| Since its listing,the Treasury bond futures have gradually played the role of price discovery and risk avoidance,which has played an important role in China's interest rate liberalization reform and the improvement of the Treasury bond yield curve.The high-frequency quantitative trading can enhance the liquidity of Treasury bond futures,enrich investors' asset allocation,and promote the core function of national debt futures better.In recent years,many scholars have focused on the research of high-frequency data and found that intraday high-frequency data contains more information about market microstructure and investors' behavior,which provides a theoretical basis for forecasting financial time series through historical information.However,because the high-frequency data contain a lot of information and are non-stationary,non-normal and have lots of noises,the traditional financial measurement methods have limited ability to predict the high-frequency data.Wavelet analysis,which is commonly used in engineering courses,can effectively smooth high-frequency data,while deep learning models can effectively extract effective information from a large number of data and make predictions based on their adaptability and self-learning habits.Therefore,this paper chooses the measures of wavelet analysis and deep learning models to study the intraday price prediction of national debt futures in China and provides new practical ideas for its quantitative trading.The research of this article is basically composed of 3 parts.First,in terms of research objects,this paper uses five-minute high-frequency data of the continuous main contract of 10-year Treasury bond futures with good liquidity in China.Through the descriptive analysis of the original data,it is found that the high-frequency data of Treasury bond futures have typical sharp peaks and thick tails,volatility aggregation and intraday effects.In this paper,the original data of national debt futures is decomposed in three layers by using the wavelet analysis method,and the data is reconstructed and restored by removing the high-frequency signal of the last layer.It is found that the reconstructed data effectively eliminates the noises in the high-frequency data of national debt futures.Secondly,this paper designed nine kinds of models contained a three layers of ANN model and the LSTM,GRU,Bi LSTM,Bi GRU models,and their combination model with the ANN model.The results found that by increasing the complexity of the neural network can effectively enhance the prediction ability,the best model is Bi GRU-ANN model,the NMSE on test set can reach 0.3159.Finally,this paper designs a strategy with a long-short position trading mechanism,uses various models trained in the previous step to compares the performance of each model in the strategy.Through the comparison of indicators such as trading signals,descriptive statistical analysis of the predicted value of the model and the return of the strategy,this paper finds that the model with high prediction accuracy triggers more trading signals,and the predicted value is nearly the same with the real rate of return,and also has more outstanding return in the strategy.The innovations of this article are: 1.The innovation of the research object.In this paper,different deep learning models are applied to the intraday price prediction of China's Treasury bond futures products,which expands the research of deep learning models in the price prediction of financial assets.2.The innovation of research content.By simulating the characteristics of Treasury bond futures trading,this paper designed relevant quantitative strategies to expand the research of Treasury bond futures products in quantitative trading.3.The innovation of research methods.Beacause high frequency data have a lot of noises and is difficult to predict,this paper adopts the method of combining wavelet analysis with different neural network models to expand the research on high frequency data of financial time series. |
PDF Full Text Request