Avaliação do uso de inteligência computacional para previsão de vendas

Abstract

The growing digitalization of retail processes has expanded the availability of operational data and reinforced the need for more accurate sales forecasting to support strategic decisions and optimize planning, inventory management, and financial performance. This scenario highlights the opportunity to apply computational intelligence techniques capable of handling complex patterns, seasonality, and exogenous variables, overcoming the limitations of traditional statistical methods. This study aimed to evaluate computational intelligence models applied to sales forecasting and compare them with a statistical reference model. The methodology involved collecting real sales and customer flow data via SAP and the Digifort API, performing preprocessing procedures, imputing artificial zeros using Poisson GLM, creating seasonal and contextual variables, building different versions of training datasets, and applying the Random Forest, XGBoost, and SARIMAX models, assessed through an 80/20 temporal holdout and walk-forward cross-validation. The results showed that computational intelligence models presented lower predictive error and greater temporal stability, with Random Forest standing out, while SARIMAX performed worse when faced with higher variable complexity. The study concludes that computational intelligence techniques are more effective and robust for sales forecasting in physical retail, offering valuable support for managerial decision-making.