A CEEMDAN-LSTM Model for Forecasting the USD/DZD Exchange Rate

This study develops a hybrid forecasting model for the USD/DZD exchange rate by combining Complete Ensemble Empirical Mode Decomposition with Adaptive Noise (CEEMDAN) and Long Short-Term Memory (LSTM) networks to address high volatility and complex temporal dependencies in currency markets. Using 310 monthly observations, the CEEMDAN procedure decomposes the series into five frequency components and a residual, which are modeled by component-specific LSTM networks. The proposed CEEMDAN-LSTM model achieved the lowest forecast errors among the tested models, with a MAPE of 0.4782%, outperforming traditional LSTM and SVM benchmarks. The 12-month forecast suggests relative exchange rate stability, with a slight decline of about 0.48%. The results indicate that decomposing the original series before LSTM modeling improves predictive accuracy by separating short-term noise from medium- and long-term dynamics. The proposed framework may support exchange-rate risk management, hedging decisions, and short- to medium-term planning in emerging-market settings.