Though the performance of the model is satisfactory, some people do not feel comfortable or confident to rely irrationally on the model. It should note that some algorithms are good at discovering specific methods where some others are appropriate for many types of methods. The choice of algorithm or set of algorithms used depends solely on user’s judgement. 4.7. Summary Data mining, which is also known as knowledge discovery in databases (KDD), is the area of attention in recent years. It is a set of techniques that exhaustively automated to uncover potentially interesting patterns from a large amount of data in any kind of data repositories. Data mining goals can be roughly divided into two main categories, verification (including explanation and confirmation) and discovery. The first step of the data mining process is to map business problems to data mining problems. Then, data to be mined is captured, studied, selected and preprocessed respectively. The preprocessed activities are performed in order to prepare final data sets to be fed into data mining model. Next, data mining model is constructed, tested, and applied. The results of this step are evaluated subsequently. If the result is satisfactory, then it will be deployed in the real business environment. Lessons learned during data mining engagement should be recorded as guidelines for future project. As data mining is developed from and driven by multidisciplinary fields, different tools and techniques can be applied in each step of data mining process. Those - 41 - tools and techniques include database algorithms, statistic algorithms, artificial intelligence and data visualization. The choice of tools and techniques depends on nature and size of data as well as types of methods to be mined. Types of data mining methods can be categorized into six groups which are data description, dependency analysis, classification and prediction, cluster analysis, outlier analysis and evolution analysis. The appropriate techniques or algorithms of each data mining method are summarized in table 4.1. Among all underlying algorithms of these methods, Apriori algorithms, decision trees and neural networks are the most familiar ones. Data Mining Methods Data description Dependency analysis Classification and prediction Appropriate Data Mining Techniques - Attribute-oriented induction - Data generalization and aggregation - Relevance analysis - Distance analysis - Rule induction - Conceptual clustering - Nonlinear regression - Rule induction - Distance-based analysis - Data normalization - Apriori algorithm - Bayesian network - Visualization - Neural network - Relevance analysis - Discriminant analysis - Rule induction - Decision tress - Case-based reasoning - Genetic algorithms - Linear and nonlinear regression - 42 - Table 4.1: Summarization of appropriate data mining techniques of each data mining method Data Mining Methods Cluster analysis Outlier analysis Evolution analysis Appropriate Data Mining Techniques - Data visualization - Bayesian classification - Neural network - Data partitioning - Discriminant analysis - Data visualization - Data cube - Discriminant analysis - Rule induction - Deviation analysis - Nonlinear regression - All above-mentioned techniques - Time-related analysis - Sequential analysis Table 4.1: Summarization of appropriate data mining techniques of each data mining method (Continued) Data mining already has its market in customer relationship management as well as fraud detection and is expected to penetrate new areas in the near future. However, data mining software packages that are currently available have been criticized for not automated enough and not user-friendly enough. Therefore, with the abundance of market opportunities, the continued improvement and growth in the data mining arena can be anticipated. - 43 - 5. Integration of Data Mining and Auditing 5.1. Objective and Structure The objective of this chapter is to identify ways that data mining techniques may be utilized in the audit process. The reasons why data mining should be integrated with auditing process are reviewed in section 5.2. Section 5.3 provides a comparison between the characteristics of currently used generalized auditing software (GAS) and data mining packages from auditing profession’s perspective. In section 5.4, each possible area of integration is discussed in more details including possible mining methods and required data sets. Lastly, a brief summary of data mining trend in auditing profession is provided. 5.2. Why Integrate Data Mining with Auditing? As mentioned in the first chapter, auditors have realized the dramatically increase of transaction volume and complexity of accounting and non-accounting transactions. The greater amount of the transactions is resulted from new emerging technologies especially business intelligent systems such as enterprise resource planning (ERP) systems and supply chain management (SCM) systems. Now that transactions can be made flexibly online without time constraint, such growth can be unsurprisingly anticipated. Besides online systems and transactions, other hi-technology devices make accounting and managerial transactions more complicated. As transactions are made, recorded and stored electronically, the advanced tools to capture, analyze, present and report are required. Dealing with intricate transactions in large volume, it requires the considerable more effort of professional stuffs and that cannot be cost-effective. Besides, in some cases, professional judgement along might not be sufficient due to human brain’s limitation. Therefore, the capability to automatically manipulating complicated data through data mining is of great interest to the auditing profession. - 44 - On the other hand, the huge auditing market presents tremendous opportunity for data mining business as well. Auditing is one of many application areas that the explosive growth of data mining integration is predicted to continue. Therefore, the opportunity to have data mining tools as an advanced computer assisted auditing tools (CAATs) can be expected before long. 5.3. Comparison between Generalized Auditing Software and Data Mining Packages As mentioned above, nowadays auditors rely exceedingly on generalized audit software (GAS). The objective of this section is to make it more transparent the differences between currently used GAS and data mining packages available in the market considering in auditing profession perspective. This section was mainly based on the features of auditing software gathered from the software developers’ websites and some software review publications. The software packages include the followings: - ACL - Audit Command Language (ACL Services Ltd., 2002) - IDEA - Interactive Data Extraction and Analysis (Audimation Services Inc., 2002) - DB2 Intelligent Miner for Data (IBM Corporation, 2002) - DBMiner (DBMiner Technology Inc., 2002) - Microsoft Data Analyzer (Microsoft Corporation, 2002) - SAS Enterprise Miner (SAS Institute Inc., 2002) - SAS Analytic Intelligence (SAS Institute Inc. 2002) - SPSS (SPSS Inc., 2002) The publications include the followings: - Audit Tools (Needleman, 2001) - How To Choose a PC Auditing Tool (Eurotek Communication Ltd., 2002) - 45 - - Information Systems Auditing and Assurance (Hall, 2000) - Software Showcase (Glover, Prawitt & Romney, 1999) - A Review of Data Mining Techniques (Lee & Siau, 2001) - Data Mining – A Powerful Information Creating Tool (Gargano & Raggad, 1999) - Data Warehousing, Technology Assessment and Management (Ma, Chou & Yen, 2000) 5.3.1. Characteristics of Generalized Audit Software Though the features of each GAS are different from one another, most packages, more or less, share the following characteristics: - All-in-one features: GAS packages are designed to support the entire audit engagement including data access, project management that can be used to manage the engagement, and all audit procedures. - Specifically customized for audit work: As audit work generally follow some predicable approaches, GAS packages can be designed to support those approaches. It means that the auditors do not need to alter the programs before employing them and are able to understand how to work with them easily. Of all the features, the audit trail might be the most valuable one. - User friendliness features: Most GAS packages have a user friendly interface that include easy to use and understand features as well as high presentation capability. - No or little technical skill required: Due to GAS’s user friendly interface and it is specifically designed for audit work, it requires no or little technical skills to work with. - Speed depending on the amount of transaction input: Nearly all GAS packages available nowadays are designed for processing huge amount of transactions that could reach millions. transaction. However, the processing speed depends on the input - 46 - - Professional judgement required: Audit features that are built into GAS packages include sorting, querying, aging and stratifying. However, they still require auditors to interactively observe, evaluate and analyze the results. There are many GAS packages available in the market nowadays. The return on investment for GAS packages is considered high especially comparing to expenses of professional staffs. Therefore, most auditing firms nowadays rely on them a lot. However, experience and professional judgement of auditors are still indispensable. That is, GAS can reduce degree of professional staff requirements but cannot replace any level of professional staffs. 5.3.2. Characteristics of Data Mining Packages Among a plethora of data mining packages available, some characteristics of data mining packages in general are: - Automated capability: The ideal objective of data mining is to automatically discover useful information from a data set. Though today’s data mining packages are still not completely automated, only the guidance to ensure that the results are interesting and evaluation of the results require intensive human efforts. - High complexity: How data mining algorithms work is sometimes mysterious because of their complexity. Their poor self-explaining capability results in low confidence of the result by the users. - Scalability: It could be said that data warehousing is the foundation of data mining evolution. Data mining, therefore, is designed for unlimited amount of data in the data warehouse that makes scalability one of the key features of the data mining characteristics. - Ability to handle complex problems: As its capability not limited by the human brain, data mining is able to handle complex ad hoc problems. - Opportunity to uncover unexpected interesting information: When performing audit work, auditors normally know what they are looking for. This might result in the limited scope of tests. On the other hand, data mining can be used even when users do not have a preconceived notion of what to look for. Therefore, with data - 47 - mining, some interesting information hidden in the accounting transactions can be revealed. - Learning capability embedded: Many data mining algorithms have learning capability. Experiences and mistakes from the past can be used to improve the quality of the models automatically. - Technical skill required: Substantial technical skill is mandatory of the data mining software users. First, users must know the difference between various types of data mining algorithms in order to choose the right one. Second, they are supposed to guide the program to ensure that the findings are interesting. Finally, the result of data mining process must be evaluated. - Lack of interoperability: There are numerous data mining algorithm methods or data mining techniques to be employed. However, software packages currently available are of data mining software users tend to focus on one method and employ only a few techniques. Interoperability between different data mining algorithm methods still presents significant challenges to data mining software developers. - Lack of self-explanation capability: In general, data mining processes are done automatically and the underlying reasons of the result are frequently subtle. From an auditing perspective, this is a major problem because the audibility, audit trails and replicability are key requirements in audit work. - Relatively high cost: Though data mining software has becoming cheaper, it is still somewhat expensive comparing to other software. Besides, in performing data mining, the users have to incur data preparation cost, analyzing cost and training cost. Although data mining is frequently considered as a highly proficient technique in many application areas, it has not been widely adopted in auditing profession yet. However, it is expected to gain increasing popularity in audit. The automation potential of data mining suggests that it could greatly improve the efficiency of audit professionals including replacing level of professional staffs involvement. - 48 - 5.4. Possible Areas of Integration Recently, people in auditing profession have started realizing that technology advancement can facilitate auditing practices. Data mining is one of those technologies that has been proven to be very effective in application areas such as fraud detection, which can be thought of as a part of auditing. However, the integration of data mining techniques and audit work is relatively immature. Based on the above-mentioned theories in chapter 2 and 4 as well as my personal experiences, I tried to go through all auditing steps and list out the possibilities that data mining techniques can be applied into those steps. The opportunities and their details are summarized in table 6.1. Notice that though, in my opinion, there are many audit steps that data mining techniques are likely to be capable of assisting, enhancing or handling, such potentials may not seem realistic at this moment. One might argue that some listed steps can be done effortlessly by accustomed manual procedures with a little help from or even without easy-to-use software packages. The examples of those steps are client’s acceptance, client’s continuance and opinion issuance. I have nothing against such opinion especially when the financial figures or data sets required of those steps are not that much and electronic data is proprietary. However, I still do not want to disregard any steps because