The Home Depot Paint Department Analysis GEORGIA TECH SENIOR DESIGN FINAL REPORT: FALL 2013 Team Members Bobby Faulk Bryce Ferguson Michael Gilkenson Jing Mei Ho Drew Keller Lauren Kley Melanie Ostis Silvana Vivanco Rfaulk6@gatech.edu Bferguson31@gatech.edu Mgilkenson3@gatech.edu Jho35@gatech.edu Dkeller7@gatech.edu Lkley3@gatech.edu Mostis3@gatech.edu Svivanco6@gatech.edu Advisor Dr. Steve Hackman Shackman@isye.gatech.edu Client Contact Will Welch William_S_Welch@homedepot.com Cory German Cory_German@homedepot.com Jennifer Smith Jennifer_L_Smith@homedepot.com This document has been created in the framework of a student design project. The Georgia Institute of Technology does not officially sanction its content. Executive Summary The Home Depot’s (THD) paint department (D24) is a major driver of store revenue, and its success relies on superior customer service. The Home Depot’s Store Support Center (SSC) believes that increased profits can be realized through an improvement in the customer experience. During data collection at three Atlanta-area locations, the senior design team observed an average of 3% of customers reneging and 18% of customers waiting more than 2.5 minutes, supporting the SSC’s belief. To supplement current staffing tools, the senior design team has developed a Paint Department Analysis Tool (PDAT). This tool allows the SSC to analyze different staffing scenarios and their effects on profit and customer service metrics. Through analysis of PDAT outputs, the senior design team has developed a number of service protocol recommendations which are outlined below: Associate Roles – The senior design team suggests that associates be assigned to either the desk or the aisle. Through several protocols, associates’ role assignments will change to reflect customer traffic and needs. Desk Queue Threshold – When the desk queue reaches a specific threshold, aisle associates will come to the desk to assist with mixing paint. Additional Accessory Sales – Accessories are the main drivers of profit in D24. By staffing associates in the aisle, they have the opportunity to suggest additional accessory purchases to customers. By applying these recommendations in a simulation model, the average percent of customers reneging decreased from 3.3% to 3.1%, while the average percent of customers waiting more than 2.5 minutes for service decreased from 17% to 14%. Additionally, the number of additional accessories sold to customers increased, leading to an average incremental daily profit due to accessory sales of $13.53. This incremental daily profit translates to roughly 5 additional accessories sold per day per store. Staffing recommendations suggested by the tool led to an average reduction of approximately 1.65 labor hours per day. By following these recommendations and using the PDAT, THD will realize an annual incremental profit of $10.4 million (pre-tax) from recovered lost sales and increased revenue on existing transactions. This conservative estimate does not account for labor changes. A more optimistic scenario will provide incremental profit of $29.1 million (pre-tax), but this can only be achieved if shifts are perfectly in sync with labor modification suggestions. More realistically, the value of this project lies somewhere within that range depending on how closely THD can match their staffing schedules to PDAT suggestions. i Table of Contents 1. Business Overview ............................................................................................................. 1 1.1. Paint Department Overview................................................................................................. 1 1.2. Current Staffing Process ....................................................................................................... 1 2. Project Scope ..................................................................................................................... 2 2.1. Problem Overview ................................................................................................................ 2 2.2. Solution Overview ................................................................................................................ 2 3. Design Approach................................................................................................................ 4 3.1. Data Analysis and Modeling ................................................................................................. 4 3.2. Customer Arrival Model ....................................................................................................... 5 3.3. Customer Renege Model ...................................................................................................... 6 3.4. Customer Profile Model ....................................................................................................... 7 3.5. Associate Role Model ........................................................................................................... 7 3.6. Desk Queue Length Threshold Model .................................................................................. 8 3.7. Accessory Sales Model ....................................................................................................... 10 3.8. Staffing Adjustments Model............................................................................................... 10 4. Deliverables..................................................................................................................... 12 4.1. Paint Department Analysis Tool ......................................................................................... 12 4.2. Simulation........................................................................................................................... 13 5. Recommendations and Valuation .................................................................................... 15 5.1. Recommendations ............................................................................................................. 15 5.2. Valuation ............................................................................................................................ 15 6. Appendices ...................................................................................................................... 18 6.1. Appendix A: Customer and Mixer Data Collection Procedure ........................................... 18 6.2. Appendix B: Mixer Utilization Data Analysis ...................................................................... 