Demand Forecasts Using Process Models and Item Class Parameters: Application of Ancillary Variables.

摘要

Theoretical and statistical results are presented on forecasting a time series (Dt) in conjunction with a correlated series (Ht). In the particular problem Dt is demand for a part which is on an aircraft which flies Ht hours in period t. Reasonable recursive models of the underlying demand process are postulated and it is shown that theoretically rigorous or heuristically satisfactory forecast algorithms can be obtained by applying Kalman filter or weighted moving average techniques to 4 time series: Dt, Dt/Ht, log Dt, log Dt/Ht. An important forecasting parameter - denoted k - is developed for the structural models; k is the ratio of the noise variance of a process to the variance of random changes in the process mean. It is found that forecasts utilizing flying hours do give improved performance; the best algorithm is a Kalman filter with a varying weighting parameter which depends upon the flying hours in a period and k, which is determined by the item's demand frequency class. When the ancillary variable (program variable) is end item density rather than flying hours, the algorithm is identical but with different k-values. Projected savings, over the current Army method of forecasting demands on the wholesale supply system, were 1.8 million dollars annually on the 10,000 parts in the data base.