they can provide some ideas for further research in the time when those data becomes overflow. Another attention to be drawn is that it is by no mean the table below is definite answer for the integration between auditing procedures and data mining techniques. However, as far as I concern, the major and apparent notions of the topic should be already covered. - 49 - Audit Processes Appropriate Mining Methods Client Acceptance or Client - Classification and prediction Continuance - Evolution analysis Data Sets Required - Previous years financial statements Possibility - By using financial ratios, business rating and industry - Business rating rating, client can be classified - Industry rating as favorable or unfavorable - Economic index (by using classification and - Previous years actual costs (in prediction methods). Then, case of client continuance) together with estimated cost based on previous years’ records and economic index (by using evolution analysis), the acceptance result can be reached. Planning - Risk assessment - Dependency analysis - Classification and prediction - Previous years financial Table 5.1: Possible areas of data mining and audit processes integration statements - By using dependency analysis, risk triggers (e.g. financial - 50 - Audit Processes Appropriate Mining Methods Data Sets Required Possibility - Business rating ratios, business rating, - Industry rating industry rating and controls) - Economic index can be identified. Then, the - System flowcharts level of risk of each audit area can be prescribed by using the risk triggers as criterion (using classification and prediction methods). - Audit program preparation - Classification and prediction - Client’s system information - Results of risk assessment step - The appropriate combination of audit approach for each audit area can be distinguished based on client’s information gathered and risks identified in risk assessment step. Table 5.1: Possible areas of data mining and audit Processes integration (Continued) - 51 - Audit Processes Appropriate Mining Methods Data Sets Required Possibility Execution and Documentation Tests of Controls - Controls identification - Classification - Data description - System information - Controls can be identified from normal activities by using classification analysis; the characteristics of such controls can be identified by data description method. - Control reliance assessment - Classification and prediction - Results of risk assessment step - Results of control identification step - The control reliance level of each area can be categorized based on risks and control over such risks identified in previous steps. - Sample Selection - Cluster analysis - Outlier analysis - Accounting transactions Table 5.1: Possible areas of data mining and audit Processes integration (Continued) - Accounting transactions with similar characters are grouped - 52 - Audit Processes Appropriate Mining Methods Data Sets Required Possibility through clustering. Samples can be selected from each cluster, together with unusual items identified by outlier analysis method. - Controls Testing - Cluster analysis - Outlier analysis - Results of sample selection step - Either by grouping ordinary transactions together or by identifying the outliers, the unusual items or matters can be identified. - Results evaluation - Classification - Results of control testing step - The test results from previous step can be classified as either satisfactory or unsatisfactory. If unsatisfactory, further investigation can be done by iterating the test or using other Table 5.1: Possible areas of data mining and audit Processes integration (Continued) - 53 - Audit Processes Appropriate Mining Methods Data Sets Required Possibility techniques including interviewing responsible personnel, review of supporting documents or additional investigative works. Analytical Review - Expectations development - Classification and prediction - Evolution analysis - Previous years accounting transactions - The expectations of each balance can be predicted based - Business rating on previous years’ balances, - Industry rating current circumstances of the - Economic index business, the state of its industry and the economic environment. - Expected versus actual - Classification figures comparison - Outlier analysis - Results of expectations development step Table 5.1: Possible areas of data mining and audit Processes integration (Continued) - The differences between expected and actual figures are - 54 - Audit Processes Appropriate Mining Methods Data Sets Required - Accounting transactions Possibility grouped. Those that do not fall into acceptable range should be identified and further investigated. - Results evaluation - Classification - Results of expected versus actual figures comparison step - The test results from previous step can be classified as either satisfactory or unsatisfactory. If unsatisfactory, further investigation can be done by iterating the test or using other techniques including interviewing responsible personnel, review of supporting documents or additional investigative works. Table 5.1: Possible areas of data mining and audit Processes integration (Continued) - 55 - Audit Processes Appropriate Mining Methods Data Sets Required Possibility Detailed Tests - Sample selection - Cluster analysis - Outlier analysis - Accounting transactions - By Accounting transactions with similar characters are grouped through clustering. Samples can be selected from each cluster, together with unusual items identified by outlier analysis method. - Sample testing - Cluster analysis - Outlier analysis - Results of sample selection step - Either be grouping ordinary transactions together or by identifying the outliers, the resulting unusual items or matters arising can be identified. Table 5.1: Possible areas of data mining and audit Processes integration (Continued) - 56 - Audit Processes - Results evaluation Appropriate Mining Methods - Classification Data Sets Required - Results of sample testing step Possibility - The testing results from previous step can be classified as satisfactory or unsatisfactory. In case of unsatisfactory, further investigation can be done by iterating the test. Documentation - Data description - Results of all results evaluation steps - The characters of test results and matters arising can be described and recorded by data description method Completion - Opinion Issuance - Dependency analysis - Classification and prediction - Results of all results evaluation step Using dependency analysis, circumstances or evidence that will affect the types of opinion Table 5.1: Possible areas of data mining and audit Processes integration (Continued) - 57 - Audit Processes Appropriate Mining Methods Data Sets Required Possibility issued can be collected. Then, based on the test results, audit findings, matters surfaced and other related circumstances, types of opinion can be rendered. - Lesson learned record - Data description - Results of all results evaluation steps - The nature of tests, test results, audit findings, matter surfaced and other relevant circumstances can be described and recorded. Table 5.1: Possible areas of data mining and audit Processes integration (Continued) - 58 - Notice that the size of the data set required is generally small. One might argue that when the size of data sets is not that massive, it is no use to change from familiar GAS to complicated data mining packages. One attention to be paid, however, is that the data sets specified are those required for the current year auditing processes. However, to build a data mining model, the training data sets based on historical data are essential as well. Historical data includes both the previous year data of the client itself and that of other businesses in similar and substitute industry. Therefore, the data sets used could be considerably larger in the first audit period. Besides, data sets required of some steps can be substantially large, such as sample selection step with accounting transactions as data sets required. 5.5. Examples of Tests In reality, the worst limitation is the lack of data availability, especially in the first year audit, which makes some steps of the table in previous section do not sound promising. The only certain data available for all audit engagements is general ledger or accounting transactions for the audited period. Therefore, this section is focused on what data mining can contribute when data available is only current period general ledger. As a general note, data mining methods that require historical data as a training data set cannot be done. Examples are classification and prediction, dependency analysis and evolution analysis. However, in some cases, data from previous months can be used as training data sets for the following months. To put it simplistically, the audit steps which are performed at the beginning of the audit engagement require data from previous years to train the model and, thus, are not feasible when only general ledger in the current period is available. Those steps include client acceptance and planing steps. Therefore, execution phase is the only possible phase to use data mining technique in the first year audit when available data is limited. The structure of general ledger of each company may be different. To avoid confusion, the general ledger this section will base on is a simply flat file as shown in figure 5.1 with the common attributes as follows; - 59 - Journal Transaction Number Number Date Document Number Description Account Responsible Authorized Profit Amount Number Person Person Center 0001 01 1/7/2002 E00198 Rental fee 1101 95,000 SCS SVC 100 0001 02 1/7/2002 E00198 Rental fee 7801 - 95,000 SCS SVC 100 0002 01 5/7/2002 S00059 Sales - Customer A 1209 765 SCS SCS 403 0002 02 5/7/2002 S00059 Sales - Customer A 6103 520 SCS SCS 403 0002 03 5/7/2002 S00059 Sales - Customer A 4103 - 765 SCS SCS 403 0002 04 5/7/2002 S00059 Sales - Customer A 1303 - 520 SCS SCS 403 0003 01 6/7/2002 P00035 Purchase - Company Z 1307 7,300 SCS SCS 215 0003 02 6/7/2002 P00035 Purchase - Company Z 1312 450 SCS SCS 215 0003 03 6/7/2002 P00035 Purchase - Company Z 2106 - 7,750 SCS SCS 215 Figure 5.1: Basic structure of general ledger - Journal Number -- or recorded number which can be used as a reference of each balanced transactions. - Transaction Number – include in order to make each record in the file unique. - Date -- ideally, the transaction date and the recorded date should be the same. If not, at least transaction date has to be identified. - Document Number -- refers to source or supporting documents. - Description -- or explanation of each journal. - Account Number - Amount -- normally the debit amount has positive value while credit amount has negative value. - Responsible Person -- or person who prepares or keys in the transaction. - Authorized Person -- some transactions with certain characteristics may require approval. - Other Additional Information -- such as profit center, customer group, currency and spot exchange rate. In this case, profit center is selected as the example. - 60 - Based on this general ledger structure, the detailed examples of tests using data mining techniques in each audit steps of execution phase are presented in table 5.2. However, it is important to note that the examples of interesting patterns in table 5.2 did not include some patterns that can be identified effortlessly by manual procedures or GAS packages. Examples are sample selection based on significant amount, transactions that occurred on pre-identified date (e.g. weekends, holidays) and differences between sub-ledger and general ledger systems. Besides, some audit processes that are easily done manually or by using GAS such as detailed testing and result evaluation are not included. Audit Processes Examples of Tests Test of Controls - Sample selection - Applied techniques: Grouping all accounting transactions with all relevant attributes by using clustering technique. - Examples of interesting patterns: - Transactions approved by unauthorized person. - Transactions that almost reach the limited of authorized person and occurred repeatedly in sequence. - Type of transactions that are always approved by certain authorized person such as transactions of profit center A always approved by authorized person B. - Transactions that are approved by authorized person but not the same person as usual such as transaction of profit center A always approved by authorized person B Table 5.2: Examples of tests of each audit step in execution phase - 61 - Audit Processes Examples of Tests but there were a few cases that the transactions were approved by authorized person C. - Control testing - Applied techniques: By using clustering technique, grouping samples according to more limited of relevant attributes. - Examples of interesting patterns: - Range of transaction amount prepared by each responsible person. - Range of transaction amount approved by each authorized person. - Distribution of transaction amount in each profit center. - Distribution of transaction amount grouped by date. - Relationship between responsible person and authorized person. - Relationship between responsible person or authorized person and profit center. - Date distribution of transactions prepared by each responsible person or approved by each authorized person. - Date distribution of transactions of each profit center. - Integration between some patterns above. Table 5.2: Examples of tests of each audit step in execution phase (Continued) - 62 - Audit Processes Examples of Tests Analytical Review - Expectation development - Applied techniques: Predicting the figures of the following months based on the previous months using time-dependent analysis. However, the technique can be more effective if the other concerns; such as season, inflation rate, interest rate and industry index, are taken into account. - Examples of interesting patterns: - Expectation figures that have very stable pattern. - Expectation figures that have very unstable pattern. - Other general analysis - Applied techniques: Consider time variable, trying to cluster accounting transactions in each category (e.g. assets, liabilities, sales, expenses) differently. - Examples of interesting patterns: - Some small transactions that occurred repeatedly in the certain period of the month. - Same types of