19 6.3. Appendix C: Customer Arrival Regression.......................................................................... 21 6.4. Appendix D: Customer Renege Model ............................................................................... 22 6.5. Appendix E: Recommendation and Valuation Data ........................................................... 23 6.6. Appendix F: PDAT User Guide ............................................................................................ 24 6.7. Appendix G: Associate Task Statistical Distributions ......................................................... 27 6.8. Appendix H: Simulation Flow Detail ................................................................................... 28 6.9. Appendix I: Simulation Validation Hypothesis Testing ...................................................... 30 ii 1. Business Overview 1.1. Paint Department Overview The Home Depot’s paint department sells paint and accessories that help in the painting process. Associates serve customers by processing paint orders at the desk, answering project questions, and selling accessories in the aisle. D24 contributes to 20% of all THD transactions and $5.9 billion in annual revenue. Accessory sales, with an average profit margin of 55%, have the most impact on department profit, while the margin for paint is only 30%. In addition to enhancing department revenue, it is imperative to provide an excellent customer experience, as it is one of THD’s core values. 1.2. Current Staffing Process The Home Depot uses the Forecasting and Scheduling Tool (FaST) to allocate labor two weeks in advance. FaST compares the previous four weeks of transaction data against the previous year’s data to forecast the upcoming number of customer transactions. Each THD location is unique due to the types of customers and service needs specific to that store. However, FaST cannot reflect the intricacies of individual store activity. FaST assigns the same number of labor hours to D24 for each transaction. A customer that only buys a paint brush and a customer that orders a five gallon bucket of paint will be allocated the same amount of labor, even though these customers clearly have different service needs. 1 2. Project Scope 2.1. Problem Overview The Home Depot has an opportunity to improve customer service in D24. They have had difficulty finding objective ways to quantify their level of customer service, so the senior design team looked into two THD-defined key metrics: The percentage of customers who wait for service at the paint desk longer than 2.5 minutes The percentage of customers who renege, or leave without service The Home Depot would like for these values to be as close to zero as possible. The senior design team collected customer service data at three Atlanta THD stores to evaluate these metrics (see Appendix A). As shown in Table 1, D24 is failing to realize its customer service goals. Table 1: This table shows metrics shown by customer data collected by the senior design team at the Austell, Buckhead, and Midtown stores over 8 days. Number of Observed Customers % Customers Waiting % Customers % Customers that (Purchasing Paint) > 2.5 Minutes Waiting > 5 Minutes Renege 706 18% 7% 3% The biggest opportunity to improve service is to increase THD’s visibility into D24. They have little data about the specific customers that shop in the department or the customer service they provide to those customers. For example, in a recent attempt to improve D24 customer service, the SSC added one full-time employee to a quarter of all stores, costing $12.5 million. The assignment of additional labor was not based on data or specific scenarios, and the project was unsuccessful. The Home Depot’s insufficient staffing policies are perpetuated by this lack of department visibility. As described in section 1.2, THD does not differentiate between customers when assigning labor, which prevents their staffing levels from appropriately meeting customer needs. While investigating long wait times and customer reneges, the senior design team addressed both staffing policies and potential bottlenecks caused by paint mixing machines. Analysis suggests that the current utilization of paint mixers in the stores does not cause a bottleneck (Appendix B), suggesting that these metrics result from staffing policies. 2.2. Solution Overview To improve customer service in D24, the senior design team explored various approaches to allow THD to staff the department to closely match customer needs. The team defined a method for modeling a store’s customers to understand when they are shopping, why they might leave 2 before receiving service, and what their service needs are. The team then defined associate protocols, focusing on how defining associate roles, responding to the number of customers waiting for desk service, and selling of additional accessories in the aisle can impact the customer experience. These solutions address not only customer service, but also revenue enhancement opportunities. When a customer reneges or an associate misses the chance to sell additional accessories, potential revenue is lost. By improving customer service, THD can drive customers sales to increase store revenue. The senior design team defined recommendations for improving customer service and department profit and also designed a Paint Department Analysis Tool (PDAT) that will give THD’s Labor Engineering and Allocation Team visibility into the effects of daily staffing decisions in D24. This tool will help the corporate team to understand how changes in staffing levels or service protocols will impact the customer experience prior to implementing those changes in stores. 