Forecasting Export Performance in Emerging MENA Economies: A Comparative Study between Panel ARDL and Bi-LSTM Neural Networks

Abstract

This study examines the determinants of export performance in emerging MENA economies during the period 1985-  2024, comparing the forecasting accuracy of traditional econometric models with that of Bidirectional Long Short-Term Memory (Bi-LSTM) neural networks. Using panel ARDL estimation, the empirical findings confirm a long-run cointegrating relationship among exports, exchange rates, inflation, and GDP. The exchange rate has the strongest positive effect on exports (0.5504, p < 0.01), while GDP has a negative long-term impact (-41.874, p < 0.10). These results support the Domestic Absorption Hypothesis. This negative effect is more pronounced in oil-producing countries, reflecting a shift in resources toward domestic consumption.

The Bi-LSTM model substantially outperforms traditional approaches, achieving superior predictive accuracy, as evidenced by R² = 0.9016, MAE = 0.0103, and RMSE = 0.0135. Forecasts for the period from 2025 to 2029 reveal significant heterogeneity across countries: Bahrain maintains the highest export levels, while Egypt experiences the sharpest decline, reflecting structural constraints. This study provides empirical evidence on the long-term determinants of exports in the region and shows that Bi-LSTM networks are better at capturing nonlinear export dynamics. The findings highlight the importance of country-specific policies to strengthen export resilience and diversification.