Investor sentiment and stock portfolio construction: A study based on the black-litterman model

Abstract

This study integrates the Black-Litterman (BL) model with deep learning techniques to construct a stock portfolio that incorporates investor sentiment. We develop a weighted investor sentiment index using comment data from the Eastmoney Stock Bar for eight representative stocks in the SSE 50 Index. The sentiment index is constructed through both a dictionary-based approach and the BERT model. Long Short-Term Memory (LSTM) networks are then employed to predict stock returns, which are incorporated as investor views into the BL framework for portfolio optimization. Empirical results demonstrate that incorporating investor sentiment significantly enhances stock price prediction accuracy, and the BERT-based sentiment index achieves the lowest prediction error. Within the BL model, the portfolio integrating the BERT sentiment index (BERT_BL) achieves an annualized return of 106.58% during the backtesting period, along with superior risk-adjusted performance metrics: a Sharpe ratio of 2.64, and a Sortino ratio of 4.18. The model remains robust even after accounting for transaction costs, parameter adjustments, and across different market environments. This study validates the application of textual sentiment analysis in quantitative investment, offering a methodological innovation by integrating unstructured data with classical asset allocation models. The findings provide practical insights for investors seeking to optimize portfolio management strategies.