Support Vector Machine for Demand Forecasting of Canadian Armed Forces Spare Parts

The need to reduce inventory costs and increase system operational availability is the main motivation behind improving forecast accuracy of military spare parts demand. In this paper, we assess the potential of Support Vector Machine (SVM) approach for forecasting the demand of Canadian Armed Forces (CAF) spare parts and we introduce a forecasting evaluation method using inventory cost performance curves based on over and under forecast error. We compare, using a well-known use case presented in the literature, the results given by SVM algorithm to those given by several popular forecasting approaches. We find that SVM performs better than, or equivalently to, the other methods for this use-case. We also perform some forecasting experiments using the historical data of forty CAF spare demand series with 84 periods (months) each. The results of the experiments show that SVM may offer forecasting improvements over many other methods however the performance of SVM is not quite as good on intermittent data (time series with a high Average Demand Interval-ADI and low Coefficient of Variation-CV).