Prediction: Ecommerce Corporation Evidences Demand

This research focuses on the applicability of machine learning tools in demand forecasting. The existing academic literature provides an incomplete answer to the question of validity of these tools in practice. Due to the large number of factors that can affect the implementation of demand, as well as the need to make forecasts for each commodity unit, this work focuses on comparing the forecasts made by ARIMA models and its modifications with random forest, at the level of each commodity item. The results were analyzed at the level of each SKU, and then combined by data quality, ABC segmentation, and product type. The sample used data from the second largest player in the Russian e-commerce market, under NDA. The main sample used 520 product positions with a historical depth of 166 weeks. The results obtained from the research are significant evidence of the applicability of machine learning approaches for products with low data quality and product groups whose demand determinants are not linear combinations of price indices.