FORECASTING FLUCTUATING DEMAND IN MANUFACTURING

THE OBJECTIVE OF THIS PROPOSED PROJECT IS TO INVESTIGATE METHODS FOR FORECASTING "LUMPY" DEMAND FOR PRODUCTS AND SERVICES. LUMPY DEMAND IS USUALLY ZERO BUT INTERMITTENTLY NON-ZERO. STANDARD FORECASTING SYSTEMS CANNOT COPE WITH LUMPY DEMAND, WHICH IS A COMMON PROBLEM IN PRODUCTION AND INVENTORY CONTROL. OUR INTEREST IN THE PROBLEM IS A RESPONSE TO OUR CUSTOMERS IN MANUFACTURING. EXISTING APPROACHES TO LUMPY DEMAND FORECASTING EITHER CONVERT LUMPY DATA INTO CONVENTIONAL DATA FOR CONVENTIONAL TREATMENT OR MAKE PARAMETRIC MODELS OF THE LUMPY DEMAND (E.G., POISSON). THERE HAS NOT BEEN A SYSTEMATIC COMPARISONOF THESE ALTERNATIVES. WE WILL EMPIRICALLY ASSESS THE APPROPRIATENESS AND ACCURACY OF THESE ALTERNATIVES. WE WILL ALSO INVESTIGATE TWO INNOVATIVE APPROACHES. ONE IS TO EXPLORE WHETHER ASSOCIATIONS ACROSS SERIES OCCUR IN REAL DATA AND HOW THEY MIGHT BE EXPLOITED. THE OTHER IS TO ANALYZE THE METHOD USED MOST OFTEN IN PRACTICE: "SALES-FORCE COMPOSITE" FORECASTING, WHICH TOTALS THE ESTIMATES OF DEMAND MADE BY THE SALES FORCE FOR EACH POTENTIAL CUSTOMER. COMBINATION OF SALES FORCE ESTIMATES WITH TIME-SERIES TECHNIQUES MAY HAVE SPECIAL MERIT WHEN STATISTICAL DATA ARE SPARSE, AS WITH LUMPY DEMAND. THERE MAY ALSO BE WAYS TO WORK DIRECTLY WITH THE SALES FORCE ESTIMATES, SUITABLY DEBIASED.