transactions that are recorded differently (e.g. recorded in the different account numbers). Table 5.2: Examples of tests of each audit step in execution phase (Continued) - 63 - Audit Processes Examples of Tests - Sales figure of some months that are considered excessively higher or lower than others. - Expenses that are extremely unstable during the year. - Repeatedly purchase of fixed assets. - Repeatedly re-financing transactions especially loan to and from related companies. Detailed Tests of Transactions - Sample selection - Applied techniques: Grouping all accounting transactions in each area with all relevant attributes by using clustering technique. It may be a good idea to include the results of both control testing and analytical review of each area. Examples of interesting patterns: - Transactions that do not belong to any clusters, or outliers. - Group of transactions that has large percentage of the population. - Group of transactions that has unusual relationship with the results of control testing and analytical review. Table 5.2: Examples of tests of each audit step in execution phase (Continued) - 64 - Audit Processes Examples of Tests - Applied techniques: By using clustering technique, grouping samples according to more limited of relevant attributes. Examples of interesting patterns: - Large amount of transactions that refer to the same document number. - Large amount of transactions that occur in the same date especially when it is not month-end or any pre-identified peak date. - Group of non-regular basis transactions that occurred repeatedly. An example is fixedassets purchasing transactions that occurred at the same amount every month-end date. - Time difference between document date and record date that is different from normal patterns. For example, during the second week of every month, time gap will be 5 days longer than normal. Detailed Tests of Balances - Sample selection - Applied techniques: Grouping all accounting transactions in each area with all relevant attributes by using clustering technique. It may be a good idea to include the results of both control testing and analytical review of each area. Table 5.2: Examples of tests of each audit step in execution phase (Continued) - 65 - Audit Processes Examples of Tests Examples of interesting patterns: - Transactions that do not belong to any clusters, or outliers. - Group of transactions that has large percentage of the population. - Group of transactions that has unusual relationship with the results of control testing and analytical review. - Applied techniques: By using clustering technique, grouping samples according to more limited of relevant attributes. Examples of interesting patterns: - “Cast at bank” ending balances of some banks that are different from others such as large amount of fixed balance though the company does not have any obligation agreements with the bank. - Customer that has many small balances of the same product. For example, insurance customer whose balance is comprised of many insurance policies bought repeatedly in 2 weeks. - Inter-company balances that pricing pattern are significant different from normal transactions. Table 5.2: Examples of tests of each audit step in execution phase (Continued) - 66 - Generally, it is more effective to analyze data in the aggregated form and allow iterating the test in more detail if needed. The reason is that, by running the test in aggregated manner, the exploration will be faster and more comprehensible in the sense that the number of results is in a manageable range. Besides, the over-fitting problems can be prevented. For example, the higher hierarchy of account numbers should be used in the first clustering analysis of detailed testing. Then, only interesting groups of accounts will be selected to further testing. The detailed account numbers can be used in this run. As noticed in the table above, sample selection might be the most promising step where data mining can be applied. The possibility is that by using data mining package to perform clustering analysis, auditors can select the samples from some representatives of the groups categorized in the way that they have not distinguished before and the obviously different transactions from normal ones or outliers. Then, further tests remain a matter of professional judgement. Auditors can consider using data mining packages to further the test or obtain the samples selected and test them with other GAS packages. From the discussion above, it came to a conclusion that, at present, data mining cannot be a substitution of GAS or other computerized tools currently used in auditing profession. However, it might be incorporate as an enhancement of GAS in some auditing steps and if it does make sense, the development and research in this field to make a customized data mining package for auditing can be anticipated. 5.6. Summary Recently, data mining has become an area of spotlight for auditing profession due to the cost pressure while auditing profession provided another promising market for data mining. Therefore, the integration between auditing knowledge and data analysis techniques is not far from the truth. As seen in Table 5.3, generalized audit software (GAS) and data mining packages have some different characteristics from the audit perspective. At present, GAS already has a market in audit profession. The capability to assist overall audit process with little technical skill required is the major reason for its success. However, - 67 - Characteristics GAS Data Mining Package Customized for audit works Yes No Support entire audit procedures Yes No User friendly More Less Require technical skill Less More Automated No Yes Capable of learning No Yes Lower Higher Cost Table 5.3: Comparison between GAS and data mining package characteristics GAS also has been criticized that it is only an old query tool with more efficient presentation features -- that it could make some tasks easier but it could not complete anything by itself. Data mining, on the other hand, promises automated works but is quite difficult to employ. However, data mining tool remain promising in a variety of application areas upon further research, improvement and refinement. If the appropriate data mining tools for the auditing profession are developed, it is expected to be able to replace some professional expertise required in some auditing processes. Though data mining seems to be feasible in almost all steps of audit procedures, the most practical and required one is the execution phase especially sample selection step. It can be done by mapping audit approaches, including tests of controls, analytical review and detailed tests, into data mining problems. - 68 - 6. Research Methodology 6.1. Objective and Structure The objective of this chapter is to provide the perspective of the actual study -empirical part. The research period is specified in section 6.2. The material of the study is discussed in section 6.3 and the research methods including the reason for the chosen study and the techniques used are provided in section 6.4. Next, section 6.5 specifies the chosen software used in the testing phase. Then, the methods of result interpretation and analysis are identified. Finally, all above-mentioned substances are summarized in section 6.6. 6.2. Research Period The research period covers a twelve-month period of accounting transaction archive started from January 2000. 6.3. Data Available For this thesis, the most appropriate data set is an accounting transaction archive. Since, though does not require, it makes more sense to apply data mining with vast amount of data sets so that the automation capability of data mining can be appreciated, the expected data set is relatively large. The data set used in the study was provided courtesy of PwC PricewaterhouseCoopers Oy (www.pwcglobal.com/fi/fin/main/home/index.html). To preserve the confidentiality of data, the data set was sanitized so that the proprietor of the data was kept anonymous and sensitive information was eliminated. Sensitive information such as account names, cost center names, account descriptions and structures and basic nature of transactions. Besides, the supporting documents that contained confidential information such as chart of account were not provided. According to PwC, the data set was captured from a client’s general ledger system. However, since the purpose of PwC research was to analyze the relationship between expenses and cost centers, only transactions relevant to the research were obtained. Although the data set does not represent complete general ledger, it is considered complete for cost center analysis. A more detailed account of this matter is - 69 - provided in section 7.4 -- Restrictions and Constraints. However, it is worth nothing that, due to the incomplete data set, missing information and limited supporting knowledge and information, the scope of test was limited to a few areas. This is somewhat different from a normal audit engagement where the information limitation is a serious matter that could prevent the auditors from rendering an opinion. The initial data set consists of accounting transactions taken from the general ledger system. It contains 494,704 tuples (or transactions), with forty-six attributes (or columns) where seven are calculation numeric attributes and only two are of explanation in nature. The list of columns is provided in Appendix A. Due to the limitation of available data, the area of research was the sample selection step of the “test of controls” phase. The detailed discussion of the research area was identified in section 7.3.2 -- Data Understanding. 6.4. Research Methods This study is focusing on the use of data mining techniques to assist audit work. The data set was tested using both data mining software and generalized audit software (GAS) in order to compare whether the data mining software can be used as an enhancement or a replacement of GAS. As stated in Chapter 4, data mining process consists of six phases: business understanding, data understanding, data preparation, modeling, evaluation, and deployment. However, the modeling phase, which also includes building a model for future use, is time-consuming. Due to the time constraint, data mining techniques are applied to the data so that the usefulness of such techniques can be evaluated without building the model. Besides, as the results of test have to be compared, interpret and analyzed based on the hypotheses in section 7.4 -- Results Interpretation, the evaluation phase is not included in the research process section. Finally, the deployment phase is considered out of scope of this thesis. In my opinion, data mining process can be applied to all software usage but the first three phases are especially remarkably valuable for the users to understand how to benefit from the software efficiently. However, for ready-to-use software packages such as GAS, modeling phase can be thought of as the deployment phase. To make the - 70 - process of both data mining and GAS compatible, the fourth phase, which is the last phase of research process, will be called software deployment. For all practical purpose, the first three phases were performed once and both GAS and data mining software packages were used. The prepared data was tested on GAS and data mining software, and the result was then evaluated and compared on its usability and interestingness. 6.5. Software Selection 6.5.1. Data Mining Software For data mining software, the DB2 Intelligent Miner for Data Version 6.1 was chosen. It was developed by IBM Corporation (www.ibm.com) and was distributed in September 1999. The distinction of the product is its capability to mine both relational data and flat file data. However, as implied by its name, this product works better with DB2’s relational data. Figure 6.1: IBM’s DB2 Intelligent Miner for Data Version 6.1 screenshot As shown in figure 6.1, the mining template of DB2 Intelligent Miner for Data supports six methods of data mining algorithms. The methods include association, classification, clustering, prediction and sequential patterns and similar sequences. The - 71 - underlying algorithms as well as other parameters and constraints can be chosen and specified for each mining session. Besides, the mining results can be viewed in many formats with the enhanced visualization tools. Another interesting feature of DB2 Intelligent Miner for Data is its interoperability. It can be integrated with other business intelligent products such as SPSS (Statistical Package for the Social Sciences) and operational information systems such as SAP. 6.5.2. Generalized Audit Software ACL (Audit Command Language), the most well-known generalized audit software, was selected as the representation of GAS. It was developed by ACL Service Ltd. (www.acl.com) and the version chosen was ACL for Windows Workbook Version 5 that was distributed in 1997. I would not go into details about ACL because almost all features were mentioned in section 3.3 -- Generalized Audit Software. However, it is worth noting that, besides the statistical and analytical functions, the preeminent feature of ACL is its automatic command log. It captures all activities including commands called by the users, messages and the results of the commands. 6.6. Analysis Methods The auditing results from both software packages were interpreted, analyzed and compared. The focus was on the assertions from the auditing point of view. The elements of comparison include the following: - Interestingness and relevance of the results - Time spent - Level of difficulty - Level of required technical knowledge - Level of automation - Risks and constraints - 72 - However, as auditing assertions are frequently linked to materiality and judgment, to base the analysis solely on a single opinion was considered not sufficient. Thus, to strengthen the interpretation of the test results and to avoid bias, the opinions, suggestions and comments on the above-mentions arguments from experienced auditors and competent professionals were also gathered. 