Estimativas de arrecadação do ICMS do Estado do Maranhão usando algoritmos de machine learning

Abstract

Tax collection forecasting is a cornerstone of fiscal planning and efficient public management. The Tax on Circulation of Goods and Services (ICMS) constitutes the main source of revenue for Brazilian states, and its accurate projection is crucial for allocating resources to strategic areas. However, the complexity of its dynamics, influenced by non-linear macroeconomic variables, and the lack of studies applied to the reality of the state of Maranhão pose a challenge for public administrators. This work aims to address this gap by investigating how machine learning techniques can improve the accuracy of forecasting monthly ICMS revenue in Maranhão. The overall objective is to develop and validate advanced computational models using a historical series of economic and social data from January 1997 to April 2024. This quantitative and applied research adopted the CRISP-DM framework. Data were collected from public sources such as SEFAZ-MA, IBGE, and the Central Bank. Initially, nineteen independent variables were considered, and a Multiple Linear Regression model was used to select the most relevant ones, such as GDP, diesel consumption, and electricity consumption indicators. Four machine learning algorithms were implemented, compared, and validated: Random Forest, Decision Tree, Linear Regression, and XGBoost. Performance evaluation was performed using the RMSE, MAE, MAPE, SMAPE, and R² metrics, using the k-fold cross-validation technique (with k=10) and a data split of 80% for training and 20% for testing. This study contributes a practical and validated model that can be integrated into the state's budget planning process, promoting more transparent, efficient, and data-driven fiscal management