Hybrid GARCH-LSTM Forecasting for Foreign Exchange Risk

Abstract

This study proposes a hybrid forecasting model that integrates the Generalized Autoregressive Conditional Heteroskedasticity (GARCH) model with a Long Short-Term Memory (LSTM) neural network to estimate Value at Risk (VaR) in the Rwandan foreign exchange market. The model is designed to capture both volatility clustering and temporal dependencies in daily exchange rate returns. Using daily data on USD, EUR, and GBP from 2012 to 2024, we evaluate the model’s performance relative to standalone GARCH(1,1) and LSTM models. Empirical results show that the hybrid model improves VaR estimation accuracy by up to 10%, especially during periods of elevated market volatility. These improvements are validated through MSE, MAE, and backtesting statistics. The enhanced accuracy is particularly relevant in emerging markets, where exchange rate dynamics are highly nonlinear and sensitive to external shocks. The proposed approach offers practical insights for financial institutions and regulators seeking to improve market risk assessment in emerging economies.