Financial Factors Mining Based On Bootstrap

In recent years,factor investment has been recognized by more and more fund managers and investors.The proportion of factor products including Smart Beta indexed investment and multi-factor strategy has also increased year by year,and the basis of all this is the mining of effective factors.The common method of factor mining is to construct factors based on the observed pricing error from the perspective of economics or behavioral finance.This paper attempts to find effective factors using data mining method and use multiple test frameworks to exclude the result of data fitting.It's a way to quickly find effective factors.This paper selects 92 X indicators and 11 Y indicators with high coverage from three financial statements,and uses the six factor-synthesis formulas to calculate 5850 financial derivative indicators.And then construct long and short combinations based on these 5850 indicators,and then use the long-short portfolio returns as the factor returns of the indicator.Finally,these factors were tested using a multiple-testing framework based on the bootstrap method to obtain 166 factors that were still significantly effective after excluding the influence of data fitting.In order to verify the effectiveness of this method,I divided the sample data into two periods of equal length,and then I tested the effectiveness by observing the factor validity transfer matrix,the difference of long and short portfolio and the performance of pure long-position portfolio based on the selected factors.From the factor validity transfer matrix,we can see that 50.65%effective factors in the early stage are still valid in the later stage,and this ratio is significantly larger than the unconditional expectation of the transition probability by 20%;30.50%effective factors in the previous period remain ineffective in the later stage and the ratio is also higher than the unconditional expectation.It indicates that the selected factors have good persistence.From the perspective of the difference between long and short portfolio returns,we can find that the monthly mean of the difference between the long and short portfolio returns is 0.58%,and the t-statistic is 5.2.That is to say,the selected factors do have a good discrimination.Looking at the performance of the pure long-position portfolio based on the selected factors,we can see the annualized return reaches 21.9%,the maximum retracement is 40.3%,and the Sharpe ratio is 1.621,and the performances are better than the performance of the CSI 500 Index.Therefore,factors can provide excess returns.In addition,the paper also analyzes the selected factors and finds that many of the top 100 effective factors are the four types of indicators:indicators related to income and profit,indicators related to advance receipts,indicators for payroll employees,and indicators related to changes in leverage ratio.The logic of such factors is relatively clear,and most of them are the year-on-year or chain growth rate of financial indicators.Such as payroll employee compensation,it's a good proxy indicator for employee compensation and benefits,and its growth rate reflects the expansion of the company's business.Therefore,the factor mining method in the paper can quickly find good factors from the massive indicators.At last,this paper also analyzes the source of factor returns from the perspective of behavioral finance.The results show that the factor returns selected by this method come from pricing error rather than risk compensation.