Perakende satış tahmini için dinamik zaman bükme yöntemi ile benzerlik gruplaması ve içgörü geliştirilmesi

Retail industry is a dynamic industry with many observable and hidden drivers of sales. Accurately forecasting short and long term sales is a major advantage to retailers as it helps with decision and policy making. Along with forecasting, determining and understanding hidden drivers of sales is invaluable for identifying indicators of change in trends. In this thesis, we propose a method for increasing forecasting accuracy of the model proposed by Gür Ali and Pınar (2013), and provide a new method for automatic similarity identification to gain insights. The base model is a multi-period two-layer pooled regression model. The first layer considers marketing, inflation and seasonal effects. In the second layer, the residuals from the first layer are extrapolated to identify trend-cyclical components of the sales. The pooling is done according to characteristics of stores, predetermined by the company. We propose behavioral pooling to improve the accuracy of the forecasts. Behavioral pooling groups the stores according to their similarity in movement patterns. We use Dynamic Time Warping method to quantify similarity between stores. The resulting pooling significantly improves the accuracy of the base model. Dynamic Time Warping is also used as a similarity measure between the residuals and environmental and socio-economic indicator time series to gain insights about other potential drivers of sales.