Evaluating predictive count data distributions in retail sales forecasting

Massive increases in computing power and new database architectures allow data to be stored and processed at finer and finer granularities, yielding count data time series with lower and lower counts. These series can no longer be dealt with using the approximative methods that are appropriate for continuous probability distributions. In addition, it is not sufficient to calculate point forecasts alone: we need to forecast entire (discrete) predictive distributions, particularly for supply chain forecasting and inventory control, but also for other planning processes. However, tools that are suitable for evaluating the quality of discrete predictive distributions are not commonly used in sales forecasting. We explore classical point forecast accuracy measures, explain why measures such as MAD, MASE and wMAPE are inherently unsuitable for count data, and use the randomized Probability Integral Transform (PIT) and proper scoring rules to compare the performances of multiple causal and noncausal forecasting models on two datasets of daily retail sales.