6.7. Summary For the study, a data set provided by SVH PricewaterhouseCoopers Oy was tested with data mining software and generalized audit software (GAS). IBM’s DB2 Intelligent Miner for Data Version 6 was selected to represent data mining software while ACL for Windows Workbook Version 5 was chosen for GAS. Based on the data available, the study focused on sample selection for control testing. The data set was tested with DB2 Intelligent Miner for Data and ACL for Windows Workbook. The results of the tests were interpreted concerning relevance, interestingness, time spent, required knowledge, automated level, risk and constraints. The interpretation was also confirmed by a reasonable number of auditors and competent professionals. - 73 - 7. The Research 7.1. Objective and Structure This chapter aims to provide the information about the research. The hypothesis is described in section 7.2. Then, all facts about the work performed are documented in section 7.3. The results of the research are summarized in section 7.4 and interpreted in section 7.5. Finally, a brief conclusion of this actual study is discussed in section 7.6. 7.2. Hypothesis The hypotheses of this study are as follows: H1: Data mining will enhance computer assisted audit tools by automatically discovering interesting information from the data. H2: By using clustering techniques, data mining will find the more interesting groups of samples to be selected for the tests of controls comparing to sampling by using generalized audit software. 7.3. Research Process The research process for this thesis consists of four phases: they are business understanding, data understanding, data preparation, and software deployment. The first three phases were performed only once and the last phase was performed using both software packages. Details of each phase are as follows: 7.3.1. Business Understanding In auditing, the term “business” might be ambiguous as it can be thought of as the business of the proprietor of the data, the auditing, or both. Since the proprietor of the data is kept anonymous in this case, the focus is auditing. However, it is important to note that normally all requirements from both businesses should be taken into account. And once the business objective is set, it should be rigid so that it will be easier and more logical to perform the following phases. - 74 - The main purpose of this thesis is to find out whether data mining is useful to auditing profession. However, it is impossible to test all assertions since data and time are limited. Therefore, only one practical area of research where data is available is selected for testing. If this research shows that data mining can contribute something to auditing profession, then further research using more complete data may be anticipated. In this research, as the possible areas of tests rely on the data available and the detailed business objective cannot be identified until the data is studied; consequently, this phase has to be performed in parallel with the data understanding and data preparation phase. 7.3.2. Data Understanding In practice, data understanding and data preparation phases go hand in hand. In order to do the data understanding phase, the data needs to be transformed so that the software will accept the data as the input. On the other hand, the data has to be understood before performing the cleaning phase, otherwise some useful information might be ripped off or the noises might be left out. As a result, the processes of data understanding and data preparation have to be iterated. Going back and forth between these two phases allows user to revise the data-mining objective in order to meet business objective. As mentioned before, the business objective of this thesis is ambiguous because the selected test depends on the data available. Therefore, the business understanding phase is also included in all iteration. To make it simple, the details of the iteration are documented only once in section 7.3.3 -- Data Preparation. I chose to pre-study the data by using SQL (Structured Query Language) commands in DB2 Command Line Processor. The studied process is only to analyze the relevance and basic nature of attributes by querying each attribute in different ways. For example, all unique records of each attribute were queried so that the empty attributes (with null values) and the attributes that have same value in all records can be eliminated. Notice that, in ACL, these queries are embedded under user-friendly builtin functions. Therefore, the results of querying will be the same. - 75 - Another attention to be paid is that, in reality, the data can be studied much more extensively. After the preliminary analysis, the interesting matters will be further investigated. Such investigation includes aggregated analysis (analyzing the studied data by aggregating it with other relevant information), supporting material review and information querying from responsible person. However, this can not be done in this research due to the confidentiality constraint. Since the available data is the general ledger of a certain period, the possible areas of research were limited to those stated in table 5.2. However, due to data and time constraint, the research was restricted to only sample selection step of the control testing process. That is because the knowledge about the data available is insufficient which made the analysis process of control testing and analytical procedure more complicated, or even infeasible. In addition, the data cannot be tested in aggregated format because the structures of neither accounts nor cost centers were available. Therefore, the detailed test of transactions and detailed test of balances cannot be performed effectively either. In a normal audit engagement, the sample size of the control testing is varied depending on the control reliance. As the control reliance of this data cannot be identified, the sample size is set to fifty transactions, which are considered as a mediumsize sample set. 7.3.3. Data Preparation The objective of this step is to understand the nature and structure of the data in order to select what to be tested and how it should be tested. It includes three major steps, which are data transformation, attribute selection and choice of tests. Details of each step are as follows: 7.3.3.1. Data Transformation As the data file provided by SVH PricewaterhouseCoopers Oy (PwC) was in ACL format, it can be used with the ACL package directly although the version difference caused some problems. However, to test the data with Intelligent Miner for Data, such data had to be transformed into the format that can be placed in DB2 environment. Therefore, in this research, this step was only needed for data mining test. - 76 - Notice that IM can also mine data in normal text files. However, it is more complicated as it requires users to specified the length of record, the length of columns and the definition of each column. Besides, it is more convenient to study the data and to test the correctness of imported data in DB2 command processor because the commands are based on SQL. As it is out of the scope of this paper, the details of each transformation step will not be provided. In short, data was converted into text file with a certain format and then imported into DB2 database afterwards. However, it is worth nothing that this process was time consuming especially when the structure of the data is unclear. 7.3.3.2. Attribute Selection It is always better to scope the data set to only relevant attributes so that the data set can be reduced to more manageable size and the running time of the software is minimized. However, this step is not considered significant to ACL. This is because the sample selection algorithm of ACL is quite rigid and is not based on any particular attribute except for those specified by users. Therefore, this step is necessary for data mining test only. This step aims to understand the data better so that the choice of test can be specified. As mentioned in the data understanding section, the most appropriate test is the sample selection step of the control testing process given the data constraints. However, it is crucial to identify the relevant attributes as the selection criteria. The risk is the relevant ones might be eliminated which will affect the interestingness and the accuracy of the result. On the other hand, the remaining irrelevant attributes can detract the data structure. Many iterations of this step were performed. At the end of the iteration, the possible choices of test are reviewed regarding the remained data set. A brief summary of each iteration is documented below: a) First Iteration: Eliminating all trivial attributes -- This iteration includes the following: - 77 - - Eliminate all attributes that contained only null value -These include altogether seven attributes. - Eliminate all attributes that contained same value in all records -- These include altogether twelve attributes. - Eliminate all redundant attributes that are highly correlated with the more detailed ones (e.g., Month attribute can be replaced by Date attribute) -- These attributes will not add additional information and may distort the result. They include altogether six attributes. b) Second iteration: Eliminating irrelevant attributes that might cause noises or that the structure is unclear -- This iteration includes the following: - Eliminate attributes that mainly contained null or meaningless (e.g. “*****”) value -- This step can be considered risky as the exception might be outlier or interesting pattern. However, as the data cannot be analyzed in detail, to keep these attributes would contribute nothing but data structure detraction. Thus, the attributes that contained more than forty-percent empty value were removed. These include altogether six attributes. - Eliminate attributes with certain structure such as sequential numbers grouped by specific criteria -- These include altogether four attributes. At the end of this stage, only complete and non-deviating attributes were remained in the data set. However, the choice of tests from these eleven attributes could not be specified due to two reasons. First, although data mining software can handle many divergent attributes, the test result will be excessively incomprehensive. Second, some of the attributes that have unknown or unclear definition will reduce the - 78 - accuracy of the test result and make the result more complicated to analyze. c) Third iteration: Eliminating attributes that the structures are unclear which might depreciate the accuracy of the results -This is very risky because it depends on judgement. Therefore, after the selection, confirmation from responsible person at PwC was also obtained. This step includes the following: - Eliminate attributes that require aggregated analysis -- As mentioned earlier, attributes that are complicated such as having many unique values should be analyzed in aggregate format. Therefore, three attributes that require the understanding of structure were eliminated. - Eliminate attributes that were added up for internal analysis by the company whose data belongs to -- These include altogether three attributes. At the end of this phase, the scope of test was reduced to only six attributes. Those attributes include “Batch Error”, “Cost Center”, “Authorized Person”, “Document Number”, “Document Date” and “Amount (EUR)”. Aside from using DB2 Command Line Processor to analyze each attribute more extensively, a small clustering test (optimizing mining run for time) was also run in order to ensure the relevance and to preliminarily examine the relationship among attributes. However, the results of the test are, as expected, incomprehensible especially when further analysis can not be made. The example of the testing results is shown in figure 7.1. Therefore, the other iteration of this step was required. d) Fourth iteration: Eliminating attributes for the final test -According to PwC, this data set was taken from general - 79 - Figure 7.1: Result of neural clustering method with six input attributes. ledger system for the cost center analysis. However, after the preliminary test, it cannot be established that the data set is a representation of the complete general ledger; a limitation. that will be addressed in section 8.4 -- Restrictions and Constraints. Finally, the other three attributes were eliminated so that the final test will include only relevant attributes that maximize the efficiency of the test. Details of each eliminated attribute are as follows: - Cost Center -- As this analysis is important for PwC, it was not eliminated despite the structure is ambiguous. However, this attribute consisted of 748 unique values so it is difficult to analyze without an understanding of its hierarchical structure. - Document Number -- This is the only reference number at the transaction level of this data set. The first assumption is that it is journal number referring to the original - 80 - document. However, the sum of all the transaction balances of each document number is not zero. Besides, the transactions of each document number referred to a different Document Date. Thus, it was concluded that these are not journal numbers and it would be of no use to leave it in the data set. - Document Date -- As mentioned earlier, the document date of the transactions were different even when they referred to the same Document Number. Besides, the knowledge of this attribute was insufficient so further analysis regarding this attribute was not feasible. 