Research On Application Of Machine Learning In Quantative Investment

Quantative investment(QI)is a collection of investment strategies that perform analysis and modeling by applying mathematical methods to financial market data.Machine Learning(ML),however,requires that a particular computer program would improve its capability of tackling a certain task by learning on given data.Apparently both ML and QI share the common of extracting information from data.Stimulated by recent success of ML,both industry and academe have grown interest in the research of QI-ML combination.This paper analyzed the pros and cons of ML in the scenario of QI along with introduction of relevant technique background,upon which predictive models on excessive return of stocks were constructed from deep learning methods.This paper used 'divide-concat' structure as general framework and long short term memory(LSTM)network as core to construct deep models,and multiple loss functions were tried to train these models.An automatic model rolling-refreshing mechanism was also implemented.Historical data from 2012 to 2018 in the stock market of China was used to train and test the models,and factors known as Alpha191 which are generated from price and volume data were used as model input.The back-test stage started from April 2014 and ended at November 2018,during which period,a model raised from this paper gained a cumulative return of 265.7% and annualized return of 32% from April 2014 to November 2018,and the CSI 300 index gained a far less competitive cumulative return of 43% and annualized return of 8.1%This paper,based on deep learning methods,implemented a research framework of 'data processing-model construction-predictive signals generation-trading strategy – back test analysis' and discussed a series of crucial issues including model structure under the framework of 'divide-concat' structure,model selection criteria based on the correlation of model output and its supervising value,model evaluation independent of back-test and trading strategy,etc.Several portfolio-oriented loss functions were proposed as an innovative exploration of loss functions in financial scenario,among which,the negative cosine loss function has proven to be practical and significantly stronger in the term of stability and return.The same empirical analysis also revealed the characteristics and limits of these loss functions.As one of the current focuses of investment industry,the trend of QI-ML combination and utilization strongly motivated the study of this paper.The work of the paper is highly practice-oriented and might provide references to further practices.