A Stock Market State-aware Model Based On Concept Drift Detection

According to the "2022 US Fund Industry Yearbook",the global fund scale is growing rapidly.A group of leading fund companies with the largest asset management scale in the world,represented by Bridgewater Fund,widely uses computer technologies such as data analysis and artificial intelligence to help them make better investment decisions.This reflects that under the objective economic environment such as inflation and economic cycle,people’s demand for asset value preservation and appreciation is increasing.Quantitative investment has a considerable market size and broad development prospects.Quantitative investment is an important application field of computer technology such as machine learning and deep learning,using advanced artificial intelligence technology to participate in investment decision-making is a popular research direction.In recent years,the field of stock price series forecasting has achieved fruitful results.However,compared with other time series such as wind power and electric power,there is still a big gap in the effect of existing time series forecasting models applied to stock price time series.The stock price series has a very prominent concept drift problem,which means that it will continue to change with the development of the financial market.As a result,the data distribution and regular changes of the stock price series can cause the prediction performance of models to decrease or even fail.Aiming at the above problems,this paper proposes a stock forecasting framework based on concept drift detection,which is mainly divided into three parts:pattern division,concept drift detection and stock price trend prediction.Firstly,the extreme point method is used to divide the historical data of the stock price series into two types of data distribution corresponding to the rising trend and the falling trend of the stock price.Afterwards,building a Bayesian ensemble model to predict the data distribution of the next trading day,comparing it with the current distribution to determine whether concept drift occurs,and adjusting the data distribution of the training set of the stock price trend prediction module if concept drift occurs.Finally,the stock price trend prediction module takes the trend deterministic data preparation layer as the input feature of the model and builds a support vector machine to predict the rise and fall of the stock price in the next trading day.In addition,as a complement to the application and practical aspects of the model framework,this paper also presents two trading strategies in conjunction with a conceptual drift detection mechanism.This article selects 10 index component stocks and fund holding stocks respectively,a total of twenty representative stocks in the Chinese market as experimental dataset.The experimental results show that the model in this paper has reached or even surpassed the state-of-the-art such as FEDformer and VML in terms of accuracy,sharpe ratio,maximum drawdown and return.In the stock prediction framework proposed in this paper,the model will timely adjust the data distribution of the training set in the stock price trend prediction module when the concept drift detection module finds that the data distribution of the stock price series changes,so that it is more closer to the data distribution that changes with time,so as to improve the prediction performance of the model.Among them,this paper proposes a concept drift detection mechanism based on the extreme point method,which is more suitable for stock price sequences compared to traditional concept drift detection methods,and has better interpretability and learning ability.At the same time,in order to make the model better adapt to the data distribution changes of stock price series,the model of this paper adopts methods such as ensemble learning and dynamic adjustment of training sets.Most importantly,this paper emphasizes that making timely adjustments to the model when the data distribution of stock price series changes is a feasible idea to improve the existing stock forecasting model.