Measuring Demand Forecasting Errors in EOQ Environments
Şevket Günter is an Assistant Professor in the Operations Management & Information Systems Area of the Department of Management of Bogazici University. He previously taught as an Assistant Professor at West Viginia University, WV, at Temple University, PA in the USA. He holds BA degrees in Mathematics, in Economics and in Business from Colby College, ME, an MA degree in Economics from the University of Rochester, NY, and a PhD degree in Operations Management from Syracuse University, NY. Dr. Gunter was a founding partner of Teknix Inc., a Philadelphia-based global financial forecasting consulting firm, and one of the founders of Rooster Graphics, a PA-based local graphic/website design firm. His primary research interests are in the areas of forecasting, production planning and project management. He is among the top-ten most-published researchers in the world in the area of the "combinations of forecasts". His research has appeared in International Journal of Production Research, European Journal of Operational Research, International Journal of Forecasting, Journal of Forecasting, Economic and Financial Modelling, Economic and Financial Computing (plus one that was accepted for publication by Management Science, and 9 other referreed papers published in the proceedings of various international conferences). He was an invited speaker or session/track organizer in several International Symposia on Forecasting (ISF) as well as being a member of the Organizing Committee of the 16th International Symposium on Forecasting in Istanbul. Dr. Gunter is also the developer of MAPACS and OPCIL software for Operations, COMFOR, TSG and ODSOX software for forecasting, and REACTION software for event studies.
Accuracy of forecasts are traditionally measured by error metrics such as the mean squared error (MSE) or the mean absolute percentage error (MAPE). We base our research on Clive Granger's (1969) proposition that the cost of a forecasting error is not necessarily proportional to the square or the absolute value of the error, and Judge and Yancey's (1986) ensuing proposition that that such measures should somehow reflect the true cost of using erroneous forecasts in the economic decision making problem of interest. We develop a new "decision-theoretic" forecast error measure for EOQ-like inventory control models that reflects the true costs of making forecasting errors more accurately than traditional forecast error measures such as the MSE or MAPE. The MRSDr (Mean Relative Squared Differences of Roots) is analytically demonstrated to be the correct forecast error measure to use for demand forecasting for inventory management in EOQ-type settings. The MRSDr is quite robust and reflects the correct format for evaluating the errors in various other inventory management problems, too. We also empirically demonstrate the sensitivity of the forecasting method selected by a manager to the error metric used in the selection via the 1428 monthly data series used in the infamous Makridakis "M3" forecasting competition.