7.3.3.3. Choice of Tests As mentioned in the data understanding subsection, the only possible testing area is sample selection step of the control testing. This step can be performed differently by grouping the transactions using different sets of attributes. However, since the knowledge of the data is insufficient, only three relevant attributes were remained in the data set. Therefore, it cannot be studied extensively. The focus of the test is how data mining will group the accounting transactions according to those three relevant attributes and whether the result brings out any interesting matters. Although the most appropriate data-mining method for this step is clustering, there is also a problem of its lacking of selfexplanation. In other words, the results of clustering functions are clusters that are automatically grouped by underlying algorithms of data mining method but the criteria of grouping are left not addressed. For the purpose of cross-validation, tree classification method was chosen to test because it shows how the rules are conducted. In conclusion, the mining methods chosen are clustering and classification. Descriptions of each method according to IBM’s Intelligent Miner for Data help manual are as follows: a) Clustering There are two options of clustering mining functions -demographic and neural. Demographic clustering - 81 - automatically determines the number of clusters to be generated. Similarities between records are determined by comparing their field values. The clusters are then defined so that Condorcet’s criterion is maximized. Condorcet’s criterion is the floating number between zero and one and is the sum of all record similarities of pairs in the same cluster minus the sum of all record similarities of pairs in different clusters (IBM, 2001a, 6). Put another way, the more Condorcet value, the more similar all records are in the same cluster. Similarly, neural clustering groups database records by similar characteristics. However, it employs a Kohonen Feature Map neural network. Kohonen Feature Map uses a process called self-organization to group similar input records together (IBM, 2001a, 6). Beside the underlying algorithms, there are two major differences between these two functions according to “Mining Your Own Business in Banking Using DB2 Intelligent Miner for Data” (IBM, 2001, 63). First, demographic clustering has been developed to work with non-continuous numeric (or categorical variables) while neural clustering techniques works best with variables with continuous numeric values and maps categorical values to numeric values. Second, For Neural clustering, users have to specify the number of clusters that they wish to derive while, with demographic clustering, the natural number of clusters is automatically created based on the use specifying how similar the records within the individual clusters should be. - 82 - b) Classification Two algorithms classification of and classification neural method classification. are The tree neural classification employs a back-propagation neural network, which is general-purpose, supervised-learning algorithm, to classify data. This simply means that the results will be as ambiguous as the clustering techniques. Since this method was just for gaining a deeper understanding of the structure of the records and cross-validation, only tree classification function was chosen. Tree classification utilizes historical data containing a set of values and a classification for these values. By analyzing the data that has already been classified, it reveals the characteristics that led to the previous classification (IBM, 2001a, 7) In sum, three mining functions, namely demographic clustering, neural clustering and tree classification, were chosen to complement each other and to validate the derived results. 7.3.4. Software Deployment This is the most critical step in this research and will provide the trail for further research. Therefore, more space will be devoted to explain each software deployment process. Details are as follows: 7.3.4.1. IBM’s DB2 Intelligent Miner for Data Before proceeding, it is important to note that the explanations in this subsection are mainly based on the following: - “Intelligent Miner for Data” (IBM, 2001a) - “Data Mining for Detecting Insider Trading in Stock Exchange with IMB DB2 Intelligent Miner for Data” (IBM, 2001b) - 83 - - “Mining Your Own Business in Banking Using DB2 Intelligent Miner for Data” (IBM, 2001c) - Online help manual of the software As a general note, each function is run more than once because there is no golden rule as to what are the most appropriate parameter values to be used. The parameters set for each run were updated and revised in order to make the result of the next run more comprehensible or to ensure that the current run is the most satisfactory. Only interesting graphical results are illustrated in the discussion below while all high-resolution graphical results and the descriptive results are provided in Appendix B. Three functions of DB2 Intelligent Miner for Data, namely the demographic clustering, neural clustering and tree classification were chosen to test. Details of each function are as follows: a) Demographic Clustering As mentioned above, the number of clusters to be generated is determined automatically. It finds the optimum combination of values that maximizes the similarity of all records within each cluster, while at the same time maximizes the dissimilarity between clusters it produces or, in other words, it tries to maximize the value of Condorcet (IBM, 2001a, 64). However, there are four parameters that users have to specify which the details are illustrated in table 7.1. For clustering result, the derived clusters are presented in separate horizontal strips ordered by size. The combination of each attribute is shown in each strip ordered by their important to the individual cluster (IBM, 2001c, 70). The categorical variable will be shown as pie chart. The inner circle shows the percentage of each variable value of - 84 - Parameter Maximum Passes Definition Default Value 2 The maximum number of times the function goes through the input data to perform the mining function. Maximum Clusters 9 The largest numbers of clusters the function generates. 2 Accuracy The minimum percentage of Improvement improvement on clustering quality after each pass of the data. It is used as the stopping criterion. If the actual improvement is less than the value specified, no more passes occur. Similarity Threshold 0.5 It limits the values accepted as best fit for a cluster. Table 7.1: Definitions and default values of demographic clustering parameters the cluster while the outer one shows the percentage of the same variable value but to the whole population. On the other hand, the numerical variable will be shown with histogram. The highlighted part is the distribution of the population while the line shows the distribution of the cluster. Notice that this result pattern also applies to neural clustering as well. In the first run of this method, the default values of all parameters were chosen. The graphical result of this run is shown in figure 7.2. As seen from the figure, eight clusters were derived from this run. The interesting matters are as follows: - The largest cluster (Cluster0) contained 493,846 transactions or 99.83% of the population. This simply - 85 - Figure 7.2: Graphical result of the first run of demographic clustering (Parameters values: 2, 9, 2, 0.5) means that almost all of the records follow the same pattern. It includes being not error and 86.37% of transaction amounts are between 0 and 50,000. - The second largest cluster (Cluster2) is where all error transactions are. It also shows that all of them did not have authorized person to approve the transactions. The Condorcet value of this cluster is 0.9442, which means all of them are almost identical and extremely dissimilar to other clusters. - Cluster3 and Cluster4 should have a close relationship because the distribution of authorized person of Cluster 3 and Cluster4 are almost the same. Besides, the distributions of the absolute transaction amount are almost equal. - 86 - - Cluster5 and Cluster7, which mainly contain the extremely high and low transactions, include only seven transactions. It is interesting that these small amounts of transactions were grouped separately from others. In other words, these two clusters can be thought of as outliers. The global Condorcet value at 0.6933 is considered satisfactory. However, for comparison purpose, another run was performed. Because it seems that the result is already detailed, in the second run, the maximum number of clusters was reduced to five while other parameter values remained the same. Except for the number of derived clusters, the result of the second run is almost as same as the first one. This is especially true of the exact Condorcet value and the distribution of the major clusters including Cluster0, Cluster2, Cluster3 and Cluster4. This means that neither the fewer nor the greater number of clusters provides better result in this case. However, as the result of the first run shows more detail, it will be used as a representative of demographic clustering for the comparison analysis. b) Neural Clustering For neural clustering, the users have to specify two parameters -- the maximum passes and the maximum clusters. The default value of those parameters are five and nine, respectively. Moreover, the input data are selected by default and normalization of the data is strongly recommended. The input data normalization means that the data will be scaled to a range of zero to one in case of continuous or discrete numeric field and converted into a numeric code in case of categorical field so that they can be presented to the neural network. - 87 - All default values were used in the first run of this method. The result is shown in Figure 7.3. Notice that the result of neural clustering is different from the result of demographic clustering. The reason why different clustering techniques produce different view of transaction records is that their functions are based on different algorithms. Normally, the choice of technique depends on user’s judgement; however, most of the time more than one techniques are used together for comparison purpose. Figure 7.3: Graphical result of the first run of neural clustering (Parameter values: 5, 9) From Figure 7.3, seven clusters were produced. The interesting patterns of this result are as follows: - The most interesting cluster is Cluster4, which contains 573 transactions or 0.12% of the population. It contains all error transactions which previously known from the preliminary analysis and the demographic clustering that they were not authorized. However, it also contains the - 88 - transactions authorized by the other person (LOUHITA) which might means that this person had approved a set of transactions that has the same pattern as the erroneous one. Thus, if it is the real case, this is the cluster where auditors should focus on. - The smallest cluster, Cluster1, which contains only twenty-eight transactions and some were not authorized. In addition, the major transaction amounts of this group are extremely low. For comparison, two other runs with the maximum clusters of four and sixteen were produced. For the third run, the result is considered too detailed and does not show anything more specific than the first one. In the second run, the error transactions were grouped with the majority of the records (64.86%). However, the most interesting cluster is the cluster that contains ninety-five transactions including those not authorized but also not recorded as errors. However, the result is relatively similar to the Cluster1 of the first run. Therefore, only the result of the first run is selected as the representative of neural clustering result. c) Tree Classification As mentioned above, tree classification technique was chosen for cross-validation purpose. The advantage of this technique is its rule explanation. For DB2 Intelligent Miner for Data, the binary decision tree visualization is built as an output of the tree classification. Like other classification techniques, tree classification requires training and testing data set to build the model, or classifier. However, in this test, all records were used as the training data set. That is because the objective of this test is - 89 - not to build a model. Instead, it is to find the decision path where the model is built on. The only parameter that user has to specify when running tree classification test is the maximum tree depth value. It is the maximum number of levels to be created for the binary decision tree. The default value for maximum tree depth is unlimited. However, the choice of data fields has to be specified as well. The relevant attributes are chosen from available fields as input fields and class labels. The class label represents the particular pre-identified classification based on the input fields. In this case, the Batch Error attribute was used as the class label, while the other two attributes were specified as input fields. Once more, the default value of maximum tree depth was used in the first run. Clearly, the result of the first run with sixty-seven tree-depth levels is incomprehensible. Although there is a feature of the Intelligent Miner for Data that allows pruning the tree, it only allows pruning down the groups that have their population less than the certain amount but not the other way around. By doing so, small clusters that normally are more interesting in terms of different patterns will be pruned down. Therefore, with the same choice of data fields, the maximum tree depth of the second run was set at ten in order to reduce the level of the tree while remaining all records to be grouped. Figure 7.4 shows the binary tree of the classification result. The most interesting nodes are the second, the fourth and the fifth leaf nodes because the majority of their population are error transactions. An example of the interpretation is that, for the second leaf node, the transaction whose value is - 90 - Figure 7.4: Graphical result of the second run of tree classification (Maximum tree depth: 10) greater than –21,199.91 but less than –8,865.80 has an error possibility of 83.3%. However, the tree path begins with the null value of authorized person attribute, which simply means that all authorized transactions were not include in the tree path. Therefore, only amount attribute was taken into consideration. Due to the fact that the value of transaction amount is varied by its nature, this result does not contribute to that interesting pattern. Finally, the third run of tree classification was done. In this run, the maximum tree depth remained the same but the Authorized Person attribute was chosen as the predefined class label and the Batch Error attribute was switched to the input field list. However, the result came out with 59.69% error rate, which is not considered as satisfactory. - 91 - In conclusion, clustering techniques fit the objective of this research more than classification techniques. The first run of both demographic clustering and neural clustering were chosen as representatives of data mining result to compare with the result of ACL in section 7.4 -- Results Comparison. 