| Summary: | 碩士 === 國立交通大學 === 資訊管理研究所 === 107 === Option pricing has long been researched over the past years. In the past, the estimation of the underlying asset price distribution was usually resolved by statistical and stochastic processes. However, these traditional methods made some strict economic assumptions. These strict assumptions have proven to be inappropriate in past studies. The previous researches have already confirmed that the application of deep learning can correctly deal with option pricing after training with historical transaction data.
These models, which are trained from a large amount of knowledge in historical series data, do not require any premise assumptions. Such characteristics allow deep learning models to achieve superior performance in terms of pricing accuracy over traditional models. There are two different methods Based on this idea proposed in this article. The first method inherits the framework of the traditional Black-Scholes model. And by extending its framework to achieve a better pricing result. The second method estimates the underlying asset price distribution by learning the discrete implied distribution first then adjust by the option price directly.
The methods proposed in this paper are rolling test on the TAIEX options in 2017. When the rolling test is applied, there is an instability observed that one neural architecture can not deal with all time period. To achieve a more stable result, this article utilizes automatic deep learning techniques, which is called neural architecture search(NAS), and by using the training data as an input to the automatic neural architecture search, different architectures for the different time period can be generated through a single controller. The whole proposed pricing system can generate better pricing results calculated in mean absolute error in 2017. In addition to the pricing results, the pricing model proposed in this article is also used in an options trading strategy, and the model can achieve a better trading performance than the traditional
|
|---|
