66th ISI World Statistics Congress

66th ISI World Statistics Congress

Enhancing HAR-Based Volatility Forecasting with Transformers: A Hybrid Approach

Author

T
Carlos Cesar Trucios Maza

Co-author

  • D
    Dan Anibal Aguero Cerna

Conference

66th ISI World Statistics Congress

Format: WSC 2027: IPS Abstract

Keywords: deep learning, hybrid modeling, realized volatility

Abstract

Volatility forecasting plays a central role in key economic and financial applications, including portfolio allocation, financial and macroeconomic risk management, hedging, and derivative pricing. Consequently, a vast literature on volatility forecasting has emerged over the years, with the Heterogeneous Autoregressive (HAR) model, an approach exploiting information from high-frequency data, being one of the most successful and easy-to-implement approaches, consistently outperforming traditional models based solely on daily data. Despite its simplicity and remarkable success, the emergence of modern deep learning techniques raises a natural question: can modern deep learning techniques deliver superior forecasting performance than the widely adopted HAR model? To address this question, we conduct a large-scale empirical study comparing the HAR model against several cutting-edge deep learning techniques, including Long Short-Term Memory (LSTM) networks, Transformers, Multi-Transformers, and hybrid architectures. Based on both volatility and tail-risk forecasting, our findings indicate that while the HAR model is difficult to outperform using stand-alone deep learning techniques, hybrid approaches blending HAR specification with deep learning techniques consistently yield significant performance gains, offering a highly promising alternative for both researchers and practitioners.