7.3.4.2. ACL As mentioned earlier, ACL software is customized especially for audit work. Therefore, the sampling feature, which is illustrated in Figure 7.5, is provided for sample selection step. Notice that the discussion below is mainly based on the help manual of the software. Figure 7.5: Sampling feature of ACL The sampling functions of ACL are quite rigid. Users have to first specify the sample type, either monetary unit sample (MUS) or record sample. As the name suggests, MUS function biases higher value items. Simply put, a transaction with higher value has higher possibility to be selected than a transaction with lower value. It is useful for detailed testing because it provides a high level of assurance that all material items in the population are subject to testing. Notice that the population for MUS is the absolute value of the field being sampled. On the other hand, the record sample is unbiased which simply means that each record has an equal chance of being selected and the transactional value is not taken into account. Record sampling is most useful for control or compliance testing where the rate of errors is more important than the monetary unit. Therefore, the record sampling is selected to test in this research. - 92 - Next, sample parameter has to be chosen to specify the sample type or the specific method to be used to draw the samples. The methods include fixed interval, cell and random methods. The brief explanations of each method are as follows: - Fixed internal sample: An interval value and a random start number must be specified. It implies that the sample set will consist of the start number record and every item at the interval value order thereafter. - Cell sample: The population is broken into groups the size of the interval and one random item is chosen from each group based on the random seed specified. Therefore, this method is also known as a random interval sample. - Random Sample: The size of the sample, a random seed value and the size of the population have to be specified. The process is that ACL will generate the required number of random numbers between one and the population specified based on the random seed and then the selection will be made using the prescribed random numbers. To put it simplistically, by using ACL to generate samples, auditors have to either have sampling criteria in mind or to let the software choose for them randomly. This might be very effective when auditors have the clear understanding of the data they have. On the other hand, if the structure of the data is ambiguous, only monetary unit sampling and random sampling can be used and, thus, the interesting transactions may be completely skipped. As the focus of this research is the test of controls and the structure of the available data is extremely unclear, random record sampling was selected. However, notice that the samples can also be selected differently based on the preliminary analysis from the data understanding step. follows: Those possibilities are as - 93 - - Randomly selected from the whole population: By using this strategy, every transaction has an equal chance to be selected regardless of the error or the authorized person. - Randomly selected from all records except for those that are recorded as errors. In this case, all of the non-error transactions have the same possibility to be selected. However, the potential of interesting pattern discovery is uncertain and depends mainly on the experience and skill of the auditor. - Randomly selected from the transactions that were not authorized. erroneous According to the preliminary analysis, all transactions are unauthorized transactions. Therefore, the chance that potentially inaccurate transactions will be selected is assumed to be similar to the chance that the error transactions will be selected. - Randomly selected from the transactions that were not authorized and were not recorded as errors. In this case, it might be a chance that the inaccurate records, which were not recorded as errors, will be selected. From my point of view, the only possible benefit from examining error records is that the patterns of errors can be specified. However, it is not considered an efficient way especially when the certain number of error transactions is large. On the other hand, if a small sample size is selected from a large population, it is difficult to find any correlation among those samples. Therefore, the last option was chosen for testing so that the samples are selected from the most potential inaccurate transactions. Fifty samples were selected from a population of 258 based on 100 random numbers. Details of the sample are provided in Appendix C. distribution of the transaction amounts of the sample is shown in Figure 7.6. The - 94 - Figure 7.6: The transaction amount distribution of ACL samples Notice that, from the fifty samples, three transactions have relatively high amount and only one has very low amount. Otherwise, the transactions are not that different from the average amount of –72.67. 7.4. Result Interpretations Before proceeding further, it might be worth reviewing some of the interesting matters of the result. The details are as follows: 7.4.1. IBM’s DB2 Intelligent Miner for Data From demographic clustering (the first run), the most interesting clusters are Cluster5 and Cluster7 where only four and three transactions are identified, respectively. These two clusters share the same pattern, which is a set of small number of transactions with the extremely high absolute values. These transactions can be considered as the outlier because there are only a few of them and they were grouped separately from the other transactions. From neural clustering (the first run), there are two outstanding clusters which are Cluster4 and Cluster1. Cluster4 is comprised of 457 error transactions and 116 transactions authorized by an authorized person, “LOUHITA”. Although the - 95 - reason why these transactions were grouped together is not provided and is subject to further research, it is still irrefutably interesting. The other cluster, Cluster1, is consisted of twenty-eight transactions which some were not authorized. Besides, the amount of transactions in this cluster is relatively low. As the size of this cluster is only 0.01%, it can be considered as outlier as well. The fifty samples selected based on mining clustering, in my opinion, would be comprised of the following: a) Seven transactions from Cluster5 and Cluster7 of the demographic clustering result. b) Twenty-eight transactions from Cluster1 of the neural clustering result. c) Five randomly selected transactions from 116 transactions of Cluster4 of the neural clustering result. d) One transaction from each cluster other than the above except for Cluster2 of the demographic clustering result which contains all error transactions. However, there is a chance that some records in a) and b) might be the same. One of each double sample should be eliminated so that it will be counted only once. In this case, the substituting samples should be selected from c). Unfortunately, the transactions in this data set do not have identification information. Therefore, there is no chance to know whether there are any double samples until the further investigation is conducted. 7.4.2. ACL The sample set derived by ACL is consisted of fifty transactions randomly selected from the transactions that were not authorized but also not recorded as errors. It includes four transactions with outstanding absolute values and fifty-six indifferent small ones. - 96 - Based on the results from both DB2 Intelligent Miner for data and ACL, the comparison between sample selection is illustrated in table 7.2. Issue Intelligent Miner for Data ACL Interestingness The number of samples can be It is fair to say that the and Relevance varied from each cluster. The samples selected by ACL are more focus can be put on the relatively ordinary. The more interesting ones but at chance to find interesting least one sample is selected transactions is low. Besides, from each group of the although the test is population. However, the satisfactory, it cannot verify criteria of clustering are in the correctness of all question and, thus, auditor’ population, not even at the judgement is still required to group level. Thus, the preliminarily analyze the decision of sample size is interestingness and the really important in order to relevance of the clusters. ensure that the samples are, to some extent, good representative of the population. Time Spent Excluding time spent in learning Running the sampling how to use the software, the function takes only a few running time for data mining minutes for ACL. Besides, test takes only a few minutes for only sample size and seed each test. However, in order to value must be specified. find the optimum parameters for Therefore, it does not require each type of test, many much time to spend on iterations have to be run and that decision process. requires extra time for analysis and decision process. Table 7.2: Comparison between results of IBM’s DB2 Intelligent Miner for Data and ACL - 97 - Issue Intelligent Miner for Data ACL Required Technical knowledge about The features of ACL are Technical database and understanding somewhat similar to general- Knowledge about data mining are required purpose software packages. in all data mining processes Thus, it does not require especially in data understanding much efforts to learn how to phase. Besides, for auditors that use. do not have the technical background, it is still not easy to learn how to use data mining software despite it is much more user-friendly at present. Level of Although the most important As the features of ACL are Automation and the most difficult process -- not that flexible, auditors clustering -- is performed need to only specify a few automatically, auditors are still parameter values in order to required to measure whether the run the test. Otherwise, all clustering result is interesting tasks are done automatically. and whether the other run should be performed. Besides, the choice of samples from each clusters has to be identified by the auditors as well. Risk and As the reason why each cluster The sample size is very Constraints is grouped together is important in random record ambiguous, to measure whether sampling. If the sample size the clustering result is is too low, the chance of interesting is a matter of finding interesting matter is judgement. However, as the low. On the other hand, if the Table 7.2: Comparison between results of IBM’s DB2 Intelligent Miner for Data and ACL (Continued) - 98 - Issue Intelligent Miner for Data ACL samples are selected from each sample size is too large, it is cluster that has some similar not feasible to perform the characteristics. It provides more test. Besides, the population assurance that the sample set is where samples to be selected a better representative of the is also significant and population. required auditors’ judgement. Table 7.2: Comparison between results of IBM’s DB2 Intelligent Miner for Data and ACL (Continued) From the discussion above, one may concludes that the samples selected by clustering function of Intelligent Miner for Data is more interesting than the samples selected by ACL. However, using Intelligent Miner for Data requires more technical knowledge than using ACL. In my opinion, the required technical knowledge for using Intelligent Miner for Data is tolerable especially when the result is satisfactory. Comments and suggestions about the results were also solicited from five auditors at different experience levels. For confidentiality reason, however, their names are kept anonymous. All of them agreed that the result from Intelligent Miner for Data is more interesting and, if the mining result is available, they will certainly choose the sample selected from it. Nevertheless, the result frequently is not specific enough and additional research is still required. Two out of five said that the required technical knowledge makes data mining less appealing. If, and only if, data mining software is more customized for audit work, would they consider using it. Finally, professional judgement is not an issue because it is required in any case. In conclusion, based on the result of this research, there are some interesting patterns discovered automatically by data mining technique that cannot be found by generalized audit software. This finding confirms the H2 hypothesis. However, it is worth nothing that it does not prove whether those interesting patterns are material matters. Although data mining techniques are able to draw out potentially interesting information from the data set, it is, however, not realistic to say that data mining is an advance computer assisted auditing tools due to its deployment difficulties and the - 99 - ambiguous results. Therefore, H1 hypothesis is partly confirmed and is subject to further research. 