Multi-factor Stock Selection Scheme Design Based On XGBoost And LightGBM Algorithm

With computer science and technology's growing by leaps and bounds,quantitative investment has mushroomed in China.In the general sense,quantitative investment includes quantitative stock selection,quantitative timing,asset allocation and algorithmic trading.Compared with the traditional stock selection methods represented by fundamental stock selection and technical analysis stock selection,quantitative stock selection is relatively more objective and has the advantage of not being easily affected by investor sentiment and subjective ideas.The machine learning methods represented by integrated learning methods have also been effectively applied in the capital market.At the micro level,applying cuttingedge machine learning algorithms to securities investment,constructing stock selection models that can effectively improve investment efficiency and investment returns,and study specific factors that have a great contribution to the models have both theoretical and practical significances.At the macro level,in-depth research on quantitative stock selection methods will be helpful to broaden investment channels for investors,improve capital market pricing mechanisms,and improve capital market mechanism construction.Both XGBoost and LightGBM belong to the integrated learning model,which have the advantages of being efficient,accurate and easy to visualize in solving classification problems.This paper looks at stock selection from the perspective of classification,and systematically studies the stock selection ability of XGBoost and LightGBM.Firstly,use the grid search method to optimize the parameters and determine the optimal parameters.Secondly,perform annual and monthly rolling tests to compare and analyze the model effectiveness,decision tree structures,and factor importance of XGBoost and LightGBM.Thirdly,use industry-neutral stock selection strategy to deeply explore the effectiveness of XGBoost model and LightGBM model in the field of quantifying stock selection.Finally,under the framework of industryneutral stock selection,a single-factor test of 5 different integrated learning models including LightGBM,XGBoost,GBDT,Ada Boost and Random Forest was performed,and the differences in time efficiency and accuracy of the 5 models were compared.This paper uses the monthly frequency data and corresponding monthly return data of all 423 factors in 10 categories for all A-share listed companies from January 31,2008 to July 31,2019 for modeling analysis.The study found that:(1)The XGBoost and LightGBM classification models have good prediction capabilities,and the LightGBM classification model has a stronger prediction capability than XGBoost;(2)The XGBoost and LightGBM classification models have many parameters,but in actual tuning,there are very few parameters that can significantly improve the predictive ability of the model,and other parameters can just use the default values.(3)Integrated learning models such as LightGBM,XGBoost,GBDT,Ada Boost,and Random Forest all have good stock selection capabilities.LightGBM and XGBoost have better stock selection capabilities than other integrated learning models.(4)The LightGBM and XGBoost classification models are faster than other integrated learning models such as GBDT,Ada Boost,and Random Forest when applied to quantify stock selection,and the LightGBM model runs faster than XGBoost.(5)Factors such as Mkt Value(market value),GREC(change tendency of recommended rating score by analyst)and DAREC(changes of recommended rating score by analyst)have greater impacts on stock price returns,and Mkt Value has been the most effective factor in the market until 2017,of which importance far exceeds other factors but has declined rapidly after 2017.