3 3. Design Approach 3.1.Data Analysis and Modeling Using in-store data collection results, transaction and associate schedule information, and associate time study data, the senior design team developed a number of models to support the project solution, all of which are incorporated in an Arena simulation of the D24 customer service processes. The simulation includes an understanding of how much time each customer service task requires, who a store’s customers are, and how adjusting associate protocols can best meet customer needs. The senior design team used 300 hours of associate time study data provided by THD to define 107 unique customer service tasks in the paint department. The team then fit these tasks to statistical distributions to model the amount of time required by each task. The simulation model is based on these customer service tasks. The next element to modeling D24 is understanding who is shopping in the department. The senior design team modeled customer arrival times, customer renege decisions, and customer profiles (Table 2). Table 2: This table describes the methodology used to develop customer models. Model Data Used Customer Arrival Model In-Store Data Collection Transaction Timestamps Customer Renege Model In-Store Data Collection Customer Profile Model Itemized Transaction Data Methodology Used Regression Logistic Regression Analysis of Transaction Data Based on this understanding of the customer, the model then evaluates the effects of associate protocols on the customer experience, including changes in defining associate roles, responding to the number of customers waiting in the desk queue, and selling accessories in the aisles (Table 3). Table 3: This table describes the methodology used to develop associate protocol models. Model Data Used Methodology Used Associate Role Model Number of Staffed Associates Simulation Desk Queue Length Threshold Model Simulated Customer Service Simulation Accessory Sales Model Kurt Salmon Retail Study Simulation Finally, the model allows the user to evaluate the impact of increased or decreased labor on the customer experience by making changes to the number of scheduled associates throughout a given day. 4 3.2.Customer Arrival Model In order to staff D24 properly, THD needs to understand when customers are arriving to the paint desk. Currently, the only data they have is point-of-sale transaction timestamps, but this data does not directly indicate when customers arrive to the desk. Using transaction timestamps provided by THD and customer arrival times collected in Atlantaarea stores, the senior design team developed a regression equation to model arrival times from transaction data. Figure 1 shows the cumulative D24 transactions and arrivals on October 10, 2013 at the Midtown location. Arrivals precede transactions by approximately 30 minutes. Figure 1: This graph shows an example of arrivals that have been regressed from transaction data. Using linear regression as explained in Appendix C, the number of arrivals in a fifteen-minute interval t can be modeled from the number of transactions occurring in intervals t+1 and t+2. To estimate the arrival time for one customer based on his transaction timestamp, the regression coefficients were scaled to determine the likelihood that a given customer arrives in either one or two fifteen-minute intervals prior to their transaction time. The arrivals are assumed to be uniform in each time interval (Table 4). Table 4: This table shows how long customers tend to stay in the store between their arrival to D24 and their transaction timestamp. Distribution of Time Between D24 Arrival and Likelihood of this Distribution Transaction Timestamp (Minutes) 49% Uniform(15,30) 51% Uniform(30,45) 5 3.3.Customer Renege Model In addition to knowing when customers arrive to D24, THD must also understand under what conditions a customer might leave prior to receiving service. During data collection, the senior design team observed that a higher percentage of arriving customers reneged from the paint desk when there were more customers ahead of them in the system. A logistic regression was used to determine the likelihood that a customer reneges. Figure 2 shows the observed likelihood that a customer reneged during data collection (in gray) and the expected likelihood that a customer reneges (in orange) according to regression results. Figure 2: When there are more customers ahead of an arriving customer, the arriving customer is more likely to renege. The gray points represent the observed probability that an arriving customer reneges, while the orange line shows the expected renege probabilities. When collecting data, the senior design team ensured that no customer who was counted as having reneged returned on the same day. Because there is no way to accurately estimate how many of these customers returned on a later date to make the purchase, the renege model assumes that no reneging customers return. Appendix D shows further calculations regarding this model. 6 3.4.Customer Profile Model Because different types of customers have unique service needs, the senior design team defined customer profile types that are derived from transaction data. Table 5 shows the four types of D24 customers that have been identified with approval from THD. Table 5: This table describes the breakdown of customer profiles and the characteristics of each group. Customer Profile Description Definition Service Requirements Professional Tend to buy large Purchase at least 3 Require desk service quantities of paint paint items or at Do not require aisle least 5 gallons of Have significant product service paint knowledge Purchase a “contractor pack” accessory Do-It-Yourself (DIY) Have little knowledge of Purchase both paint Require desk and aisle With Accessories paint products and what and accessories service kinds of accessories their Do not meet the May agree to increase projects may require professional purchase size when Need help with paint requirements associates suggest selection additional accessories Most important factor to them is customer service1 Do-It-Yourself (DIY) Similar to “DIY With Only purchase Only require desk Without Accessories Accessories” customers paint service Do not intend to purchase Do not meet the May be willing to acccessories (possibly due professional purchase accessories in to lack of product requirements addition to paint knowledge) Aisle Only interested in Only purchase Only require aisle purchasing accessories accessories service Do not affect the paint Do not meet the desk queue professional requirements 3.5.Associate Role Model Currently, THD does not staff specific associate roles within D24. This sometimes results in associates waiting for customers to approach them rather than looking for meaningful customer interaction. The senior design team recommends that THD defines two roles: desk and aisle associates. These roles will serve as customer service home bases for associates. Desk associates will mix paint and fulfill customer orders. Aisle associates will answer customers’ questions and 1 According to a study done by J.D. Power and Associates, 80% of home improvement customers request help when looking for a product. Additionally, 27% of customers rate store staff as the most influential factor to their satisfaction, more than any other factor. 