7.5. Summary The hypotheses of this thesis are that clustering techniques of data mining will find the more interesting groups of sample than those sampled by using generalized audit software and, thus, data mining may considered as an enhancement of the computer assisted audit tools. The research process consists of business understanding, data understanding, data preparation, and software deployment phases. Many iterations of the first three phases were performed in order to understand the data better and to determine what to study. Finally, the choice of test is the sample selection step of the control testing process concentrating on authorized persons, batch errors and the transaction balances. For methods of data mining technique, demographic clustering, neural clustering and tree classification were chosen. The prepared data was applied with the IBM’s DB2 Intelligent Miner for Data and ACL as inputs. While the sampling choice of Intelligent Miner for Data is based on auditors’ judgement regarding the derived clusters, the sample set selected by ACL is automatically generated. Auditors’ judgement is indispensable in determining which cluster is more interesting and should be focused on. Based on the results of this research, the conclusion is that the result of Intelligent Miner for Data is more interesting than the result of ACL. Although greater technical knowledge is required by Intelligent Miner for Data, it is in the acceptable level. However, the automated capability of data mining cannot be fully appreciated, as the auditors’ judgements are still required to interpret the results. On the other hand, ACL is much easier to use but the quality of result can be compromised. A comparison of the results derived from both software packages based on my opinion is summarized in table 7.3. It should be reminded, however, that all analysis is based on certain unavoidable assumptions about the data and further investigation is required to confirm these interpretations. - 100 - Intelligent Miner for Data ACL Higher Lower Slightly Higher Slightly Lower Required Technical Knowledge Higher Lower Level of Automation Lower Higher Risk and Constraint Lower Higher Issue Interestingness and Relevance Time Spent Table 7.3: Summary of comparison between sample selection result of Intelligent Miner for Data and ACL The consensus of auditors’ comments on these results is that the results of data mining clustering techniques are more interesting than the results of ACL although it requires high level of auditors’ judgement. However, they also agree that it still requires further research to determine whether the sample set from Intelligent Miner for data is a better option than the sample set randomly selected by ACL. Moreover, they feel more comfortable working with ACL due to its well-customized features and userfriendly interface. In sum, it is unarguably that, within the scope of this research, the result of data mining is more interesting than the result of normal generalized audit software even when the data was incomplete and the supporting knowledge and information was limited. However, the conclusion that data mining is a superior computer assisted audit tool cannot be made until more extensive researches are conducted. - 101 - 8. Conclusion 8.1. Objective and Structure This chapter will attempt to provide an overall perspective of this thesis including a brief summary of the whole study in section 8.2. The results and implication are discussed in section 8.3, restrictions and constraints of the study in section 8.4, and suggestions for further research in section 8.5. A final conclusion is given in section 8.6. 8.2. Research Perspective This master thesis aims to find out whether data mining is a promising tool for the auditing profession. Due to the dramatically increased volume of accounting and business data and the increased complexity of the business, auditors can no longer rely solely on their instincts and professional judgments. Lately, auditors have realized that technology, especially intelligent software, is more than just a required tool. New tools including generalized auditing software have been adopted by the audit profession. Another relatively new filed that has received greater attention from all businesses is data mining. It has been applied to use with fraud detection, forensic accounting and security evaluation in other business applications related to auditing. The allure of data mining software is its automated capability. However, although data mining has been around for more than a decade, the integration between data mining and auditing is still esoteric. The biggest cost of auditing is professional staff expense. That is why the employment of data mining seems to make good sense in this profession. In this thesis, the ideal opportunities that data mining can be integrated with audit work are explored. However, due to the restrictions and limitation of the available data for research, the test can not be done extensively. The only area of testing is sample selection step of the test of control process. The data provided by SVH PricewaterhouseCoopers Oy was studied with both data mining software (IBM’s DB2 Intelligent Miner for Data) and generalized audit software (ACL) and the results from both studies were compared and analyzed. - 102 - In general, samples selection can be done differently depending on what to focus. In this thesis, the focus is relationship between authorized persons, errors and transaction amounts. With Intelligent Miner for Data, demographic clustering and neural clustering functions were selected to draw out the relationship patterns among the data. The results were analyzed and the choice of sample was based on that analysis. For ACL, the random record samples were automatically selected based on the random number and the size of the sample set specified. The population that the samples were selected from was a group of records that has the most potential to be the error ones. 8.3. Implications of the Results Based on the feedback from the five auditors and my observations, it can be concluded that, within the scope of this research, the results derived from data mining techniques are more interesting than those derived from generalized audit software. However, this conclusion is predicated on certain unavoidable assumptions of the data set and, thus, is not conclusive. Further investigation of whether data mining result is more interesting is necessary but, unfortunately, due to the data limitation and time constraint it cannot be done in this research. However, it is by no mean the results of the generalized audit software are considered not useful. If the size of the transaction archive to be audited is not that massive and the relationship between those transactions is unambiguous, employing generalized audit software is easier to use and is not a bad idea at all. In sum, the hypothesis that clustering techniques of data mining will find the more interesting groups of sample to be selected for the test of controls comparing to sampling by generalized audit software is confirmed. However, the other hypothesis, which is that data mining will enhance computer assisted audit tools by automatically discovering interesting information from the data is still unclear. The determination whether the information derived from data mining techniques is really interesting cannot be made until the further investigation is performed. Besides, the auditors’ judgement is still a prerequisite so the level of automation is not fully appreciated. - 103 - 8.4. Restrictions and Constraints The restrictions and constraints of this research are as follows: 8.4.1. Data Limitation Although the available data was taken from general ledger system, it was not a complete collection of the general ledger transactions. The limitations include incomplete data, missing information, and limited understanding of the data. Details are as follows: 8.4.1.1. Incomplete data In the data understanding phase, the objective of which is to better understand the nature of the data in general, the data set was found out that it was not a complete general ledger. That is because the sum amount of all transactions (plus sign for debit amounts and minus sign for credit amount) does not zero out. Besides, neither the sum amount of each recorded date, nor of each document number is zero. Therefore, the assumption is that this is a partial listing of the general ledger, which is considered complete for cost center analysis. Unquestionably, for data mining that aims at finding hidden interesting pattern, it is much better when the data is complete. However, finding the accounting transactions at the reasonably large size is not simple either. Moreover, for test of controls in general, it is not that critical to have all the accounting transactions as the sampling population; this is especially true when auditors have some selection criteria in mind and the extended test is allowed. In summary, this data set is fairly satisfactory for a pilot test such as this research but more comprehensive data is required for more detailed study. 8.4.1.2. Missing information As mentioned earlier, all sensitive information such as account names were eliminated from the data set. Besides, the supporting information for aggregated analysis -- such as chart of accounts, cost center structure or transaction mapping -- is not available. Without knowing what is what, the analysis is extremely - 104 - difficult, if not impossible. Therefore, the scrutinized testing and analytical review is no longer feasible for this research. For detailed testing, it is normal to select samples for testing according to the audit areas or, in other words, account groups. Therefore, the chart of accounts, or at least the knowledge of data structure, is remarkably important. This situation limited the scope of test to only tests of controls. 8.4.1.3. Limited Understanding In the normal audit engagement, one of the most necessary requirements is the ability to further investigate when the matter arises. Generally, it can be done by reviewing supporting document and interviewing responsible person. If the result of further investigation is unsatisfactory, the scope of test may be expanded later on as well. However, further investigation cannot be done in this research. The one and only resource data is a limited version of data file obtained from SVH PricewaterhouseCoopers Oy. Although the analysis process includes the opinions gathered from many auditors and competent persons, they are based on assumptions and judgements. 8.4.2. Limited Knowledge of Software Packages Although both chosen software packages in this research are well customized and user-friendly, there is a chance that some useful features might be overlooked due to the limited knowledge about the software. However, since this research is just a pilot test of this subject, self-studies that included reading manuals and inquiring of competent persons is considered sufficient. Nevertheless, it is worth noting that when an auditing firm decided to study, use or implement a software package, it is a good idea to educate the responsible personnel by experts of such software. This includes training the real users, providing real-time support and sharing the discovered knowledge among team members. - 105 - 8.4.3. Time Constraints This might be the most important constraint of this research. The more time spent, the more extensive the test can be. Further testing with this limited data and additional knowledge may be obtained through trial and errors until the more interesting matter is found. However, in my opinion, it does not consider this as an intelligent strategy and that it will not contribute to more satisfactory results comparing to time spent. Therefore, the test was restricted to the most promising one and leave the rest to future research when more complete data is available. 8.5. Suggestions for Further Research As mentioned above, the research regarding the integration of data mining and auditing can be done extensively especially when the complete data is available. The examples are the possible areas of integration in table 5.1 and the examples of tests that can be performed in the execution phase when only general ledger transactions of the current year are available in table 5.2. However, it is important to note that it is far more feasible and efficient to work on the complete data set. This includes the privilege to scrutinize all relevant supporting information and the permission to perform more extensive investigation if necessary. 8.6. Summary Though the integration between data mining techniques and audit processes is a relatively new field, data mining is considered useful and helps reducing cost pressure in many business applications related to auditing. Therefore, this thesis aims to explore the possibility of the integration between data mining and the actual audit engagement processes. However, due to the data and other limitations, the study could not be done extensively. Only sample selection step of the test of controls was studied. From the result of this research, it does show that data mining techniques might be able to contribute something to this profession even when the data was not that complete and all the analyses were based on assumptions. However, it also does not prove that data mining is a good fit for every audit work. It requires a substantial effort - 106 - to learn how to employ data mining techniques and to understand the implication of the results. However, if the auditing firms have vast quantities of data to be audited and the auditors are familiar with the nature of transactions and expected error patterns, then data mining does provide an efficient means to surface interesting matters. However, it is still a long way from possessing “Artificial Intelligence” to fully automate the audit testing for the auditors. - 107 - List of Figures Figure 2.1: Summary of audit engagement processes 15 Figure 3.1: ACL software screenshot (Version 5.0 Workbook) 19 Figure 4.1: Four level breakdown of the CRISP-DM data mining methodology 26 Figure 4.2: Example of association rule 33 Figure 4.3: A decision tree classifying transactions into five groups 38 Figure 4.4: A neural network with two hidden layers 39 Figure 5.1: Basic structure of general ledger 59 Figure 6.1: IBM’s DB2 Intelligent Miner for Data Version 6.1 Screenshot 70 Figure 7.1: Results of neural clustering method with six input attributes 79 Figure 7.2: Graphical result of the first run of demographic clustering (Parameter value 2, 9, 2, 0.5) 85 Figure 7.3: Graphical result of the first run of neural clustering (Parameter value 5, 9) 87 Figure 7.4: Graphical result of the second run of tree classification (Parameter value 2, 9 ,2 ,0.5) 90 Figure 7.5: Sampling feature of ACL 91 Figure 7.6: The transaction amount distribution of ACL samples 94 - 108 - List of Tables Table 3.1: ACL features used in assisting each step of audit processes 20 Table 4.1: Summarization of appropriate data mining techniques of each data mining method 41 Table 5.1: Possible areas of data mining and audit processes integration 49 Table 5.2: Examples of tests of each audit step in execution phase 60 Table 5.3: Comparison between GAS and data mining package characteristics 67 Table 7.1: Definitions and defaults values of demographic clustering parameters 84 Table 7.2: Comparison between results of IBM’s DB2 Intelligent Miner for Data and ACL 96 Table 7.3: Summary of comparison between sample selection result of Intelligent Miner for Data and ACL 100 - 109 - References a) Books and Journals American Institute of Certified Public Accountants (AICPA) (1983), Statement on Auditing Standards (SAS) No. 47: Audit Risk and Materiality in Conducting an Audit. American Institute of Certified Public Accountants (AICPA) (1988), Statement on Auditing Standards (SAS) No. 56: Analytical Procedures. Arens, Alvin A. & Loebbecke, James K. (2000), Auditing: An Integrated