SAP®’s Univariate Sales Forecasting Functionality: An Empirical Study

Abstract

The accuracy of sales forecasts is a major determinant of inventory costs and service level driven revenues. Failure to establish future customer demand without a reasonable degree of certainty often leads to poor levels of customer service and/or high levels of the wrong inventory and resulting reductions in company profitability. Enterprise resource planning (ERP) systems such as SAP are widely adopted and typically contain sales forecasting functionality. The sales forecasting accuracy of ERP systems can have a critical impact on business profitability and empirical research into the effectiveness of ERP offerings should be considered a valuable endeavour. However, commonly adopted measures of forecast accuracy, such as mean absolute error (MAE) and mean absolute percentage error (MAPE), do not provide explicit costs associated with forecast errors.

This thesis adopts a quantitative case study methodology to evaluate the nine forecasting models (2 moving average and 7 exponential smoothing) of SAP®’s enterprise resource planning system (SAP® ERP). The SAP forecast models are evaluated against both common statistical measures and commercial measures of forecast error, using a 24 month and 60 month product sales data set provided by a New Zealand based importer and retailer of electrical products. The forecast models are a combination of SAP default smoothing parameters, fitted smoothing parameters, baseline statistical forecasts, and event management adjusted forecasts resulting in a total of 38 forecast model combinations. The 38 forecast models are assessed using a rolling-origin holdout approach and evaluated against mean absolute error (MAE), mean absolute percentage error (MAPE), revenue-weighted mean absolute percentage error (RW-MAPE), and two new measures developed by the author; margin-weighted mean absolute percentage error (MW-MAPE) and cost of forecast error (CFE).

The findings of the case study support the fitting of forecast smoothing parameters using historical data, the selection of forecasting models based on historical time series characteristics, i.e. level, trend, and seasonality, and length of available history. The study also supports the evaluation of sales forecasts with cost of forecast error (CFE) as a more commercially useful measure than the widely adopted mean absolute error (MAE) and mean absolute percentage error (MAPE) measures. Event management adjustment of baseline statistical forecasts was not found to be statistically significant. However, an argument is presented that significance testing should be de-emphasised in favour of effect size (forecast error cost reduction in the case of this study). However, the results should be viewed as case specific and reflect the particular time series characteristics, costs, and margins of the company in question.

The study concludes with methodological recommendations for practitioners, ERP vendors, and academics which are supported by the specific case study results and the reviewed literature.