Volatility Forecasting and Value-at-Risk in UAE Equity Markets: GARCH versus Deep Learning

Abstract

This study tested for ARCH effects in UAE equity index returns and compared conventional econometric volatility models — ARCH, GARCH(1,1), EGARCH(1,1), and GJR-GARCH(1,1) — with two deep learning models, the Long Short-Term Memory (LSTM) and Gated Recurrent Unit (GRU) networks. Using daily index data from the Dubai Financial Market (DFM) and the Abu Dhabi Securities Exchange (ADX) from January 2010 to December 2025, all models were estimated in Python with Student-t innovations on the training sample to generate genuine one-step-ahead out-of-sample forecasts. Engle’s ARCH-LM test confirmed strong conditional heteroskedasticity in both markets. Forecast accuracy was assessed using RMSE and MAE, with significance evaluated through the Diebold-Mariano (1995) test, while Value-at-Risk (VaR) estimates were backtested using the Kupiec and Christoffersen tests. Once forecasts are evaluated on a correctly aligned out-of-sample basis, deep learning models offer no statistically significant accuracy advantage over the GARCH family. On the ADX, GARCH(1,1) and LSTM are indistinguishable while EGARCH outperforms LSTM; on the DFM, every GARCH-family model outperforms the LSTM. Significant leverage effects are documented in both markets, and the asymmetric EGARCH emerges as the strongest specification, passing both the Kupiec and Christoffersen tests. By contrast, the GRU exhibited a complete failure of variance stability on the ADX, producing near-zero variance predictions despite competitive forecast accuracy — a striking dissociation between point-forecast accuracy and applied risk measurement. The study highlights the strength of asymmetric GARCH models in emerging markets and shows that forecast accuracy alone is an insufficient criterion for model selection in risk applications.