Document Type : applicative
Authors
1
Department of Finance, Imam Sadiq University, Tehran, Iran
2
Faculty Member, Faculty of Economics, Faculty of Economics, Imam Sadiq University, Tehran, Iran.
10.30473/jier.2026.78746.1554
Abstract
Accurate volatility forecasting underpins risk management, capital allocation, and value-at-risk assessment, yet the nonlinear, regime-dependent behavior of volatility in the Iranian stock market complicates the choice between classical and deep learning models, with direct implications for industrial risk management. This study models and forecasts the logarithm of daily variance, used here as a close-price-based proxy for realized volatility given the absence of intraday data, for two Tehran Stock Exchange indices, the TEDPIX and the Equal-Weighted index, over 2015 to 2025. Three classical models (ARIMA, GARCH, HAR-RV) are compared with four deep learning architectures (LSTM, CNN-LSTM, Transformer, PatchTST-lite) across three horizons of 1, 5, and 22 days, using MAE, RMSE, and the robust QLIKE loss, together with the Diebold-Mariano and Model Confidence Set tests. Results reveal a systematic divergence across evaluation spaces: ARIMA and HAR-RV attain the lowest errors in log-space (MAE, RMSE) but the worst QLIKE, since they underestimate variance during volatility spikes, which QLIKE penalizes severely. Under QLIKE, GARCH attains the lowest numerical error at short and long horizons, while deep learning models, particularly the Transformer, show a numerical edge at the medium horizon; the Model Confidence Set, however, shows none of these differences are statistically significant, with GARCH remaining competitive in all six index-horizon combinations. Deep learning models rely almost exclusively on recent information and degrade sharply during the volatile 2018-2020 boom regime. Overall, no model family dominates, but GARCH offers a robust, consistently competitive benchmark, while QLIKE proves essential for risk-sensitive model selection.
Keywords
Main Subjects
)