Research On XGBoost-ARMA-GARCH Model Stock Index Prediction Based On Bayesian Optimization Algorithm

The stock market occupies an important position in China's national economy,and it has a close relationship with the real economy.Its sharp rises and falls may have a significant impact on the real economy.At the same time,the stock market is characterized by high risks and high returns,which has attracted many financial institutions and individual investors to enter.As the "barometer" of the stock market,the stock index reflects the overall price level of the stock market and its changing trend,so the prediction of the stock index has profound theoretical and practical significance.Based on the existing literature,this paper uses the XGBoost-ARMA-GARCH model based on Bayesian optimization algorithm to predict the closing price of the CSI 300 Index.The XGBoost model was proposed by Dr.Chen Tianqi in 2016,and has the advantages of strong prediction accuracy,not easy to overfit,and easy to interpret.It can better solve the shortcomings of SVM,neural network and ARMA models.Because there are many important hyperparameters in the XGBoost model,this paper chooses the Bayesian optimization algorithm to adjust the model.The Bayesian optimization algorithm is different from the traditional grid search method or random search method.It uses the previously searched hyperparameters.The parameter information is used to determine the next sampling point.It stops iterating until the condition is met to obtain the optimal parameter group.It has the advantages of fewer iterations and small granularity.The construction idea of XGBoost-ARMA-GARCH model is: first use the XGBoost model to predict the closing price of the stock index,and then perform a white noise and stationarity test on the regression residual sequence.If the sequence is stable and not white noise,construct an ARMA ARCH test is performed on the residuals of the ARMA model.If there is an ARCH effect,a GARCH model is constructed to eliminate the conditional heteroscedasticity.Considering the impact of price indicators,technical indicators and other indicators on the CSI 300 Index,this paper selects 56 variables as alternative input features.Pearson correlation coefficient method and random forest algorithm were used to comprehensively select 40 indicators as input variables of XGBoost model.At the same time,this article will also build ARMA-GARCH,SVM-ARMAGARCH,GBDT,and BP neural networks as control models.The construction of SVM-ARMA-GARCH is similar to the XGBoost-ARMA-GARCH model.When the GARCH model is established,this paper considers three different cases where the residuals obey normal distribution,t distribution,and random error distribution.According to the AIC and BIC criteria,the most suitable residual distribution is selected for modeling.There are different evaluation criteria for the prediction effect of the model.Considering the comprehensiveness and objectivity of the evaluation effect,this paper chooses four evaluation indexes of RMSE,MAE,MAPE and accuracy to evaluate the model effect.Finally,this paper constructs a CSI 300 ETF trading strategy based on model prediction.Considering the three dimensions of the CSI 300 ETF,liquidity and tracking error,this paper chooses Huatai Barry CSI 300 ETF for backtesting analysis of trading strategies.According to the prediction value of the model in the backtesting set,and considering some transaction costs,this trading strategy has traded between the primary market and the secondary market,and achieved good results.The conclusions obtained in this paper are: 1.The ability of XGBoost to extract information is significantly better than SVM,GBDT,and BP neural networks,and the prediction effect of XGBoost-ARMA-GARCH model is significantly better than SVMARMA-GARCH,ARMA-TGARCH and other models;2.When tuning XGBoost model parameters,using Bayesian optimization algorithm can improve the accuracy and efficiency of tuning parameters,which is superior to traditional grid search method and random search method.3.The model-based prediction trading strategy constructed in this paper performs well in the backtesting set.Among them,the XGBoost-ARMAGARCH model-based trading strategy has the best backtesting results,with the least number of transactions,the highest win rate,the highest return,and the highest Sharpe ratio.