Stock Market Volatility Forecasting: Exploring the Power of Deep Learning

This study provides a comprehensive evaluation of five deep learning (DL) architectures—TiDE, LSTM, DeepAR, TCN, and Transformer—against the extended Heterogeneous Autoregressive (HAR) model for stock market volatility forecasting. Utilizing 22.5 years of high-frequency data from the S&P 500, DJIA, and Nasdaq indices and incorporating key macroeconomic variables (DXY, VIX, US10Y, and US1M), we assess predictive accuracy across multiple horizons from one day to one month. Our analysis yields three main findings. First, when macroeconomic variables are included, DL models consistently and significantly outperform the HAR benchmark, with TiDE excelling in one-day-ahead predictions and DeepAR dominating longer horizons. Second, in the absence of these exogenous variables, the statistical advantage of DL models over HAR often disappears, highlighting HAR’s enduring relevance in feature-constrained settings. Third, among the DL architectures, DeepAR emerges as the most robust and versatile performer, especially when leveraging macroeconomic data. These results underscore the conditional power of deep learning and provide practical guidance on model selection for financial practitioners and researchers.