Abstract
This experimental study provides a comparative analysis of three forecasting approaches ARIMAX (econometric), XGBoost (machine learning), and LSTM (deep learning) for predicting the Libyan monetary base using actual monthly data from January 2004 to April 2026 (268 observations). All forecasts are based on actual historical data with no simulated or future values. Accurate forecasting of the Libyan monetary base is critical for economic stability in a fragile, oil‑dependent economy where forecast errors can have significant financial repercussions. To ensure fairness, all models receive identical multivariate input features (net foreign assets, net domestic assets, and lagged monetary base values). Previous studies have focused primarily on stable economies, leaving a gap in understanding forecasting methods applicable to volatile contexts such as Libya. The experiment follows a rigorous framework with a training period (2004–2020) and a testing period (2021–2026). Performance is evaluated using MAE, RMSE, and MAPE, and statistical significance is assessed using the Diebold–Mariano test. Results show that ARIMAX(2,1,2) substantially outperforms both XGBoost and LSTM, achieving a MAPE of 25.36% compared to 28.41% (XGBoost) and 49.46% (LSTM). The Diebold–Mariano test confirms statistically significant differences (ARIMAX vs XGBoost: DM = -5.97, p < 0.001; ARIMAX vs LSTM: DM = -13.47, p < 0.001). Feature importance indicates that lagged monetary base values dominate predictions, while LSTM struggles with structural breaks and limited sample size. Sensitivity analysis shows that a 12‑month window is optimal for both LSTM and XGBoost. This study contributes an unbiased, reproducible experimental benchmark for monetary forecasting in fragile, oil‑dependent economies.