Youtube Analytics and Volatility Forecasting

Abstract

This thesis investigates whether finance-focused YouTube channels contain useful information for forecasting volatility beyond standard heterogeneous autoregressive (HAR) benchmarks. Using daily attention measures constructed from YouTube data, I examine whether these signals improve forecasts of realized and implied volatility in a rolling out-of-sample framework. First, I show that attention across channels is strongly correlated and can be summarized by a small number of latent factors. This motivates the use of both aggregate attention measures and PCA-based specifications in the forecasting models. Second, the results show that YouTube-based attention improves forecasting in a limited but meaningful way. The clearest improvement appears in one-day-ahead realized volatility forecasts. Here, the best PCA-based model achieves an out-of-sample R^2 of 0.439 relative to the HAR benchmark. At longer horizons, aggregate attention performs better, but with more modest improvements. By contrast, implied volatility is far harder to improve, with gains that are smaller and less stable. Overall, these results suggest that YouTube attention is not uniformly useful in volatility forecasting, but instead provides incremental gains that depend on the volatility target, forecast horizon, and representation of the attention signal.