Latent Common Return Volatility Factors: Capturing Elusive Predictive Accuracy Gains When Forecasting Volatility

Abstract

In this paper, we use factor-augmented HAR-type models to predict the daily integrated volatility of asset returns. Our approach is based on a proposed two-step dimension reduction procedure designed to extract latent common volatility factors from a large dimensional and high-frequency returns dataset with 267 constituents of the S&P 500 index. In the first step, we apply either LASSO or elastic net shrinkage on estimates of integrated volatility of all constituents in the dataset, in order to select a subset of asset return series for further processing. In the second step, we utilize (sparse) principal component analysis to estimate latent common asset return factors, from which latent integrated volatility factors are extracted. Although we find limited in-sample fit improvement, relative to a benchmark HAR model, all of our proposed factor-augmented models result in substantial out-of-sample predictive accuracy improvement. In particular, forecasting gains are observed at market, sector, and individual-stock levels, with the exception of the financial sector. Further investigation of the factor structures for non-financial assets shows that industrial and technology stocks are characterized by minimal exposure to financial assets, inasmuch as forecasting gains associated with factor-augmented models for these types of assets are largely attributable to the inclusion of non-financial stock price return volatility in our latent factors.