Developing Demand Forecasting Models for E-Commerce: Analyzing the Impact of Time Lags on Model Performance

Time series are an important analytical tool used in many problems today. Particularly favored in regression problems such as demand forecasting, time series enable more accurate modeling of the impact of past data on future values through various lag options. Time lag is a method used in time series analysis or machine learning models to examine the effect of past (lagged) values of a variable on current or future values. Time lag options play a crucial role, particularly in the success of demand forecasts. This study aims to develop demand forecasting models that help e-commerce businesses gain a competitive advantage by accurately predicting demand and comprehensively analyzing the impact of time delay options on forecasting performance. In this context, an interface with hyperparametric flexibility has been developed, and the effects of the lag options "Use Best N," "Use Correlation," "Use All Delays," and "Selected Delay Lag" on forecasting performance have been analyzed using demand forecasting models. Models have been created for two different months and three different products. The performance of the developed models has been evaluated using the Mean Absolute Percentage Error (MAPE) metric. The lowest MAPE value for July has been obtained with the MQRNN model developed using product A, while the lowest MAPE value for August has been obtained with the MLP model developed using product B.