Abstract:

The aim of the article is to discuss forecasting accuracy measures for intermittent data. It is emphasized that many forecasting errors are inappropriate for intermittent data. Accuracy metrics for such data should feature specific properties. Division by zero must be avoided. They ought to be scale-independent (relative) and also unbiased. It is assumed that a given accuracy metric is biased if it favors underestimated or overestimated forecasts. In this article unbiasedness is particularly emphasized. It was demonstrated that such error measures like MASE favor underestimated forecasts, often zero forecasts. Therefore, other metrics are discussed, such as mean-based errors (mean-based MAPE), Cumulative Forecast Error, Number of Shortages, Periods in Stock. Moreover, the author’s proposal is presented – namely the scaled Compound Error. All of these metrics are evaluated on the theoretical and empirical basis. In the empirical research presented measures are evaluated in terms of two forecasting methods: TSB and SES. Zero forecasts were also taken as a benchmark. Forecasts for 355 time series with high intermittence from a company selling work clothes and tools were analyzed. The final remark is that mean-based MAPE seems to be the best accuracy measure for highly intermittent data.