Machine Learning for Remittance Forecasting and Macroeconomic Dynamics in Nepal: An Integrated Analytical Framework

Background: Remittances constitute over 25% of Nepal’s GDP, functioning as a critical macroeconomic shock absorber. However, remittance inflows are highly volatile and exhibit nonlinear dependencies on global commodity prices, exchange rate dynamics, and migration policies. Traditional econometric models such as ARIMA and VAR often fail to capture regime shifts, structural breaks, and complex temporal dependencies inherent in remittance series. Objectives: This study develops an integrated machine learning (ML) framework to: (1) forecast remittance inflows with superior predictive accuracy; (2) identify nonlinear macroeconomic drivers of remittances; and (3) quantify the dynamic macroeconomic impact of remittance shocks on GDP growth, inflation, and exchange rate dynamics. Methods: Using monthly data (2000–2023) from Nepal Rastra Bank, the World Bank, and the Department of Foreign Employment, we benchmark traditional ARIMA and VAR models against Random Forest, XGBoost, LSTM, and a hybrid LSTM-Attention architecture. Forecast evaluation employs rolling origin validation, RMSE, MAE, MAPE, and Diebold Mariano tests. SHAP (Shapley Additive Explanations) is applied for model interpretability. Structural analysis is conducted through an ML augmented VAR with counterfactual simulations and generalized impulse response functions. Findings: The hybrid LSTM-Attention model achieves a 45% improvement in forecasting accuracy (4.9% MAPE) compared to ARIMA. SHAP analysis identifies exchange rate volatility (inverted U-shaped effect) and oil prices (regime-dependent post-2015) as dominant nonlinear predictors. A one-standard-deviation remittance shock increases GDP growth by 0.35 percentage points (peak at 6 months) and inflation by 0.18 percentage points (peak at 14 months), revealing delayed inflationary transmission. Conclusion: Remittances exhibit a dual macroeconomic role stimulating short-term growth while generating delayed inflationary pressures. Policy focus should shift from managing remittance levels to managing remittance volatility and its structural determinants, particularly exchange rate stability and external commodity exposure. Novelty: This is the first study to integrate LSTM-Attention-based forecasting with structural VAR counterfactual analysis in a remittance dependent economy, demonstrating that machine learning enhances not only predictive performance but also structural macroeconomic inference.