GBR And Other Six Regression Models Study The Multi-factor Arbitrage Of Digital Currency

Digital currency is the hot spot in the investment field at present.However,the huge volatility risk poses a difficult problem for investors.The multi-factor model reveals the change of return rate with a set of multi-dimensional factors,which provides a scheme for measuring assets.This paper takes Bitcoin,the representative of digital currency,as the research object,combines factor model in finance with ridge regression,Random Forest and GBR regression model in machine learning to predict future returns of Bitcoin,aiming at realizing multi-factor arbitrage in the digital currency market.In this paper,the time-frequency data of OKCoin International Exchange from July1,2017 to March 5,2019 are taken as samples.Firstly,data preprocessing and data distribution characteristics are analyzed.It is found that the right deviation of return rate presents a cusp distribution and has a trend effect.Second,the sample data are divided into the training set from July 1,2017 to December 6,2018,and data for December 2018 from 6 to March 5,2019 are for the test set.A multi-factor strategy is constructed in the training set.A factor pool is built based on 30 indicators such as moving average,Bollinger Bands,momentum and true amplitude.Regression method is used to verify the validity of single factor based on maximum withdrawal,sharp rate and IC value.After eliminating redundant factors according to factor thermodynamics,eight effective factors are determined,six models such as GBR are trained and learned.Optimal parameters of learning model.The test set predicts the yield of Bitcoin over the next three months.Finally,the validity of the evaluation strategy of the annual return rate,maximum withdrawal and other indicators is counted..The experimental results show that the performance of six regression models surpasses the real rate of return of Bitcoin,the results of CART tree,Ada boost algorithm,random forest and GBR under non-linear structure are better than Lasso regression and ridge regression under linear structure,and the prediction accuracy under random forest is the highest.Regardless of transaction costs,the cumulative yield of random forest in the next three months is 1.0915%,the annual yield is 4.43%,the maximum withdrawal rate is only 14%,the Sharp rate is 6.68,while the actual cumulative yield of Bitcoin in the same period is-0.0475%.