7 sell paint accessories. Defining and implementing these two roles will help associates to provide service that reflects changes in customer traffic throughout the day. Every fifteen minutes, the simulation assigns these roles by allocating the appropriate number of associates to work at the desk and then assigning all remaining staffed associates to the aisle. The number of associates who will be staffed at the desk is equal to the maximum number of desk associates who were helping customers at one time during the previous fifteen-minute period. Also, there must always be at least one associate at the desk and the number of desk associates can never exceed a store-defined value. Throughout the fifteen-minute interval, associates move from the aisle to the desk if the number of customers in the desk queue reaches a certain threshold. Realistically, THD stores will not redefine associate roles every fifteen minutes. By understanding the recommended desk queue threshold and how to react to changes in customer traffic, store associates will be able to mirror the logic that is modeled in the simulation. 3.6.Desk Queue Length Threshold Model The desk queue threshold is the maximum number of customers waiting in the desk queue before an aisle associate will move to the desk to assist with customer service. The senior design team simulated customer service in 12 stores over multiple months (Appendix E), comparing the baseline state (with no associate role definition or desk queue threshold) against scenarios with desk queue thresholds of 1, 2, 3, and 4. As shown in Figures 3 and 4, the percentage of customers waiting more than 2.5 minutes for service or deciding to renege was lowest when the desk queue threshold was equal to 1. This shows that THD is best meeting their customer service goals when implementing a desk queue threshold of 1. Figure 3: This figure compares the percentages of customers waiting > 2.5 min for each desk queue threshold scenario. 8 Figure 4: This table compares the percentages of customers reneging for each desk queue threshold scenario. However, there is a tradeoff between customer service and profit. Figure 5 shows the daily, average incremental profit above the baseline state. This profit is realized from reduced reneges and increased accessory sales. When the desk queue threshold is higher, more associates are able to work in the aisle rather than moving to the desk. Therefore, they are able to sell more accessories, increasing profit. Figure 5: This table shows the average daily change in profit for each desk queue threshold scenario as compared to the baseline, current state. Due to THD’s commitment to providing high customer service, the senior design team recommends that they define associate roles and implement a desk queue threshold of 1 despite the lower incremental profit. 9 3.7.Accessory Sales Model By staffing associates in the aisle, THD can significantly increase the number of accessory sales, which have the highest impact on department profit. According to a Kurt Salmon retail study, 75% of customers will increase their purchase size (upsell) by 25% when an associate encourages them to buy additional items. The accessory sales model calculates the number of customers that will agree to upsell service from an aisle associate. All DIY customers may be encouraged to purchase additional items, but the senior design team believes that “DIY without accessories” customers will be the most likely to purchase additional items. Desk associates should encourage customers to visit the aisles while their paint orders are being fulfilled. Then, aisle associates will approach customers to offer service and suggest additional items. In the simulation model, when a “DIY without accessories” customer receives customer service from an aisle associate, there is a 75% chance that the customer will increase their purchase size by 25% of the average “DIY without accessories” purchase price on the same day in that store. 3.8.Staffing Adjustments Model To provide customers with a more consistent customer experience, the senior design group analyzed the current staffing levels throughout the day in D24. After running the simulation for the baseline scenario, the customer service metrics were analyzed to strategically determine where the number of staffed associates should be increased or decreased to increase upselling, reduce reneging, and reduce unnecessary labor. Figure 6 gives an example of over- and understaffed times. Figure 6: This table shows the average understaffed and overstaffed times over 4 fiscal months for the 20 THD stores that were run to establish the best practice recommendations. 10 When running the comparison state in the PDAT, the model suggests labor changes from the baseline state, according to the following logic: Add an associate if: No one is staffed to work in the 15 minute interval. There are more than 5 “DIY without accessories” customers in the 15 minute interval, no one is assigned to work in the aisle, and no customers have agreed to upsell. There is a renege in the 15 minute interval, no one is working in the aisle, and there are currently fewer associates working than the specified maximum associates at desk protocol. Remove an associate if: The number of associates in the aisle is greater than the number of customers who agreed to upsell. 11 4. Deliverables 4.1.Paint Department Analysis Tool The Paint Department Analysis Tool provides the SSC Labor Allocation Team with visibility into daily paint department operations, giving them the ability to fundamentally change the way they staff associates and interact with customers. For a baseline scenario analysis, the user inputs past transaction data and the corresponding associate schedule. The simulation model then outputs the following expected customer service metrics and associate information in fifteen-minute intervals: The number of customers that wait more than 2.5 minutes for service The number of customers that renege Associate role assignments Associate idle time Machine utilization rates The user also has the option to adjust service protocols and the number of staffed associates for a given day. After making these changes, the user runs the alternative scenario that compares results from the updated service and labor selections against the results given in the baseline scenario analysis. This comparison analysis uses the previously defined service protocols: Associate role assignments (Section 3.5) Desk queue length threshold (Section 3.6) Additional accessory sales (Section 3.7) The PDAT then shows the user how the selections made in the alternative scenario affected customer reneges, wait times, the number of customers who agree to purchase additional accessories, and the incremental change in D24 profit. Table 6 shows an example of PDAT results on July 22, 2013 at the Plano, TX location. Table 6: This table outlines the outputs given by the PDAT for both the baseline and comparison states. Metric Baseline State Comparison State 5 4 # Customers Waiting > 2.5 Minutes 2 1 # Customers Reneging 0 7 # Customers Agreeing to Upsell $57.66 Incremental Profit – 12 By defining associate roles, setting the maximum desk queue threshold to 1, adding upsell logic, and removing one associate from 8:30am to 11am, this comparison state created $57.66 in incremental profit. See Appendix F for more in depth instructions on using the PDAT. 4.2.Simulation The PDAT’s output comes from an Arena simulation that models D24 customer service processes. The model contains 107 distinct customer service tasks, each of which uses a statistical distribution to determine service time allotted to that task. The senior design team fit these distributions from over 300 hours of associate time study data provided by THD. Appendix G shows a selection of these distributions. Figure 7 depicts the general flow of the simulation, where each block represents a submodel in the customer service process. Figure 7: This flowchart shows an overview of the simulation model. Customer orders are processed through the model based on information provided by the PDAT. The simulation model reads customer arrivals, associate schedules, and resource capacities from the PDAT. Throughout the day, the number of staffed associates is updated to reflect the number of staffed associates. For a more detailed explanation of the customer service process within the simulation, see Appendix H. This model focuses solely on customer service tasks in D24. Other tasks are also relevant, like packdown (restocking) and cleaning, but THD prioritizes customer service. To ensure that associates are staffed with appropriate time to complete these less time-sensitive tasks, the simulation model also tracks and reports the amount of idle time that associates have throughout the day. The Home Depot sets standards for how much time should be spent doing non-customer service tasks. Therefore, tracking associate idle time will help THD to determine whether or not the number of staffed associates allows for non-customer service tasks to be completed effectively. To validate the simulation model, the senior design team compared simulation output to the data that the team collected throughout the course of the project. The FaST baseline simulation model was used for validation because it reflects current state protocols (i.e. it does not include associate role definition or require that the desk is never left unattended). Existing transaction 13 and associate schedule data were used to run the PDAT for each day to ensure that renege rates and wait times were representative of what was observed during data collection. For each day that the senior design team collected data, the simulation was run 10 times to get an average representation of simulation output. The average results are shown in Table 7. Table 7: This table shows the validation results from comparing observed customer service metrics against simulation service metrics for 8 days. Metric Observed Simulated Difference 2% 3% 1% % Customers that Renege 18% 24% 6% % Customers Waiting > 2.5 Minutes 7% 8% 1% % Customers Waiting > 5 Minutes The senior design team used hypothesis testing to determine whether or not the mean of the observed and simulated metrics were equal. The results show that the difference between the two means is insignificant with a 95% confidence interval, validating the model. See Appendix I for an explanation of these calculations. 14 5. Recommendations and Valuation 5.1.Recommendations The senior design team ran the PDAT for 20 stores for multiple months in 2013 (Appendix E). Each store and day was run for the following scenarios: FaST Baseline Scenario: In this scenario, no associate roles were defined. This scenario is represents the current state. Desk Queue Threshold Scenario of {1, 2, 3, 4}: In these four scenarios, associate roles were defined and tested with desk queue thresholds of 1, 2, 3, and 4. Desk Queue Threshold Scenario of {1, 2, 3, 4} and Additional or Reduced Labor: In these four scenarios, associate roles were defined and tested with desk queue thresholds of 1, 2, 3, and 4, and labor levels were adjusted based on the algorithms in the Staffing Adjustments Model (Section 3.8). The senior design team recommends that THD defines associate roles, implements a maximum desk queue threshold of 1, encourages the purchase of additional accessories, and continues investigating opportunities for reallocating labor. 5.2.Valuation Due to the intricacies of staffing individual shifts, the senior design team defined two different project values: one without labor modifications and one with labor modifications. The project value without labor modifications is the more conservative estimate of value, while the project value with labor modifications is the more optimistic scenario and can only be achieved if shifts could be perfectly in sync with the labor modification suggestions. The Home Depot will be able to use the average suggested labor changes to understand where they are currently overstaffed and understaffed. This will allow them to staff associates to create a more consistent customer experience, improving customer service and profitability. 