Prediction of stock returns based on cross-sectional multivariable model

A mathematical model is often used to manage stock portfolio of pension and investing funds. For example, a prediction model of stock returns, which can be constructed by analyzing a large amount of financial data, is used to select stocks incorporated into the portfolio. Two types of prediction models have been constructed by analyzing financial data. One is based on learning theory such as GA (Genetic Algorithm) and GNP (Genetic Network Algorithm). The other is based on ARMA model using historical price data. A problem with the former model was that it only predicts when a stock should be bought or sold, and thus it can not be applied to selection of investing stocks based on modern portfolio theory used practically because it doesn't predict return of stocks. A problem with the latter model was that it uses only stock price data to predict future returns and thus cannot achieve high accuracy. We propose a new prediction method of stock returns based on a cross-sectional multivariable model where explanatory variables are financial statement data, macro indexes and so on, and an explained variable is a future stock return. To achieve precise prediction, candidates for explanatory variables were first selected by using various test statistics including t-stats and adjusted R-square. Optimal combination of explanatory variables was then determined over time in order to maximize expected portfolio return and prediction accuracy. Coefficients of the prediction model were determined by regression analysis using historical data. A long-short portfolio, in which stocks with high predicted return are bought and stocks with low predicted return are sold short, was constructed to evaluate the proposed prediction method. The simulation test using historical data showed that the proposed method could achieve high portfolio performance, that is, an average annual portfolio return of about 20 % and information ratio of about 3.0. This also shows that the proposed method had high accuracy of prediction. Moreover an excellent property that an expected portfolio return is almost identical to an actual portfolio return was confirmed. © 2011 The Institute of Electrical Engineers of Japan.