| In recent years,with the development of the domestic financial market and technological advancements,the proportion of quantitative investment in the domestic stock trading market has been increasing,while the market environment has become increasingly complex.Many previously developed quantitative strategies may only be applicable to certain environments or simple market situations,and are not suitable for today’s complex market conditions,resulting in reduced returns.Whether there is a way to further improve these models and increase returns has become an important issue.Therefore,this thesis proposes using ensemble learning to integrate various types of basic market timing strategies,including not only machine learning but also traditional strategies and deep learning,to study the combination of several strategies with different complexities in an attempt to achieve higher returns.The main work of the thesis includes:(1)This thesis plans and implements the overall scheme design of ensemble learning algorithms.The overall design includes two parts:the algorithm and the backtesting system.The algorithm part explains the design scheme of the ensemble algorithm,including the construction optimization of various basic strategies and the selection and optimization of the ensemble strategy.The backtesting system describes the construction of the ensemble system and various backtesting indicators.(2)This thesis plans and implements the overall scheme design of ensemble learning algorithms.This thesis selects the daily trading data of the CSI 300 Index from April 2005 to February 2023 as the data sample,uses eleven machine learning algorithms to optimize the three best-performing machine learning strategies,constructs a traditional quantitative market timing strategy by optimizing the MACD moving average strategy,and constructs a more complex deep learning strategy using LSTM.All of the above strategies perform better than the CSI 300 returns,and it is found through indicators that strategy returns and complexity indeed show a positive correlation.(3)This thesis adopts a heterogeneous ensemble learning classifier for construction and analyzes the backtesting results.This thesis adopts two heterogeneous ensemble learning methods:hard voting and Stacking.In this stage of the experiment,the hard voting method constructs 11 strategy combinations,while Stacking constructs 26 strategy combinations.Both methods greatly improve the basic strategies,as evidenced by increased annualized returns,reduced maximum drawdowns,and higher Sharpe ratios.Through experiments,this thesis demonstrates that the quantitative market timing scheme based on heterogeneous ensemble learning has high practical value in market timing strategies.Compared to various basic timing strategies,its backtesting performance has improved in all aspects,especially in effectively increasing the return rate. |
