FORECASTING THE CUMULATIVE EFFECT OF PROMOTIONAL CAMPAIGNS USING MACHINE LEARNING METHODS

Bystrov Aleksandr Igorevich

This work aims to improve the accuracy of forecasting the final promotional effect in the fast-food industry. The study proposes an innovative approach that involves isolating the sales increase driven by promotional campaigns into a separate time series and subsequently modeling this value using machine learning methods. To solve this problem, the gradient boosting algorithm (CatBoost) is employed, which effectively handles categorical and numerical features, including discount depth, promotion mechanics, product characteristics, and regional specifics. The methodology includes constructing a base forecast using exponential smoothing to estimate "non-promo" sales, followed by determining the difference between actual sales and the baseline. This difference serves as the target variable for training the model, allowing the avoidance of information loss characteristic of traditional time series models. The experimental part, based on data from 760 unique promotions, demonstrates that the proposed approach leads to an average improvement of 8% in the 1 – WAPE metric compared to the classical approach. The results confirm the possibility of more accurate demand forecasting, contributing to optimized inventory management and enhanced operational efficiency.

Keywords:demand forecasting, promotions, gradient boosting, promotional effect, fast-food industry