Approach, New Jersey: Prentice-Hall. Bagranoff, Nancy A. & Vendrzyk, Valaria P. (2000), The Changing Role of IS Audit Among the Big Five US-Based Accounting Firms, Information Systems Control Journal: Volume 5, 2000, 33-37. Berry, Michael J. A. & Linoff, Gordon S. (2000), Mastering Data Mining, New York: John Wiley & Sons Inc. Berson, Alex, Smith, Stephen & Kurt, Thearling (2000), Building Data Mining Applications for CRM, McGraw-Hill Companies Inc. Bodnar, George H. & Hopwood, William S. (2001), Accounting Information Systems, New Jersey: Prentice-Hall. Committee of Sponsoring Organizations (COSO) (1992), Internal Control Integrated Framework. Connolly, Thomas M., Begg, Carolyn E. & Strachan, Anne D. (1999), Database Systems – A Practical Approach to Design, Implementation, and Management, Addison Wesley Longman Limited. Cross Industry Standard Process for Data Mining (CRISP-DM) (2000), CRISP-DM 1.0 Step-by-Step Data Mining Guide, www.crisp-dm.org/. - 110 - Gargano, Michael L. & Raggad, Bel G. (1999), Data Mining – A Powerful Information Creating Tool, OCLC Systems & Services, Volume 15, Number 2, 1999, 81-90. Glover, Steven, Prawitt, Douglas & Romney Marshall (1999), Software Showcase, The Internal Auditor, Volume 56, Issue 4, August 1999, 49- 56. Hall, James A. (2000), Information Systems Auditing and Assurance, SouthWestern College Publishing. Han, Jiawei & Kamber, Micheline (2000), Data Mining: Concepts and Techniques, San Francisco: Morgan Kaufmann Publisher. Hand, David, Heikki, Mannila & Smyth, Padhraic (2001), Principles of Data Mining, MIT Press. IBM Corporation (2001a) Intelligent Miner for Data - Data Mining, IBM Corporation. IBM Corporation (2001b) Data Mining for Detecting Insider Trading in Stock Exchange with IMB DB2 Intelligent Miner for Data, IBM Corporation. IBM Corporation (2001c), Mining Your Own Business in Banking Using DB2 Intelligent Miner for Data, IBM Corporation. Lee, Sang Jun & Keng, Siau (2001), A Review of Data Mining Techniques, Industrial Management & Data Systems: Volume 101, Number 01, 2001, 44-46. Ma, Catherine, Chou, David C. & Yen, David C. (2000), Data Warehousing, Technology Assessment and Management, Industrial Management & Data Systems: Volume 100, Number 3, 2000, 125-135. McFadden, Fred R., Hoffer, Jeffrey A. & Prescott, Mary B. (1999), Modern Data Management, Addison-Wesley Educational Publisher Inc. Moscove, Stephen A., Simkin, Mark G & Bagranoff, Nancy A. (2000), Core Concept of Accounting Information System, New York: John Wiley & Sons Inc. - 111 - Needleman, Ted (2001), Audit Tools, The Practical Accountant: March 2001, 3840. Rezaee, Zabihollah, Elam, Rick & Shabatoghlie, Ahmad (2001), Continuous Auditing: The Audit of the Future, Managerial Auditing Journal: Volume 16, Number 3, 2001, 150-158. Rud, Olivia Parr (2001), Data Mining Cookbook, New York: John Wiley & Sons Inc. - 112 - b) Web Pages ACL Service Limited (2002) ACL for Windows, www.acl.com/en/softwa/softwa_aclwin.asp (Accessed on January 4, 2002) Audimation Services Inc. (2002) IDEA – Setting The Standard in Case of Use, www.audimation.com/idea.html (Accessed on January 4, 2002) DB Miner Technology Inc. (2002) DBMiner Insight – The Next Generation of Business Intelligence, www.dbminer.com/products/index.html (Accessed on February 2, 2002) Eurotek Communication Limited (2002) How To Choose a PC Auditing Tool, www. Eurotek.co.uk/howchoose.htm (Accessed on March 12, 2002) IBM Corporation (2002) DB2 Intelligent Miner for Data, www- 3.ibm.com/software/data/iminer/fordata/ (Accessed on February 5, 2002) Microsoft Corporation (2002) Microsoft Data Analyzer – The Office Analysis Solution, www.microsoft.com/office/dataanalyzer/ (Accessed on February 2, 2002) SAS Institute Inc. (2002) Uncover Gems of Information – Enterprise Miner, www.sas.com/products/miner/index.html (Accessed on March 12, 2002) SAS Institute Inc. (2002) SAS Analytic Intelligence, www.sas.com/technologies/analytical_intelligence/index.html (Accessed on March 12, 2002) SPSS Inc. – Business Intelligence Department (2002) Effectively Guide Your Organization’s Future with Data Mining, www.spss.com (Accessed on February 21, 2002) - 113 - Appendix A: List of Columns of Data Available No. Original Name Translated Name 1. AS_RYHMA_GLM Customer Group 2. CHARACTER1 Character 3. EROTIN1 Seperator1 4. EROTIN3 Seperator3 5. KAUSI_TUN Period ID 6. KLO_AIKA Period Date 7. LAJI_TUN Type ID 8. NIPPU_JNO Batch Queue 9. NIPPU_JNO_VIRHE Batch Error 10. NIPPU_KAUSI_TUN Batch Period 11. NIPPU_KAUSI_TYYPPI Batch Technical Number 12. NIPPU_KIRJ_PVM Batch Date 13. NIPPU_MLK_1 Batch Point 1 14. NIPPU_MLK_M_JNO Batch Point Queue 15. NIPPU_MLK_T_JNO Batch Point Queue 16. NIPPU_MLK_TUN Batch Point ID 17. NIPPU_TEKN_NRO Batch Technical Number 18. NIPPU_TUN_GLM Batch ID 19. NIPPU_VALKAS_KDI Batch Currency 20. ORG_TUN_LINJA_GLM Foreign Cost Center 21. ORG_TUN_MATR_GLM Cost Center 22. PAIVITYS_PVM Transaction Date 23. SELV_TILI_TUNNUS Authorized Person 24. TILI_NO Account Number 25. TILI_NO_AN Reconciliation Account Number 26. TOSITE_KASPVM Document Date - 114 - No. Original Name Translated Name 27. TOSITE_NO Document Number 28. TUNNUS ID 29. TUN_KUMPP_GLM Partner ID 30. VAL_KURSSI Exchange Rate 31. VAL_KURSSI2 Exchange Rate 2 32. VIENTI_JNO Entry Queue 33. VIENTI_KASTILA_GLM Status 34. VIENTI_M_VAL_DES_1 Currency Amount Point 35. VIENTI_MAARA_ALKP Original Amount (FIM) 36. VIENTI_MAARA_M Amount (EUR) 37. VIENTI_MAARA_M_DK Debit / Credit 38. VIENTI_MAARA_T Amount 39. VIENTI_MAARA_VAL_1 Currency Amount 1 40. VIENTI_MAARA_VAL_2 Currency Amount 2 41. VIENTI_SELITE Explanation 42. VIENTI_SELITE2 Explanation2 43. VIENTI_VALPAIV_KE Center 44. YHTIO_TUN_GLM Company Code 45. YHTIO_TUN_GLM_AN Company Reconciliation Code 46. YHTIO_TUN_KUMPP Inter-company Code - 115 - Appendix B: Results of IBM’s Intelligent Miner for Data a) Preliminary Neural Clustering (with Six Attributes) - 116 - User Specified Parameters Mining Run Outputs Maximum Number of Passes: 5 Number of Passes Performed: 5 Maximum Number of Clusters: 9 Number of Clusters: 9 Deviation: 0.732074 Cluster Characteristics: Id Cluster Size Id Absolute % 0 129267 26.13 1 51018 2 Cluster Size Absolute % 5 264 0.05 10.31 6 43665 8.83 79163 16.00 7 31327 6.33 3 11592 2.34 8 142422 28.79 4 5987 1.21 Reference Field Characteristics (For All Field Types): (Field Types: ( ) = Supplementary CA = Categorical CO = Continuous Numeric Id Name DN = Discrete Numeric) Modal Frequency (%) Modal Value Type No. of Possible Values/ Buckets 1 NIPPU_JNO_VIRHE CA 0 99.91 2 2 ORG_TUN_MATR_GLM CA 7989 3.31 748 3 SELV_TILI_TUNNUS CA AUT.HYV 35.12 18 4 TOSITE_KASPVM CA 2000-02-29 3.41 303 5 TOSITE_NO CO 250 91.14 15 6 VIENTI_MAARA_M CO 50000 86.37 12 Reference Field Characteristics (For Numeric Fields Only): Id Name 5 TOSITE_NO 6 VIENTI_MAARA Minimum Value Maximum Value Mean Standard Deviation 1 23785 867.926 3283.79 -5.64084E7 5.171E7 9.72619 271013 - 117 - b) Demographic Clustering: First Run - 118 - User Specified Parameters Mining Run Outputs Maximum Number of Passes: 2 Maximum Number of Clusters: 9 Improvement Over Last Pass: 2 Similarity Threshold: 0.5 Number of Passes Performed: 2 Number of Clusters: 8 Improvement Over Last Pass: 0 Global Condorcet Value: 0.6933 Cluster Characteristics: Cluster Size Absolute % Condorcet Value 0 493846 99.83 0.6933 1 73 0.01 2 457 3 174 Id Id Cluster Size Condorcet Absolute % 4 117 0.02 0.6916 0.7988 5 4 0.00 0.9337 0.09 0.9442 6 31 0.01 0.7918 0.04 0.6765 7 3 0.00 0.8095 Similarity Between Clusters: Similarity Filters: 0.25 Cluster 1 Cluster 2 Similarity Cluster 1 Cluster 2 Similarity 0 1 0.46 1 7 0.39 0 2 0.44 3 4 0.48 0 3 0.41 3 5 0.43 0 4 0.39 3 6 0.43 0 5 0.34 3 7 0.36 0 6 0.34 4 5 0.36 0 7 0.35 4 6 0.43 1 3 0.42 4 7 0.47 1 4 0.43 5 6 0.48 1 5 0.44 5 7 0.56 1 6 0.42 6 7 0.45 - 119 - Reference Field Characteristics (For All Field Types): Id Name Type Modal Value Modal Frequency (%) No. of Possible Values / Buckets Condorcet Value 1 NIPPU_JNO_VIRHE CA 0 99.91 2 0.9982 2 SELV_TILI_TUNNUS CA AUT. HYV 35.21 18 0.1911 3 VIENTI_MAARA_M CO 50000 86.37 12 0.8871 Reference Field Characteristics (For Numeric Fields Only): Id 3 Name Minimum Value VIENTI_MAARA_M -5.64084E7 Maximum Value 5.171E7 Mean 9.72619 Standard Deviation Distance Unit 271013 135506.453 - 120 - c) Demographic Clustering: Second Run - 121 - Mining Run Outputs User Specified Parameters Number of Passes Performed: 2 Maximum Number of Clusters: 5 Number of Clusters: 5 Improvement Over Last Pass: 2 Improvement Over Last Pass: 0 Similarity Threshold: 0.5 Global Condorcet Value: 0.6933 Maximum Number of Passes: 2 Cluster Characteristics: Cluster Size Absolute % Condorcet Value 0 493846 99.83 0.6933 1 91 0.02 0.6803 2 457 0.09 0.9442 Id Id Cluster Size Condorcet Absolute % 3 178 0.04 0.6706 4 133 0.03 0.6519 Similarity Between Clusters (Similarity Filters: 0.25) Cluster 1 Cluster 2 Similarity Cluster 1 Cluster 2 Similarity 0 1 0.44 1 3 0.43 0 2 0.44 1 4 0.42 0 3 0.41 3 4 0.47 0 4 0.38 Reference Field Characteristics (For All Field Types): Id Name Modal Frequency (%) Modal Value Type No. of Possible Values / Buckets Condorcet Value 1 NIPPU_JNO_VIRHE CA 0 99.91 2 0.9982 2 SELV_TILI_TUNNUS CA AUT.HYV 35.21 18 0.1911 3 VIENTI_MAARA_M CO 50000 86.37 12 0.8871 Reference Field Characteristics (For Numeric Fields Only): Id Name Minimum Value 3 VIENTI_MAARA_M -5.64084E7 Maximum Value 5.171E7 Mean 9.72619 Standard Deviation Distance Unit 271013 135506.453 - 122 - d) Neural Clustering: First Run - 123 - User Specified Parameters Mining Run Outputs Maximum Number of Passes: 5 Maximum Number of Clusters: 9 Number of Passes Performed: 5 Number of Clusters: 7 Deviation: 0.0980459 Cluster Characteristics: Id Cluster Size Absolute Id % Cluster Size Absolute % 0 162528 32.85 5 79773 16.13 1 54463 11.01 6 69742 14.10 2 11221 2.27 8 19814 4.01 4 9421 1.90 Reference Field Characteristics (For All Field Types): Id Name Type Modal Value Modal Frequency(%) No. of Possible Values/ Buckets 1 NIPPU_JNO_VIRHE CA 0 99.91 2 2 SELV_TILI_TUNNUS CA AUT.HYV 35.21 18 3 VIENTI_MAARA_M CO 50000 86.37 12 Reference Field Characteristics (For Numeric Fields Only): Id Name Minimum Value 3 VIENTI_MAARA_M -5.64084E7 Maximum Value 5.171E7 Mean 9.72619 Standard Deviation 271013 - 124 - e) Neural Clustering: Second Run - 125 - User Specified Parameters Mining Run Outputs Maximum Number of Passes: 5 Number of Passes Performed: 5 Maximum Number of Clusters: 4 Number of Clusters: 3 Deviation: 0.272368 Cluster Characteristics: Id Cluster Size Absolute Id % Cluster Size Absolute 0 173749 35.12 1 95 0.02 3 320861 % 64.86 Reference Field Characteristics (For All Field Types): Id Name Type Modal Value Modal Frequency (%) No. of Possible Values/ Buckets 1 NIPPU_JNO_VIRHE CA 0 99.91 2 2 SELV_TILI_TUNNUS CA AUT.HYV 35.21 18 3 VIENTI_MAARA_M CO 50000 86.37 12 Reference Field Characteristics (For Numeric Fields Only): Id Name Minimum Value 3 VIENTI_MAARA_M -5.64084E7 Maximum Value 5.171E7 Mean 9.72619 Standard Deviation 271013 - 126 - f) Neural Clustering: Third Run - 127 - User Specified Parameters Mining Run Outputs Maximum Number of Passes: 5 Number of Passes Performed: Maximum Number of Clusters: 16 5 Number of Clusters: 9 Deviation: 0.0380213 Cluster Characteristics: Id Cluster Size Absolute Id % Cluster Size Absolute % 0 162528 32.85 11 79773 16.13 3 54463 11.01 12 69742 14.10 5 11221 2.27 14 19814 4.01 8 9421 1.90 15 74804 15.12 10 12939 2.62 Reference Field Characteristics (For All Field Types): Id Name Type Modal Value Modal Frequency(%) No. of Possible Values/Buckets 1 NIPPU_JNO_VIRHE CA 0 99.91 2 2 SELV_TILI_TUNNUS CA AUT.HYV 35.21 18 3 VIENTI_MAARA_M CO 50000 86.37 12 Reference Field Characteristics (For Numeric Fields Only): Id Name Minimum Value 3 VIENTI_MAARA_M -5.64084E7 Maximum Value 5.171E7 Mean 9.72619 Standard Deviation 271013 - 128 - g) Tree Classification: First Run - 129 - Internal Node Class Records Errors Purity 0 0 494705 457 99.9 0.0 1 715 258 63.9 0.0.0 0 137 32 76.6 0.0.1 1 578 153 73.5 0.0.1.0 1 337 38 88.7 0.0.1.1 1 241 115 52.3 0.0.1.1.0 1 238 112 52.9 0.0.1.1.0.0 0 51 22 56.9 0.0.1.1.0.1 1 187 83 55.6 0.0.1.1.0.1.0 1 36 10 72.2 0.0.1.1.0.1.1 1 151 73 51.7 0.0.1.1.0.1.1.0 0 5 0 100.0 0.0.1.1.0.1.1.1 1 146 68 53.4 0.0.1.1.0.1.1.1.0 1 70 28 60.0 0.0.1.1.0.1.1.1.1 0 76 36 52.6 0.0.1.1.0.1.1.1.1.0 0 20 6 70.0 0.0.1.1.0.1.1.1.1.1 1 56 26 53.6 0.0.1.1.0.1.1.1.1.1.0 1 34 13 61.8 0.0.1.1.0.1.1.1.1.1.1 0 22 9 59.1 0.0.1.1.1 0 3 0 100.0 0.1 0 493990 0 100.0 - 130 - h) Tree Classification: Second Run - 131 - Internal Node Class Records Errors Purity 0 0 494705 457 99.9 0.0 1 715 258 63.9 0.0.0 0 137 32 76.6 0.0.0.0 0 11 0 100.0 0.0.0.1 0 126 32 74.6 0.0.0.1.0 1 6 1 83.3 0.0.0.1.1 0 120 27 77.5 0.0.1 1 578 153 73.5 0.0.1.0 1 337 38 88.7 0.0.1.0.0 1 213 11 94.8 0.0.1.0.1 1 124 27 78.2 0.0.1.1 1 241 115 52.3 0.0.1.1.0 1 238 112 52.9 0.0.1.1.1.0 0 3 0 100.0 0.1 0 493990 0 100.0 - 132 - i) Tree Classification: Third Run - 133 - Internal Node Class Records Errors Purity 0 AUT.HYV 493990 320241 35.2 0.0 LEHTIIR 32586 23651 27.4 0.0.0 SILFVMI 30353 32+89 27.3 0.0.0.0 LEHTIIR 12599 6963 44.7 0.0.0.0.0 SILFVMI 713 379 46.8 0.0.0.0.1 LEHTIIR 11886 6308 46.9 0.0.0.1 SILFVMI 17643 12643 28.3 0.0.1 LEHTIIR 2344 279 88.1 0.0.1.0 LEHTIIR 2308 243 89.5 0.0.1.1 KYYKOPI 36 18 50.0 0.1 AUT.HYV 461404 290884 37.0 0.1.0 AUT.HYV 435505 267403 38.6 0.1.0.0 AUT.HYV 29330 9723 66.8 0.1.0.1 AUT.HYV 406175 257680 36.6 0.1.0.1.0 LINDRHA 153161 112878 26.3 0.1.0.1.1.1 AUT.HYV 253014 138822 45.1 0.1.1 LEHTIIR 25899 14121 45.5 0.1.1.0 LEHTIIR 25031 13378 46.6 0.1.1.1 SILFVMI 868 452 47.9 - 134 - Appendix C: Samples Selection Result of ACL Sample Transaction Transaction Number Number Amount 1 008 -104.28 2 011 -916.43 3 014 -660.14 4 024 639.11 5 026 -1248.53 6 029 -4030.2 7 030 -2047.32 8 039 -1091.54 9 042 -2799.32 10 050 -2565.03 11 056 -1442.37 12 063 -5886.58 13 065 -127.02 14 068 -1492.67 15 069 -660.14 16 086 318.92 17 088 67.82 18 089 -479.34 19 090 4860.15 20 101 58.02 21 116 97.13 22 117 2123.52 23 120 26859.97 24 121 1081.77 25 126 13543.16 Sample Transaction Transaction Number Number Amount 26 137 185 27 154 6179.24 28 156 435.43 29 159 43.76 30 162 48.49 31 165 1795.71 32 167 94.64 33 168 -253.8 34 174 -85.08 35 176 30.9 36 178 -325.68 37 181 -28.56 38 192 325.68 39 193 33.6 40 197 960.23 41 198 -36389.95 42 199 14555.98 43 200 3638.99 44 212 -243.2 45 228 1173.36 46 231 -205.84 47 237 41.16 48 239 -652.72 49 242 -238.79 50 250 9.18