İlaç Sektöründe Zaman Serisi ve Regresyon Birleşik Modeller ile Talep Tahmini Uygulaması

Abstract

Accurate demand forecasting is crucially important to reduce inventory and backlogging cost. In this study, we analyze howpromos, holiday statements, price changes, stock availability and date-time features (weekdays, months etc.) affect thedemand by using several forecasting methods. Data sets were collected for the products of the global pharmaceuticalcompany providing services in Turkey. Actual daily sales data for 2016, 2017 and 2018 were used in the construction of thisdata set. In order to predict the next periods demand, we used four different models which are Holt Winters, RidgeRegression, Random Forest and Xgboost. We also ensemble those models to improve forecasting accuracy. Next, byweighting inversely proportional to the error rates of the models, binary, triple and quadruple combinations of the singlemodels were compared with themselves and the single models. Our numerical results show that the lowest forecasting errorrate was obtained in ensemble models. Particularly, the lowest error rate in individual models was obtained in Random Forestwith 15.7% RMSPE (Root Mean Square Percentage Error) value, and the lowest error rate was obtained with 10.7% RMSPEvalue in Holt Winters & Xgboost models combination. Results show that ensemble of several models can increase theforecasting accuracy.