Research And Implementation On The Quantitative Investment Strategy Analysis System Based On Machine Learning

Quantitative investment is a trading method that signals trading directives through computer program.Quantitative stock selection and quantitative timing are two main quantitative investment strategies.Quantitative timing is a kind of strategy aims to find the trend of stock market through quantitative analysis on a variety of macro and micro financial indicators,and signal the corresponding trading directives.Classification methods in machine learning are effective in predicting the trend of stock price movement.How to make improvement in stock feature selection and classification model construction,in order to increase the return rate of the timing strategy,is the main problem needs to be solved in domestic study on machine learning based quantitative timing strategy.To solve the problems above,this paper builds a machine learning based quantitative investment strategy analysis system,called Timing Quant(referred to as TQuant),on the basis of research on classification technology,financial technology indexes and quantitative investment strategy.TQuant consists of two main functional modules,stock time series database management and quantitative timing module(including Hurst based timing strategy,traditional trend based timing strategy and classification based timing strategy).System tests and commissioning show that TQuant is feasible and effective.There are three main features in this paper:1)Maximal information coefficient(MIC)based stock feature selection method.Stock transaction data,technical indicators,macroeconomic indicators are key factors for predicting stock movement trend precisely.After comparing traditional feature selection methods including principal component analysis,singular value decomposition and genetic algorithm,this paper proposes the MIC based stock feature selection method.Combining with financial field knowledge,this method calculates the maximal information coefficients of 48 stock features,and selects 11 stock features with highest coefficients values as the training features.The experimental results show that this method can achieve 4.5% higher prediction accuracy than other methods mentioned above.2)The tow-class classification model,SRA-Voting.Support vector machine(SVM)has strong generalization ability on small sample data,but is sensitive to missing data.Random forest(RF)and AdaBoost(AB)are insensitive to feature selecting,but RF inclines to decrease variance,AB inclines to decrease bias.This paper proposes a tow-class classification model,SRA-Voting.According to the stock data of different kinds and different time sections,and the characters of SVM,RF and AB,SRA-Voting assigns different weights to these three models dynamically,combines with voting method to predict stock price movement trend.The experimental results show that the prediction accuracy of SRA-Voting is about 3.125% higher than the accuracy of the above three models.3)This paper implements three main timing strategies,Hurst based,moving average based and classification based timing strategies.Hurst is computed by R/S method.Moving average based timing is implemented by simple two average signaling method.Classification based timing is implemented by SRA-Voting model,and defines three timing strategy signaling buy & sell directives in day,week and month periods.Tests and commissioning show that these three timing strategies are feasible and effective.