Explainable Deep Learning for Public Revenue–Expenditure Dynamics: A Data-Driven Framework for Intelligent Fiscal Planning

Abstract

Accurate forecasting of public revenue and expenditure is essential for effective fiscal planning and budgetary management. This study investigates the applicability of machine learning and deep learning algorithms to the joint forecasting of public revenue and public expenditure. A comparative forecasting framework was developed using Random Forest, XGBoost, Long Short-Term Memory (LSTM), Gated Recurrent Unit (GRU), Bidirectional LSTM (Bi-LSTM), and Transformer models. A naïve forecasting benchmark was also incorporated to provide a reference for evaluating the predictive value of the more complex algorithms. Model performance was assessed using Mean Absolute Error (MAE), Root Mean Square Error (RMSE), Mean Absolute Percentage Error (MAPE), and the coefficient of determination (R²) under a chronological out-of-sample evaluation framework.

The empirical results reveal substantial differences in predictive performance across the evaluated architectures. For public revenue, the naïve benchmark achieved the best overall performance, with an MAE of 535,629, an RMSE of 712,275, a MAPE of 10.10%, and an R² of 0.485. Among the deep learning models, LSTM achieved the strongest performance, with an MAE of 809,035, an RMSE of 867,162, a MAPE of 17.07%, and an R² of 0.2366. For public expenditure, the naïve benchmark again produced the lowest forecasting errors, while Bi-LSTM achieved the best performance among the deep learning models, obtaining an MAE of 1,750,307, an RMSE of 1,811,825, and a MAPE of 24.08%. In contrast, the Transformer model produced the largest forecasting errors for both variables.

The findings demonstrate that increasing model complexity does not necessarily lead to superior forecasting accuracy, particularly when working with a limited number of annual observations. The results emphasize the importance of benchmark comparison, model selection, and chronological out-of-sample validation when applying artificial intelligence to fiscal forecasting. The study provides empirical evidence that recurrent deep learning architectures can capture useful temporal patterns in public finance data, while also highlighting the importance of considering simpler forecasting alternatives in fiscal planning applications.