Improving Forecast Accuracy by Optimizing Statistical Forecasting Results Usha Dasari, PhD - Ascend Performance Materials Gijoy Mathew - Cognizant Technology Solutions Learning Points Review best practices in Statistical Forecast modeling Discuss forecast accuracy improvement by using Multilevel Forecast runs Review continuous improvement of Forecasting Accuracy Agenda Overview Ascend Supply Chain Management Process Overview Why Statistical Forecasting? Implementation Approach Continuous Improvement Plan Q&A ASCEND - Markets ASCEND – Supply Chain 5 Manufacturing Plants 46 Warehouses 13 Sub Contractors 100 + Shipping Points Products Plastics Chemicals Polymers & Fibers Supply Chain Planning – Process Overview SAP Systems SAP ECC SAP SCM (DP/PPDS) SAP CRM Why Statistical Forecasting? Statistical forecasting can be very helpful in managing Ascend’s maketo-stock planning scenario Forecast accuracy at SKU level is essential to maintain Optimal Inventory Levels and improve On-Time-Delivery Ability to analyze historical data at various planning levels & tracking Signal Analysis with the help of Excel tools Combining Bottom-Up & Top-Down methods with proper disaggregation mechanisms to help improve forecast accuracy Forecast accuracy measurement reports considering various error factors and lead times Implementation Approach Data analysis of the product group and product Forecast modeling approach at various levels of data Forecast model selection using APO-DP and Excel for analysis Select the best-fit model for each data element using “Tracking Signal” methodology Apply the best-fit model, first at the product level, followed by at product group level for appropriate ratio of disaggregation Publish the Statistical Forecast in APO-DP Measurement of Forecast Accuracy Improvement Path 3 2 Continuous Improvement using Forecast accuracy analysis and action plan development Best model selection using Excel Tools Historical Data 1 Analysis/Data Patterns Data Analysis Identification of Historical Data Pattern Analyze for patterns like Trend, Seasonality, Constant, Random etc. Availability of history of sales Pareto analysis Pattern analysis - Trend, seasonality etc. Pattern same at Product group and Product level No Yes Bottom Up & Top Down Forecasting Initial forecasting at Product level to capture variation followed by forecast run at product group level Top Down Forecasting Forecasting at Product group level and disaggregate based on last 6 months of history Identify the History Pattern at Various Levels of the Data Data Analysis Historical Data Pattern 3500000 600000 3000000 500000 2500000 400000 2000000 300000 1500000 100000 0 0 200911 201001 201003 201005 201007 201009 201011 201101 201103 201105 201107 201109 201111 201201 201203 201205 201207 201209 500000 Product Group Level 200911 201001 201003 201005 201007 201009 201011 201101 201103 201105 201107 201109 201111 201201 201203 201205 201207 201209 200000 1000000 Product Level Pattern at product group level is NOT similar to product level • Step 1 :Forecasting happens at product level and data is aggregated to product group level. • Step 2 : Forecasting is carried out at product group level and data is disaggregated based on statistical forecast values of step 1 Forecast Modeling Approach Statistical Forecast is run in APO-DP with Auto Model strategy (system provides the best model fit and provides logs for various factors) Horizontal Large data sample is available, Values have equal weight & is not exponentially decreasing No Exponential Smoothing No Yes Both trend test and seasonal test negative Yes Horizontal or Random Random No Both trend test and seasonal test positive Yes Moving Average Crostons Model Trend and Seasonal Holts-Winters Method Trend Holts Model No Only trend test is positive Yes No Seasonal Pattern Winters Method Selection of Statistical Forecasting Model Forecast runs using different models and extract output for analysis in spreadsheet Check for seasonality, Trend etc. at product level Analyze the log of the forecast run for different parameters values Iterative approach to find the lowest MAPE value Forecast Model Selection Several iterations to arrive at the lowest MAPE by adjusting the forecast parameters (α, β, Ɣ) Test for model validity using “Tracking signal” test. Finalize the Forecast Model Parameters Multi-level Forecasting Process Forecasting model selection with relevant parameters for product and product group Forecast run first at product level to capture the SKU level disaggregation based on history, followed by Product group level such that the SKU level disaggregation ratio is maintained Publish Statistical Forecast Forecast Accuracy Analysis Forecast Accuracy reporting using appropriate lag Multi-level forecast Accuracy analysis Improve forecast accuracy and provide more visibility to various intricate challenges with demand forecasting Observing the market demand indicators, economy, sales etc. in statistical forecast correction Comparison of Statistical forecast with the existing Sales forecast showed an improvement of 19% (both compared to sales history) Continuous Improvement Plan User trainings/User involvements in Forecast model/parameter selection Optimize the history load and realignment processes After achieving credibility, use statistical forecast as the base for sales forecast Alerts, tolerance levels for sales forecast in comparison with statistical forecast Q&A THANK YOU FOR PARTICIPATING Please provide feedback on this session by completing a short survey via the event mobile application. SESSION CODE: 4113 For ongoing education on this area of focus, visit www.ASUG.com