Calibration of Models for Option Pricing Using Neural Networks

Abstract

This paper is focused on research into options pricing models. The most popular ones are variancegamma and Heston models. They are powerful and flexible to some extent, but they also have drawbacks. Among their shortcomings are instability and long-time calibration. The model proposed in the paper combines neural network (autoencoder) and relatively simple option pricing (mixture normal) model. The autoencoder gives flexibility to the model and reduces the number of parameters. The mixture normal model gives a certain logic to neural network and minimizes calibration time of the new model. So the resulting model eliminates drawbacks of variance-gamma and Heston models and also keeps their advantages. It has fast calibration and shows good enough precision on S&P futures options market.