A Study Of Option Pricing Based On Neural Networks Method

Financial derivatives play an important role in modern financial market. Option is the major one in fundamental derivatives. How to price the option accurately is always an important issue to many scholars.Option market is a highly complicated nonlinear dynamic system. Although the classic BS option pricing model has been considered as the crucial research achievement in recent 30 years, using it to price the option will still cause an obvious bias in comparison with the real market price. The bias is mainly due to the idealized assumptions which BS model is based on. However, artificial neural networks, as a significant non-parametric data-driven model, has been paid a lot of attentions and put into applications. It enables people to fully make use of the data. And without any restrictions and hypothesis, data can determine the model structure and parameters so that we will have a good pricing effect.This dissertation introduces the background knowledge of option pricing, the theory and the current development of neural networks in details. It presents the discussion of the structure design in neural networks model building. We take Jiangtong CWB1 as our research example and use Matlab software to build the option pricing models of BP neural network and wavelet neural network. Then we use the above two models to make one-step and five-step forecast. Finally we use MSE, MAE and MRE to evaluate the forecast accuracy of different models.According to the experimental comparison between neural networks model and BS model, we discover that neural networks model is superior to the BS model in the pricing precision. No matter using BP neural networks or wavelet neural networks, one-step forecast has a better performance than five-step forecast, which proves that neural networks is good at ultra-short-term forecast. On the other hand, no matter using one-step forecast or five-step forecast, wavelet neural networks model is superior to the BP neural networks model in a sense. The mean relative error of the wavelet neural networks one-step forecast which has the greatest pricing effect is only 6.7%.