Modelagem híbrida sarima-autoformer para previsão de demanda no varejo

Abstract

Demand forecasting plays a strategic role in retail management, as it influences decisions related to inventory management and operational planning. However, the high variability of sales time series, characterized by seasonality, trends, and nonlinear behaviors, imposes limitations on traditional statistical models. In this context, the literature has advanced toward the development of hybrid models that combine classical approaches with machine learning techniques. This study aims to evaluate to what extent residual modeling of the SARIMA model using the Autoformer architecture contributes to improving predictive performance and reducing uncertainty associated with retail decision-making. As a methodological reference, the decomposition approach proposed by Zhang (2003) is adopted, in which the statistical model captures the linear components of the series, while nonlinear models are employed to model the residuals. The proposed approach is compared with recurrent strategies in the literature based on Long Short-Term Memory (LSTM) and Random Forest (RF). The study uses univariate sales time series from the Ecuadorian retail chain La Favorita, considering different product families. Model performance is evaluated using the RMSE, MAPE, and Theil’s U metrics, in addition to the DM statistical test. The results indicate that the hybrid SARIMA–Autoformer model presents competitive performance and, in several scenarios, outperforms approaches based on LSTM and RF, particularly in capturing seasonal patterns and long-term dependencies. As a contribution, this study extends the literature by incorporating the Autoformer as a residual modeling strategy applied to retail demand forecasting