Modeling and forecasting intraday spot volatility

We propose a multiple-equation regression-based method for modeling and forecasting intraday spot volatility. In this approach, intraday intervals are treated as individual time series, deviating from the common practice of treating the data as one continuous sample. Our empirical study, which spans more than two decades and encompasses six US blue-chip stocks, employs the recent OK volatility estimator developed by Li, Wang, and Zhang (2024) to expose the dynamics of latent intraday spot volatility over time. We demonstrate that the proposed method effectively captures the intricate dynamics of intraday spot volatility and find strong evidence that it outperforms a competing regression approach, and popular tree-based machine learning (LightGBM) and deep learning (LSTM) methods, in terms of predictive accuracy as measured by the MSE and QLIKE. These improvements in predictive accuracy extend to logarithmic extensions and across multiple forecast horizons. Overall, our results indicate that the parameter flexibility inherent in the proposed method is advantageous. This flexibility comes without undue computational burden.