What are the advantages of pricing American options using artificial neural networks?

Abstract

In this research work, I will try to see if it is possible and advantageous to use artificial neural networks to predict American options, which are more difficult to predict than European options because of the possibility of early exercise. As there are only numerical or approximation methods available, the neural network is a perfect fit, as it is non-parametric and able to capture some extremely complex non-linear functional relations. Once the neural network is set up with an efficient structure, it will also be possible to vary the input features to gain information on the real contribution of the latter on the efficiency of the model. This research includes definitions and demonstrations of key concepts in the field, a literature review of current knowledge, practices and trends on the subject, a construction of an efficient neural network structure to address the pricing problem and various feature tests on this network, each network being compared with its predecessors but also with the chosen benchmarks: the Black- Scholes model and the Cox Ross Rubinstein binomial tree model. The main conclusions of this research work are that, once the right neural network structure was found, the use of the ANN to predict American options consistently outperformed its benchmarks. What this means for managers is that machine learning, and neural networks in particular, may be worth investigating for implementation, especially in a context where there is access to sufficient data to train the network properly. My research also shows that if one is in a context where one has to predict in real time many American option prices, then neural networks are advantageous. Indeed, once the learning phase is over, the prediction is instantaneous, contrary to the iterative method of the CRR binomial tree. This advantage can be massive in attempts to leverage the pricing algorithms. In terms of conclusions for the academic side, this work shows that there is a need to continue to develop techniques for pricing American equity options using neural networks, and that one should not focus solely on European options in the belief that the latter are easier to tackle. I also demonstrate in this work that including the dividend yield in the neural network inputs increases the predictive power of the neural network. This parameter, too often omitted, can make a big difference by itself. My research also shows that taking the interest rate into account increases the predictive power, although a little less than the dividend yield, and confirms that volatility (in my case implied volatility) is very important in the input features. However, I also find that some features do not add value. This is the case for the volume, which once added to the network does not increase its predictive power (and also makes the training phase more complex), and the open interest, who deteriorates the results. So there are indeed advantages to using neural networks to predict American options. These advantages are the accuracy (by outperforming benchmarks such as the Black-Scholes or the binomial tree model), the taking into account of parameters that are sometimes difficult to integrate, the fact that a Put and a Call can be priced with the same algorithm, the fact that in the money, at the money or out of the money options can be priced efficiently with a single algorithm, and the instantaneous computational speed once the network has been trained However, there are also disadvantages, namely the learning phase can be long, the fact that one sometimes has to perform trial and error techniques to see what changes improve the network or not, the fact that one needs a lot of good quality data to train the network and the fact that a neural network is a black box that is difficult to analyse. It is up to each person to weigh up the pros and cons of each argument.