A Comprehensive Machine Learning Based Modeling of Income Tax Collection

Income tax is one of the important sources of revenue for each country, income tax forecasting is thus one of the important tasks of each country. This work presents a machine learning-based method based on Gross Domestic Product (GDP) and population data to forecast income tax collection. As a result, the violin plot shows the distribution of the data, namely that population values are concentrated around the middle, while GDP has a bimodal distribution, and income tax exhibits a pattern similar to that of the population. On both training and test data, several machine learning models were assessed for accuracy and generalization using Mean Squared Error (MSE), R-squared (R²), and Mean Absolute Percentage Error (MAPE). With Train MAPE at 2.85% and Test MAPE at 5.53%, Random Forest attained a Train MSE of 25.94 and a Test MSE of 51.00, so indicating good performance but modest overfitting. Although Gradient Boosting had a higher Test MAPE of 6.89% suggesting some overfitting, it scored almost perfect Train MSE of 0.04. While performing poorly on the test data with a Train MSE of 180.50, the Decision Tree fit the training data exactly (Train MSE of 0.00). With Train MSE of 2.32 and Test MSE of 25.49, CatBoost proved constant accuracy over both datasets, it could be considered as the best model for income tax prediction based on GDP and population since it excelled generally in stability and generalization.