RegL - um modelo de predição de séries temporais otimizado para dados do setor varejista

Abstract

Demand forecasting and inventory optimization are key challenges in the retail sector, as accurate decisions directly impact operational efficiency and business competitiveness. This work presents the regL model, an optimized hierarchical approach for time series forecasting in the retail context, with an emphasis on data aggregated by groups of similar products. regL aggregates historical data into groups of similar products, reducing the impact of noise and volatility present in the individual product series. Forecasting is initially performed at this aggregated level and then adjusted through regression to the granular level of individual products, providing greater accuracy and computational efficiency. The model was evaluated using real retail data and compared to traditional statistical methods (ARIMA, Exponential Smoothing, Linear Regression), machine learning models (XGBoost, LightGBM, Prophet), and deep neural networks (RNN, NHiTS, TFT), employing relevant performance metrics. regL consistently outperformed these methods, achieving errors below 0.12 in the adjusted mean absolute percentage error (EPAA) and execution times of up to 30 seconds for 50 time series. Thus, regL is positioned as a practical and efficient solution for daily retail applications, significantly contributing to the optimization of inventory, purchasing, and logistics processes.