| Option is a very active financial derivative.Whether it is the market maker trading system of option market or the important role of option in financial market,option pricing has become an issue of great significance and value.By assuming the factors that affect the option price,researchers construct paramet-ric option pricing models,such as Black-Scholes(BS)model,stochastic volatility model and so on.With the development of machine learning and deep learn-ing,some researchers try to apply these methods to the study of option pricing,and construct some nonparametric option pricing models.Nonparametric model does not consider complex assumptions,but directly fits the relationship between inputs and outputs.Its direction mainly includes two aspects:fit the implied volatility and directly fit.the option price.Based on the Boosting,this paper proposes a two-step optimization nonpara-metric option pricing model by combining implicit volatility fitting with direct option price fitting.The model is composed of two sub models.Through the individual optimization and serial connection of sub models,more accurate op-tion pricing is achieved.Specifically,we first use a multi-layer fully connected neural networks to fit the implied volatility surface,and then construct a Gaus-sian process regression model to predict the option price based on the estimated value of the implied volatility.In addition,we use BS model to connect the two sub models,that is,it is set as the mean function of Gaussian process regression model.This setting can not only better integrate the two-step optimization of the model,get a more accurate Boosting model,but also ensure that the model has the interpretability of the BS model to a.certain extent.and meets the basic financial criteria.The huge difference of option price affects the fitting effect of many nonpara-metric option pricing models.Gaussian process regression model uses the simi-larity of data to price options,so it can not be affected by these price differences.In addition,the prediction of Gaussian process regression model is directly ob-tained by calculation rather than training.Based on this characteristic,we adopt a step-by-step pricing prediction method,which can introduce more input infor-mation equivalently,and further improve the accuracy of option pricing without increasing the computational complexity.This paper makes some empirical analysis on the SSE 50ETF European call option and the TAIEX European call option(TXO)respectively.The results show that the proposed option pricing model is much better than the traditional BS model,and the error reduction percentage is 56.65%and 52.49%respectively.The accurate pricing of the two options shows the effectiveness and robustness of the model to a certain extent.By gradually remove the conditions such as the fusion of the implied volatility model,BS model as the mean function,and the improvement of the training method,we set up the corresponding control experiments for the model.The results show that the pricing error of the control model is higher than the whole model,which shows the rationality of model design.At the same time,we also find that the pricing error of the whole model is much lower than the single sub model,that is,the model proposed in this paper is far better than the neural networks model and Gaussian process regression model,which shows the effectiveness of the two-step optimization.Finally,we make an empirical analysis on intraday TXO call options.For more volatile intraday options,the model gets a better pricing results.We compare the pricing results of day options and intraday options in the same period of time,and find that the model achieves a lower pricing error on intraday options,which is due to the step-by-step pricing prediction method used in this paper.The validity of intraday option pricing reflects the higher practical value of the model.which can provide the theoretical option price with finer time granularity for the actual option trading market. |
PDF Full Text Request