Fast Approximate Modular Inference for Option Price Prediction

Abstract

A new approximate Bayesian inferential framework is proposed that exploits multiple information sources -- daily spot returns, high-frequency spot data and option prices -- and enables fast calculation of probabilistic predictions of future option prices. This approach operates directly from the theoretical option pricing model, and does not require an explicit statistical model, or likelihood, for the observed option prices. We demonstrate that our approach achieves predictive accuracy comparable to likelihood-based modular inference, despite not explicitly modelling option-pricing errors, while delivering substantial computational gains. Predictive accuracy based on the Heston option pricing model is illustrated empirically for short-maturity options, with the rapidity of real-time updates of the predictive distributions highlighted.