A0765
Title: Forecasting stock returns using equi-correlation structures and component selection
Authors: Takayuki Morimoto - Kwansei Gakuin University (Japan)
Yohji Akama - Tohoku University (Japan)
Yoshinori Kawasaki - The Institute of Statistical Mathematics (Japan) [presenting]
Abstract: A novel factor modeling approach is proposed for stock return prediction in the Japanese equity market by utilizing dynamic equi-correlation structures derived from daily industry returns. The dynamic equi-correlation model is implemented to estimate time-varying equi-correlation coefficients across TOPIX industry portfolios and construct an industry equi-correlation (IEC) index by removing medium-term trends. Applying principal component analysis to these IEC series, latent forecasting factors are extracted. To determine the number of components to retain, four statistically grounded selection rules are incorporated: Broken-stick rule, adjusted correlation thresholding, Guttman-Kaiser, and cumulative percentage of variance, based on recent theoretical contributions by prior studies. The predictive power of the resulting IEC-based components is evaluated using panel regressions, and model performance is assessed via out-of-sample R-squared, Sharpe ratio, and certainty equivalent return. Results show that the correlation-based factors significantly outperform traditional accounting-based factor models in terms of forecast accuracy and economic utility. This framework highlights the value of correlation dynamics and dimensionally efficient factor compression in building interpretable and robust asset pricing models tailored to the Japanese market.