15 The project recommendations without labor modifications provide THD with an annual incremental profit (pre-tax) of $10.4 million (Table 8). Additionally, the improved customer services metrics are outlined in Table 9. Table 8: This table shows the itemized buckets by sales volume and the annual incremental profit gained by implementing project recommendations without labor modifications. Store Sales # of Stores Avg Daily ∆ Avg Daily Avg Daily ∆ Annual Volume Accessory Sales ∆ Renege Profit Incremental Profit 606 $6.72 $2.02 $8.75 $1.9 M A ($0-25M) 924 $12.29 $0.03 $12.32 $4.1 M B ($25-40M) 417 $25.14 $1.44 $26.58 $4.0 M C ($40-70M) 36 $25.56 $2.03 $27.59 $0.4 M D ($70-90M) 1983 $13.53 $0.97 $14.50 $10.4 M Total Table 9: This table shows improved customer service metrics as a result of implementing project recommendations. Store Sales Volume Reduction in Percentage of Reduction in Percentage of Customers that Renege Customers Waiting > 2.5 Minutes 0.3% 2.2% A 0.04% 2.2% B 0.1% 2.1% C 0.2% 3.2% D The project recommendations with labor modifications provide THD with an annual incremental profit (pre-tax) of $29.1 million (Table 10). Furthermore, the customer service metrics still improve despite the overall reduction in staffed hours (Table 11). Table 10: This table shows the itemized buckets of profit and the annual incremental profit gained by implementing project recommendations with labor modifications. Store # of Stores Avg Daily ∆ Avg Avg Daily Avg Daily ∆ Annual Volume Accessory Sales Daily ∆ ∆ Staffing Profit Incremental Renege Profit 606 $5.68 $12.56 $8.88 $27.12 $6.0 M A 924 $10.47 $9.44 $19.06 $38.96 $13.1 M B 417 $21.18 $12.76 $24.94 $58.88 $8.9 M C 36 $20.34 $12.40 $48.41 $81.16 $1.1 M D 1983 $11.43 $11.14 $17.72 $40.29 $29.1 M Total 16 Table 11: This table shows improved customer service metrics as a result of implementing project recommendations. Store Volume Reduction in Percentage of Reduction in Percentage of Customers that Renege Customers Waiting > 2.5 Minutes 2.1% 2.1% A 1.3% 1.0% B 0.1% 0.9% C 0.8% 1.7% D 17 6. Appendices 6.1.Appendix A: Customer and Mixer Data Collection Procedure The senior design team collected data regarding customer arrivals, traffic, and service times as well as mixer utilization times at three Atlanta-area THD stores. Table A1 details the locations and dates of data collection. Table A1: The table depicts the location and dates of data collection in the Atlanta area. Store Location Austell Buckhead Midtown Customer Collection Dates 8/24/2013 9/3/2013, 9/28/2013, 10/10/2013 9/13/2013, 9/14/2013, 10/1/2013, 10/10/2013 Mixer Collection Dates Store Volume Bucket 9/28/2013 B C 9/13/2013, 9/14/2013, 10/1/2013 C On these dates, the team collected timestamp data to record: Customer Data ● Customer arrivals ● Customer wait times ● Customer service times ● Customer reneges ○ A brief physical description of each customer was also recorded to ensure that customers who reneged did not return Mixer Utilization Data ● Number of mixers in use 18 6.2.Appendix B: Mixer Utilization Data Analysis The following tables depict the utilization of paint mixers in fifteen-minute intervals (Tables B1 and B2). Table B1: Mixer utilization for the Midtown THD location on September 14, 2013. Midtown Resource Utilizations - Saturday, 9/14/13 Time Bucket 10:15 AM 1 Gallon 0.00% 5 Gallon 0.00% Multi Vol 0.00% Queue Tint 0.00% 0.00% 10:30 AM 28.06% 0.00% 0.00% 0.00% 8.06% 10:45 AM 18.83% 0.00% 0.00% 0.00% 21.89% 11:00 AM 13.96% 10.89% 0.00% 0.00% 9.28% 11:15 AM 7.19% 0.00% 0.00% 6.00% 18.22% 11:30 AM 25.63% 0.00% 0.00% 39.22% 13.39% 11:45 AM 36.43% 0.00% 0.00% 64.00% 31.67% 12:00 PM 11.20% 0.00% 31.67% 54.89% 57.61% 12:15 PM 35.50% 0.00% 23.00% 34.22% 44.33% 12:30 PM 27.72% 24.78% 0.00% 11.22% 63.28% 12:45 PM 10.15% 16.50% 0.00% 0.00% 6.78% 1:00 PM 34.22% 43.61% 0.00% 20.78% 46.17% 1:15 PM 15.43% 10.89% 0.00% 7.22% 17.78% 1:30 PM 22.13% 0.00% 0.00% 21.11% 28.94% 1:45 PM 15.98% 0.00% 0.00% 0.00% 15.22% 2:00 PM 12.41% 0.00% 0.00% 10.56% 11.11% 2:15 PM 11.93% 13.50% 0.00% 49.33% 10.83% 2:30 PM 15.48% 0.00% 0.00% 14.67% 2.39% 2:45 PM 7.04% 0.00% 0.00% 11.56% 1.28% 3:00 PM 0.00% 0.00% 0.00% 0.00% 0.00% 3:15 PM 0.00% 0.00% 0.00% 0.00% 0.00% 3:30 PM 0.00% 0.00% 23.33% 0.00% 0.00% 3:45 PM 29.56% 0.00% 96.11% 0.00% 0.00% 4:00 PM 1.85% 0.00% 77.89% 0.00% 0.00% 4:15 PM 52.43% 0.00% 54.89% 0.00% 0.00% 4:30 PM 53.43% 0.00% 0.00% 0.00% 0.00% 4:45 PM 18.06% 0.00% 0.00% 0.00% 0.00% 5:00 PM 13.20% 0.00% 0.00% 0.00% 0.00% 5:15 PM 17.43% 0.00% 0.00% 0.00% 0.00% 5:30 PM 4.65% 0.00% 0.00% 0.00% 0.00% Average 18.00% 4.01% 10.23% 11.49% 13.61% Max 53.43% 43.61% 96.11% 64.00% 63.28% 19 Table B2: Mixer utilization for the Buckhead THD location on September 28, 2013. Buckhead Resource Utilizations - Saturday, 9/28/13 Time Bucket 1 Gallon 5 Gallon Multi Vol Queue Tinter 8:00 AM 0.00% 0.00% 0.00% 0.00% 69.22% 8:15 AM 0.00% 0.00% 0.00% 0.00% 0.00% 8:30 AM 0.00% 0.00% 0.00% 0.00% 0.00% 8:45 AM 0.59% 0.00% 0.00% 11.67% 5.33% 9:00 AM 9.28% 0.00% 0.00% 4.22% 6.48% 9:15 AM 6.80% 0.00% 25.22% 25.67% 3.26% 9:30 AM 46.54% 16.28% 48.78% 7.33% 50.89% 9:45 AM 13.50% 19.17% 24.11% 71.56% 18.26% 10:00 AM 6.13% 0.00% 23.11% 7.11% 15.11% 10:15 AM 14.15% 0.00% 0.00% 3.33% 7.11% 10:30 AM 16.89% 0.00% 0.00% 41.56% 20.41% 10:45 AM 34.80% 0.00% 0.00% 39.33% 11.59% 11:00 AM 15.30% 0.00% 17.89% 16.00% 2.26% 11:15 AM 0.00% 0.00% 0.00% 10.22% 5.93% 11:30 AM 29.19% 0.00% 0.00% 12.78% 3.85% 11:45 AM 0.00% 0.00% 0.00% 4.78% 5.67% 12:00 PM 0.00% 0.00% 0.00% 16.22% 3.30% 12:15 PM 0.00% 0.00% 0.00% 0.00% 0.00% 12:30 PM 0.00% 0.00% 24.44% 34.22% 20.33% 12:45 PM 0.00% 0.00% 0.00% 35.11% 1.48% 1:00 PM 10.30% 0.00% 0.00% 0.00% 28.37% 1:15 PM 14.81% 0.00% 0.00% 19.11% 1.93% 1:30 PM 43.37% 0.00% 0.00% 27.89% 31.70% 1:45 PM 43.83% 0.00% 0.00% 25.22% 18.26% 2:00 PM 22.35% 0.00% 0.00% 16.44% 13.67% 2:15 PM 6.85% 0.00% 0.00% 0.00% 0.00% 2:30 PM 0.00% 0.00% 0.00% 0.00% 0.00% 2:45 PM 1.41% 0.00% 0.00% 0.00% 0.00% 3:00 PM 27.69% 0.00% 0.00% 0.00% 0.00% 3:15 PM 24.91% 0.00% 0.00% 0.00% 14.59% 3:30 PM 6.87% 0.00% 0.00% 18.67% 33.33% 3:45 PM 16.81% 0.00% 0.00% 0.00% 0.00% 4:00 PM 3.19% 0.00% 0.00% 0.00% 29.56% 4:15 PM 27.98% 0.00% 23.22% 0.00% 0.00% 4:30 PM 24.52% 0.00% 0.00% 0.00% 0.00% 4:45 PM 0.00% 0.00% 0.00% 0.00% 0.00% 5:00 PM 0.00% 0.00% 0.00% 0.00% 28.63% 5:15 PM 0.00% 0.00% 0.00% 0.00% 16.41% Average 12.32% 0.93% 4.92% 11.80% 12.29% Max 46.54% 19.17% 48.78% 71.56% 69.22% 20 6.3.Appendix C: Customer Arrival Regression Customer arrival times are modeled from a regression equation using the senior design team’s observed D24 arrival times and existing timestamp transaction data provided by THD. This model divides the day into fifteen-minute intervals and places customer arrivals and transactions in the appropriate intervals. The regression equation defines the following variables: 𝑦𝑡 = the number of arrivals in a 15-minute time interval t 𝑥𝑡 = the number of transactions that occur in interval t Because all paint customers arrive at the cashier after arriving at the paint desk, 𝑦𝑡 can be estimated through linear regression: 𝑦𝑡 = 0.471𝑥𝑡+1 + 0.498𝑥𝑡+2 A study from the Columbia Business School2 states that the sum of the coefficients should be positive and less than or equal to 1. The sum of the coefficients in the customer arrival regression equals 0.969. Because of the potential error in data collection as well as the variability in human behavior, the team scaled the coefficients in order to account for all customer arrivals. 