Predicting Public Budget Surplus and Deficit Using a Hybrid 1D-CNN–LSTM Model

Abstract

The fiscal position of governments in rentier economies depends heavily on oil revenues. The relationship between oil prices and the budget surplus or deficit is often nonlinear and characterized by complex temporal dependencies, which may limit the predictive capability of conventional econometric models. Accordingly, this study aims to forecast the Iraqi budget surplus and deficit and compare the predictive performance of the ARDL, NARDL, LSTM, 1D-CNN, and hybrid 1D-CNN-LSTM models using oil prices as the primary predictive variable. The hybrid model integrates the feature-extraction capability of One-Dimensional Convolutional Neural Networks (1D-CNN) with the ability of Long Short-Term Memory (LSTM) networks to capture long-term temporal dependencies. The analysis is based on monthly Iraqi data covering the period 2008-2025 (216 observations), with the final year reserved for out-of-sample testing. Model performance was evaluated using the Root Mean Squared Error (RMSE), Mean Absolute Error (MAE), Directional Accuracy (DA), and the Diebold-Mariano test. The results confirm the existence of a long-run equilibrium relationship between oil prices and the fiscal surplus/deficit under both the ARDL and NARDL models. The NARDL model further reveals asymmetric effects of positive and negative oil price shocks. In terms of predictive performance, the hybrid 1D-CNN–LSTM model outperformed all competing models, achieving the lowest out-of-sample RMSE$ (4.008)$ and the highest DA $(0.636)$. The Diebold-Mariano test also indicates statistically significant superiority of the hybrid model over the NARDL and 1D-CNN models. These findings suggest that the hybrid 1D-CNN-LSTM model provides a more effective framework for modeling the nonlinear and dynamic relationship between oil prices and the fiscal surplus/deficit, making it a promising tool for fiscal forecasting and policy support in oil-dependent rentier economies such as Iraq.