Abstract
In this dissertation, we investigate and compare traditional hedging strate- gies to a Recurrent Neural Network (RNN) when hedging options including transaction costs. The traditional hedging strategies involve hedging accord- ing to market movements or methods derived from stochastic optimal control theory. Our implemented RNN uses the stock price, strike price, volatility, and RNN output at the previous time step as inputs to generalize a delta hedging strategy. The theoretical part of this dissertation investigates the underlying theory of these strategies. The practical component of this dissertation consists of a simulation study using a stock price model, where the initial stock prices, volatilities, and strike prices are sampled from an interval to add more variability to the simulated data set. This is done to determine if the RNN can generalize a delta hedging strategy for a data set that applies over a wide variety of random stock price paths. Each hedging strategy is then implemented and compared. For our simulated data set, the RNN hedging strategy minimized losses better than the traditional hedging strategies. Furthermore, it was found that using a constant boundary (more straightforward) hedging strategy can perform better than a more complicated one that depends on various param- eters and the option Greeks.
M.Sc. (Mathematical Statistics)