| The cross period arbitrage of futures refers to the arbitrage by using the fluctuation of the price difference between the distant month and the near month.Investors can make an appropriate arbitrage plan through the prediction of the cross period spread.However,a large number of studies show that the price difference series not only have the characteristics of mean reversion,but also generally have the characteristics of nonlinear and asymmetric,so it is very difficult to predict the cross period price difference.Not only that,there is not much regularity in the fluctuation of price difference.Because the non-linearity of financial time series is not considered,the traditional regression model is difficult to predict the price difference accurately.The emergence of SVR provides an idea for the prediction of cross period spread.It can transform the price difference prediction problem into regression problem,and can effectively use the high-dimensional mapping of kernel function,so as to avoid the problem of "dimension disaster" brought by traditional methods.At the same time,it can effectively prevent the "over fitting" phenomenon of the model through the introduction of relaxation variables.Therefore,this paper introduces SVR,and takes the 30 minute closing price of the current month's continuous contract and the next month's continuous contract of CSI 300 stock index futures as the research object to study the inter period price difference of stock index futures.At the same time,in order to improve the prediction ability of the model,this paper uses the time series perceptual cross validation to select the better kernel function,and optimizes the parameters of SVR model through genetic algorithm,then constructs GA-SVR-GARCH model according to whether there is arch effect in the residual,extracts the high-dimensional nonlinear relationship with SVR,and adds the linear relationship with GARCH to determine the relationship between the two price differences.At the same time,this paper will measure the effectiveness of the scheme from two aspects: on the one hand,it focuses on the prediction accuracy of the model;on the other hand,it tests the effectiveness of the scheme through simulation transactions.In terms of model prediction accuracy,this paper selects the transaction data from January 1,2015 to June 30,2019 as the sample interval.The training model and the validation model adopt the rolling prediction method.Through the comparative study of GA-SVR-GARCH and PSO-SVR-GARCH,ARIMA-GARCH and GA-BPGARCH,it is found that in the aspect of algorithm optimization,genetic algorithm is better than particle swarm optimization in the ability of super parameter optimization,this is better because particle swarm optimization algorithm is easy to fall into local optimal solution.In terms of model,SVR-GARCH model has higher prediction accuracy than ARIMA-GARCH and BP-GARCH.This is because SVR-GARCH model enhances the ability of model to analyze the nonlinear and complex structure of time series data,and overcomes the heteroscedasticity effect of SVR model in the residual series of prediction results by combining with GARCH model should be inadequate.In the empirical aspect of simulated trading,this paper selects the trading data from July 1,2019 to November 30,2019 as the back testing simulation interval,uses GA-SVR-GARCH model to build the arbitrage strategy,and compares it with the arbitrage strategy constructed by standard distance method,PSO-SVR-GARCH model,ARIMA-GARCH model and GA-BP-GARCH model,and finds that GA-SVRGARCH model can achieve better arbitrage strategy At the same time,considering the prediction error of GA-SVR-GARCH model,the threshold limit of standard distance method is introduced on the basis of the model,which further improves the winning rate of the arbitrage strategy of GA-SVR-GARCH model,reduces the maximum withdrawal of the strategy,and achieves good returns,which provides certain help for the construction and decision-making of cross period arbitrage strategy. |
PDF Full Text Request