A Meta Forecasting Methodology for Large Scale Inventory Systems with Intermittent Demand

This paper presents a meta-forecasting approach for recommending the most appropriate forecasting technique for an intermittent demand series based on a multinomial logistic regression classifier. The meta-forecaster is based on a mapping between a demand attribute space and the best forecasting technique. The demand attribute space is based on the estimates from the demand series of the following attributes: probability of non-zero demand after zero demand, probability of non-zero demand after non-zero demand, mean demand, demand variance, lag 1 correlation coefficient of the interval between non-zero demand and lag 1 correlation coefficient. Based on the mapping, the best forecasting technique for an unknown demand vector can be predicted. Given the demand series, the demand attributes are estimated and then the classifier is used to predict the best forecasting technique. After training, the classifier was tested. The results indicate an accuracy rate of 70.87% for the recommended best forecasting technique; and an 87.94%accuracy rate for the recommended top two forecasting techniques.