Previsão de demanda em uma Empresa do Setor Agropecuário baseada em um modelo de Machine Learning

Abstract

The Brazilian agribusiness sector holds a strategic position in the national economy, showing significant growth even in the face of complex challenges imposed by supply chain management, such as seasonality, climate variability, and market fluctuations. The logistical planning difficulties faced by this sector set it apart from other economic segments, as it deals with a range of factors such as production variability, infrastructure issues, and price volatility, all of which directly impact demand forecasting. Thus, this study aims to implement and compare Machine Learning models to improve the accuracy of demand forecasting for one of the main SKUs of a company in the sector. With a quantitative and applied approach, the research is characterized by an exploratory objective and uses techniques involving bibliographic research and case study. The theoretical foundation was built through a Systematic Literature Review of the most commonly used Machine Learning methods for demand forecasting. Subsequently, the following models were implemented in the company under study: Linear Regression, Lasso Regression, ElasticNet, Decision Trees, XGBoost, Gradient Boost, and Random Forest. The SKU selection was based on the construction of an ABC curve, from which the four most relevant products for the company’s revenue were analyzed according to the necessary criteria for applying the models. A historical time series of 41 months was used, with 32 months allocated for training and 9 months for testing, aiming to validate the accuracy of the forecasts generated. The implementation results, enhanced by hyperparameter optimization processes and cross-validation adapted for time series, indicated the Linear Regression model as the most suitable, showing a good fit for both training and testing data, along with optimized values in the WMAPE and R² metrics. Therefore, this study contributes by identifying the most effective forecasting model for the demand of a strategic SKU of the analyzed company, which can assist in defining an appropriate replenishment policy for the organization. Furthermore, it sets a precedent for the application of predictive models in other areas of the supply chain.