Forecast of Yearly Stock Returns Based on Adaboost Integration Algorithm

摘要

Some existing studies use one kind of forecasting variables and Fama-MacBeth Regression to predict stock returns and find very modest predictability. To get better predictability, Adaboost integration algorithm, which is a classic machine learning algorithm and can use multiple kinds of forecasting variables, is introduced to predict yearly stock returns of all the firms of A-Share market from 2011 to 2015. The predictability is considerably improved. The average yearly out-of-sample error rate from 2013 to 2015 is 22%. The average yearly out-of-sample error rate from 2011 to 2015 is 27%. By analyzing data of the five years we find phenomena as following. First, the most important forecasting variable is different in each year when the market index shows different trend. Second, industry category plays very important role in every year. Meanwhile, other important forecasting variables of the five years include trade volume, full shares and tradable shares, Price Book Ratio and net asset per share, which represent market, size, valuation and fundamental respectively.