Improved Asset Pricing Models in Behavioral Finance Contexts and Applications in Asset Management Portfolios

Abstract

The accuracy of asset pricing models directly impacts asset management effectiveness. Traditional CAPM and the Fama-French three-factor model, due to their neglect of investor sentiment biases, underexplain irrational behavior in the A-share market, explaining less than 30% of return deviations in extreme market conditions. This paper proposes an improved asset pricing algorithm with sentiment bias correction (EBC-APM). This algorithm integrates BERT and LSTM technologies to construct a sentiment quantification module and dynamic correction factors, which are embedded into the traditional pricing framework to dynamically correct for sentiment bias. Experiments are conducted using 1486 trading days of CSI 300 component stocks and 2.86 million stock forum comments from 2018 to 2023. The results are compared with two baseline models on an Intel Xeon Gold 6338 CPU and an NVIDIA A100 GPU. Results show that the EBC-APM achieved a MAE of 1.03% and a RMSE of 1.42% on the test set, representing reductions of 44.62% and 44.09% compared to the CAPM, and 32.24% and 33.33% compared to the Fama-French model. The average error volatility coefficient under different market conditions was 0.19, a 45.71% reduction compared to the CAPM. In backtesting, the portfolio achieved an annualized return of 18.2%, 5.7 percentage points higher than the CAPM portfolio and 9.9 percentage points higher than the CSI 300 Index, effectively improving pricing accuracy and portfolio returns.