Forecasting implied FX volatility surface has important applications in investment and trading, but it is a challenging task because the FX vol surface is composed of correlated high-dimensional time series and the shape of the surface can be significantly impacted by extreme market shocks that are almost impossible to forecast.
Vector Autoregression (VAR) has been utilized often in forecasting multi-variate time series, but has limitations due to high dimensional highly correlated data, which exists for this use case. We designed a new way to regulize the VAR model to improve its accuracy by training a regularized linear regression model (i.e.Ridge Regression) on the Embedded dataset, which we called Embedded Ridge (ER) model. The Embedded Ridge model was applied to forecast 8 FX implied volatility surfaces for 1 day, 1 week and 1-month forecast horizons.
Our results showed that the Embedded Ridge model outperformed the VAR model in the measure of accuracy, computations time and stability.