𝛽1 0.471 = = 49% 𝛽1 + 𝛽2 0.969 𝛽2 0.498 = = 51% 𝛽1 + 𝛽2 0.969 Additionally, it can be assumed that customers arrive with a uniform distribution within a given 15-minute interval (Table C1). Table C1: The table explains the probability of when a customer will arrive at the Point-of-Sale after arrival at the paint desk. Distribution of Time Between D24 Arrival and Likelihood of this Distribution Transaction Timestamp (Minutes) 49% Uniform(15,30) 51% Uniform(30,45) 6.4.Appendix D: Customer Renege Model 2 Lu, Yina, Andres Musalem, Marcelo Olivares, and Ariel Schilkrut. Measuring the Effect of Queues on Customer Purchases. Columbia Business School, 6 Nov. 2012. 21 This model determines the likelihood that a customer arriving to the paint desk decides to renege based on the number of customers ahead of him in the system. Some customer reneges that occurred during data collection happened when associates left the paint desk unattended. The senior design team recommends that in the future, associates never leave the paint desk unattended, preventing this type of renege from occurring. Therefore, of the customers who reneged during data collection, only those who reneged when an associate was at the desk are included in the following calculations. A logistic regression was used to calculate the probability that a customer will renege given the number of customers ahead of him in the system at that time. The expected probability that an arriving customer reneges given the number of customers ahead of them in the system at that time, n, is calculated using a logistic regression shown in Figure D1. Figure D1: The logistic regression equation used to calculate the probability that a customer will renege. Table D1 shows the observed and calculated probabilities that a customer will renege. Table D1: Observed and calculated renege probabilities Customers at Desk Ahead Number of of Arriving Customer Observed Customers 156 0 166 1 152 2 121 3 53 4 58 ≥5 706 Total Observed P(Arriving Customer Reneges) Expected P(Arriving Customer Reneges) 0.0% 1.8% 4.0% 5.0% 1.9% 6.9% 2.8% 0.0% 2.2% 2.8% 3.4% 4.3% 5.3% The simulation model uses the expected likelihood that an arriving customer reneges to assign a renege decision to each customer based on the number of customers ahead of them at the paint desk. Reneging customers do not receive customer service or make a purchase. 22 6.5.Appendix E: Recommendation and Valuation Data Table E1: These stores/dates were run to define recommendations. Store Volume A 0424 A 1924 A 0816 A 4031 A 0144 A 1119 B 0539 B 2740 B 0366 B 3312 B 6301 C 8444 Months January, April, July, October April, October April, October January January, April, July, October April April, October April, October January, April, July, October January January April, October Table E2: These stores/dates were run to define project value. Store Volume A 0424 A 1924 A 0816 A 4031 A 0144 A 1948 A 1119 B 0539 B 2740 B 0366 B 3312 B 0550 B 6001 B 6301 B 1803 C 8444 C 0245 C 1234 C 6306 D 2667 Months January, April, July, October April, July, October January, April, July, October January January, April, July, October January, April, July, October April, July, October April, July, October April, July, October January, April, July, October January January, April, July, October January, April, July, October January, April, July, October January, April, July, October April, July, October January, April, July, October January, April, July, October January, April, July, October January, April, July, October 23 6.6.Appendix F: PDAT User Guide Instructions Select the Instructions tab. Enter the store number and date to be analyzed. Enter the file location for the current state simulation, future state simulation, and data file. If the data file does not already exist, the PDAT will create the required file at the defined location. In both Arena simulations (current and future), navigate to Advanced Transfer in the project bar, then click File. Change the file extension to match the location used for the data file. Input Transactions Select the INPUT Transactions tab. 24 Clear any prior entries. Insert the correct data into columns A-F. Columns G and F will be automatically generated when the program is run. Input Schedule Select the INPUT Schedule tab. Clear any prior entries and input schedule data into columns A-K. 25 Staffing Recommendations Select the Staffing Recommendations tab. Type appropriate value for both service protocol options and machine options under the “Current State”, and then click “Run FaST Benchmark.” Type appropriate value for both service protocol options and machine options under the “Comparison Scenario”, and then click “Run Comparison Scenario.” 26 6.7.Appendix G: Associate Task Statistical Distributions The senior design team used time study data provided by THD to identify statistical distributions for each of the 107 tasks in the customer service process. Table G1 shows examples of these distributions (in minutes). Table G1: A sample of associate task distributions broken down by item and task. Task Distribution LOGN(0.285, 0.193) Gallon: Queue Manager ERLA(0.367, 2) Gallon: Retrieve Paint Base for Customer LOGN(0.178, 0.151) Sample: Tint Paint WEIB(0.323, 1.57) Quart: Tint Paint 2 * BETA(0.957, 4.67) Gallon: Tint Paint EXPO(0.777) 5 Gallon: Tint Paint ERLA(0.125, 2) Gallon: Remove Lid WEIB(0.165, 1.34) Gallon: Place Base in Mixer EXPO(0.303) Gallon: Dab/Dry Sample on Can Lid 3.4 * BETA(0.47, 4.89) Gallon: Give Mixed Paint to Customer 27 6.8.Appendix H: Simulation Flow Detail Enter D24 Two entities are created to start the simulation. The first entity moves through the Read Number of Resources block, where the number of resources and user-defined inputs (Queue Managers, Tinters, One Gallon Mixers, Five Gallon Mixers, Multi-Volume Mixers, Maximum Capacity of Employees and Minimum Customer Queue Threshold) are read. From there, the entity is duplicated twice and each enters two separate loops. The first loop reads in the associate data (schedule time and number of associates) and then updates the schedule. In the current state model, there are no defined associate roles, while in the future state model, both desk and aisle roles are defined and scheduled separately. The associates are staffed at the desk based on the maximum