Forecasting realized volatility with the implied volatility surface: an image-based approach

This paper explores how to extract information about an asset’s future risk from the entire implied volatility surface (IVS).We convert the IVS into a standardized matrix and model it with an image-based approach – namely, a convolutional neural network (CNN) – thereby establishing a direct link between the IVS and future realized volatility (RV). Moreover, the forecasts generated by the CNN model, designated the CNN-IVS factor, are integrated into a heterogeneous autoregressive model of realized volatility (HAR-RV) framework. The experimental results demonstrate that the extended HAR model with the CNN-IVS factor significantly enhances out-of-sample performance for RV forecasting. Further interpretive analyses show that the CNN model automatically extracts predictive information from the overall structure of the IVS, captures economically meaningful signals relating to jump risk and downside risk, and provides unique incremental information for volatility forecasting.