Forecasting a customer's Next Time Under Safety Stock

Abstract

Under intermittent demand, demand time series can contain noise and have intermittent characteristics which render their modeling difficult. Models for producing forecasts in these conditions are optimized on errors that are calculated on out-of-sample forecasts of demand time series. This is insufficient in an intermittent demand context, as intermittent demand time series tend to provide poor representations of the underlying stock behavior. In this paper, we propose a new, intuitive, and robust method of error measurement for evaluating intermittent demand time series forecasting models by estimating stock-out situations from demand forecasts. This metric is called the Next Time Under Safety Stock. Applying our error measurement to common intermittent time series forecasting models reveals that forecast combination remains the most competitive modeling framework. A case study allows for a comparison between available sources of demand information when forecasting for these stock-out events. Significant correlation is also found between our new error measurement and common time series error measurements, which allows them to serve as replacements, but with reduced efficacy.