FORECASTING REGIONAL PERIODIC TAX REVENUES USING MACHINE LEARNING ALGORITHMS

Abstract

Accurate forecasting of tax revenues is essential for regional fiscal planning, budget management, and effective allocation of public financial resources. This study investigates the applicability of machine learning and time-series forecasting approaches to regional periodic tax revenue forecasting using monthly tax revenue data from Syrdarya Region, Uzbekistan. The dataset contains observations for 12 major tax categories covering the period from January 2023 to December 2025. Two forecasting approaches were evaluated: Ridge Regression with temporal trend and seasonal sine-cosine features, and the ARIMA(1,1,1) time-series model. To preserve the chronological structure of the data, observations from 2023–2024 were used for model training, January–June 2025 for validation, and July–December 2025 for testing. The models were evaluated using Mean Absolute Error (MAE), Root Mean Squared Error (RMSE), and Mean Absolute Percentage Error (MAPE). On the test set, Ridge Regression achieved an MAE of 4038.43, RMSE of 4860.45, and MAPE of 32.10%, while ARIMA(1,1,1) obtained an MAE of 4030.45, RMSE of 6082.61, and MAPE of 38.55%. Although ARIMA demonstrated a slightly lower MAE, Ridge Regression provided lower RMSE and MAPE, indicating better overall forecasting performance. The results also show that model performance varies considerably across individual tax categories, suggesting that the characteristics of each revenue series should be considered when selecting a forecasting method. The study demonstrates the potential of machine learning-based forecasting for supporting regional tax revenue planning and provides a basis for developing more advanced hybrid forecasting systems. The primary out-of-sample evaluation was conducted over a six-month forecasting horizon from July to December 2025, while the final prospective forecasting exercise covered a 12-month horizon from January to December 2026.