Arquitetura híbrida para predição de vendas no varejo com seleção automatizada de modelos

Abstract

This work presents a hybrid architecture for time series forecasting in the retail context, based on the automated selection of the most suitable model for each series. The approach combines the extraction of statistical and structural features from the series with the application of multiple predictive models, whose performances are used to train a classifier capable of intelligently identifying the most effective algorithm for each situation. This strategy mitigates the impact of data heterogeneity and volatility, avoiding the use of a single, often insufficient model, as well as the exhaustive evaluation of all models, which incurs a high computational cost. The architecture was evaluated using real sales time series and compared with traditional methods and modern machine learning models, employing metrics widely accepted in the literature. The results show that the automatic selection achieves performance equivalent to the exhaustive process, with marginal differences in metrics such as RMSE, MSE, MAPE, and R², while significantly reducing execution time.