number of associates needed during the past 15 minute interval. The remaining associates will be staffed in the aisles. The second loop is used to check the utilization of all of the resources every minute. This value is used to record the average utilization every fifteen minutes, which shows overall idle time. All of this data is then output to the PDAT for analysis. The second entity, the order entity, moves through the Read Customer Information block, where all of the transaction data and user-defined values from the PDAT (Arrival Time, Transaction Number, Customer Type, Number of Samples, Quarts, Gallons, Five Gallons, and Accessories) are read. The customer entity is held at a delay block until that customer’s arrival time. From there, a customer number is assigned and the order enters D24. Need Paint If there are any paint items, the order is directed to the desk and enters the desk system. If there are no paint items, then the order is directed to the aisle. Renege Decision The customer may decide to renege based on probabilities defined in the Customer Renege Model. Once the customer enters the desk system, if there are no associates at the desk, the customer will renege and go to Leave D24. If associates are present, the probability that the incoming customer will renege increases according to the number of customers in the queue. If a customer does not renege, he will go to Process Order. Process Order The Process Order submodel is the first process at the paint desk. In the current model, the order is processed based on available associates. If there are no available associates, the customer will enter the queue. In the future state model, if the number of customers in the desk queue reaches the desk queue threshold, an associate will move from the aisle to the desk. The order is then 28 input into the system by seizing a queue manager. The distribution used at the Process Order block is a weighted average based on processing time from all paint items in the order. Retrieve Decision This submodel takes into account the maximum number of paint base items that an associate can carry at one time. An associate can carry 4 samples, 2 quarts, 2 gallons, or 1 five gallon per trip. Tint Decision A temporary variable is assigned to the order based on the total number of each volume of paint item in the order. Every item is processed (tinted) with a distribution specific on its volume. The temporary variable is reduced by one as each item is processed. The order will wait in the queue if all of the tinters are being utilized. Mix Decision Similar to the Tint Decision submodel, the Mix Decision submodel uses a loop to reduce the total number of paint items by one until all items are processed. If one of the normal mixers (1 gallon or 5 gallon) is available, the paint item will be mixed in that resource for 2.5 minutes. If all normal mixers are busy and the multi-volume mixer is available, the paint item will be mixed in the multi-volume mixer for 2 minutes. The multi-volume mixer can process 12 samples, 9 quarts, 4 gallons, or 1 five gallon paint item. Counter Service Logic This submodel includes all of the task distributions that occur at the desk counter (e.g. testing the paint color and giving it to the customer). Once the interaction between the associate and the customer is complete, the associate is released and the total order service time is recorded. All customers leaving the paint desk enter the Aisle Logic submodel. Aisle Logic “DIY with accessories” and “aisle” customers receive desk service from associates in the Aisle Logic submodel. In the future state model, if an aisle associate is available when a “DIY without accessories” customer finishes his service at the desk, the aisle associate will encourage the “DIY without accessories” customer to purchase additional items. Professional customers will go to Leave D24 because they do not require aisle service. If no associates are available, DIY customers will also go to Leave D24. Leave D24 In this submodel, the customer’s statistics are output for PDAT analysis. Then, all services are complete and the customer leaves the paint department. 29 6.9.Appendix I: Simulation Validation Hypothesis Testing The senior design team compared customer service metrics that were observed during in-store data collection against simulated metrics for the same stores and dates. Tested metrics include the percentage of customers deciding to renege, the percentage of customers waiting more than 2.5 minutes for desk service, and the percentage of customers waiting more than 5 minutes for desk service. The average difference between simulated and observed metrics is shown in Table I1. Ideally, each of these differences would be equal to 0. However, no matter how accurate the simulation model is, there will always be some discrepancies due to variability in human behavior. Table I1: The observed and simulated service metric statistics. Statistic % Customers who % Customers who Renege Wait > 2.5 Minutes 2% 18% Observed Mean 3% 24% Simulated Mean Difference Between 1% 6% Simulated and Observed Means % Customers who Wait > 5 Minutes 7% 8% 1% To validate the model, the senior design team used hypothesis testing to compare observed customer service metrics against simulated metrics. The initial hypothesis for each metric was that the observed metric is equal to the simulated metric. Table I2 shows hypothesis testing calculations for each metric. Each of the metrics failed to reject the initial hypothesis that the observed and simulated metrics were equal, thus validating the model. 30 Table I2: Two-tailed hypothesis testing results used to validate the simulation. Statistic % Customers who % Customers who Renege Wait > 2.5 Minutes The difference between The difference between observed and simulated observed and simulated H0 means is not statistically means is not statistically significant. significant. 2% 18% Observed Mean Observed Standard 2.6% 8.9% Deviation Number of Observed 8 8 Dates 3% 24% Simulated Mean Simulated Standard 1.2% 10.9% Deviation Number of Simulated 8 8 Dates 0.0203 0.0995 Sp -0.907 -1.167 T (observed) T statistic (2-tailed, df 2.145 2.145 = 14, α = 0.05) Is T (observe) > T No No statistic? Fail to Reject H0 Fail to Reject H0 Result 95% Confidence -3.1%, 1.3% -16.5%, 4.9% Interval % Customers who Wait > 5 Minutes The difference between observed and simulated means is not statistically significant. 7% 3.5% 8 8% 8.3% 8 0.0635 -0.431 2.145 No Fail to Reject H0 -8.2%, 5.4% 31