Neural Network Models for Option Pricing on Non-Liquid Markets: an Empirical Evidence from Moscow Stock Exchange

The motivation of this paper is to investigate the prediction power of artificial neural networks in comparison to Black and Scholes model for option pricing when considering non-liquid derivatives market. Recently, deep learning techniques gained significant attention in finance, in particular, in financial derivatives pricing. However, they can be rarely seen to be used when non-liquid markets are considered. Using the data from Moscow Exchange the empirical study was conducted examining American options on futures in eight assets of different liquidity levels. It was found that neural networks are very effective in predicting prices of options with long time to maturity, even in case of limited amount of data available for training. The Quantile Regression Neural Network model appeared to be very effective for pricing options with long time to maturity including options on illiquid assets. For options with trading volume concentration closer to maturity the classical Black and Scholes model with GJR-GARCH-type volatility forecasts is a good approximation. In case of options on futures on USD/RUB exchange rate we the deviation from this rule was observed, as most of the options of this type are traded closer to maturity. This makes it more difficult to outperform Black-Scholes model. The study also finds that generated liquidity parameters increased neural networks performance accuracy for some of the considered assets.