Neural Networks in FX Derivatives Trading: Analyzing Artificial Intelligence Techniques to Price GBP/USD Options

Abstract

In the FX options market, the Garman-Kohlhagen (GK) model is the most widely used pricing model, but involves a set of unrealistic assumptions (e.g. constant volatility, etc.) that have been shown to cause systematic bias when compared the actual market prices. Using the GK model as a basis for comparison, this study explores three artificial neural networks (ANNs) models to price GBP/USD options traded on the Philadelphia Stock Exchange. These data-driven models have an advantage over parametric models since they make no initial assumptions and can adapt to changing markets. The first pricing model uses a standard ANN and is trained directly on market prices. The second pricing model utilizes a hybrid ANN that is trained to estimate the residual between the actual market price and the price estimated by the GK model. The last model uses a standard ANN to model the implied volatility curve on a given date in order to produce volatility estimates to be used in the GK model. The standard ANN pricing model is shown to perform worse than both the GK model and hybrid ANN model, while the hybrid ANN pricing model is shown to sometimes outperform the GK model in terms of mean squared error, but not mean absolute error. The implied volatility model is shown to perform equally as well as the GK model, while showing superior pricing performance for deep ITM and deep OTM calls/puts. The results of this thesis show that ANNs are more effective when used to model implied volatility rather than market prices, and demonstrate the potential of using ANNs to accurately price FX options through modeling implied volatility.