Providing a Portfolio Optimization Model Based on Prediction Considering Investment Sentiment

Abstract

The purpose of this study was to develop an integrated model for stock return prediction and portfolio optimization by combining historical market data, technical indicators, and investor sentiment extracted from social media in order to improve investment decision-making accuracy in financial markets. This study was conducted using daily data from 16 selected U.S. stock market symbols during the period from September 30, 2021 to September 29, 2022. Price data, trading volume, and social media textual information were collected and preprocessed into 105 features, including 92 technical features and 13 sentiment-related features. Binary Grey Wolf Optimizer (BGWO) was employed for feature selection, while the XGBoost algorithm was used for stock return prediction. Subsequently, the predicted 31-day returns were incorporated into the Markowitz mean–variance optimization framework. Monte Carlo simulation with 100,000 admissible weight vectors was used to determine the optimal portfolio composition. Model performance was evaluated using MSE, RMSE, MAE, MAPE, and R² indices. The findings demonstrated that integrating technical indicators with investor sentiment features significantly improved prediction performance. The highest R² values were observed for KO and COST, with values of 0.8945 and 0.8915, respectively, whereas TSLA and BX showed comparatively weaker predictive performance. The BGWO–XGBoost framework successfully selected informative features and generated a stable low-redundancy predictive structure. In the portfolio optimization stage, only symbols with R² values above 0.70 were included in the Markowitz framework. Simulation results revealed that the optimal portfolio achieved an annual Sharpe ratio of 3.1585, an expected return of 0.7956, and volatility of 0.2424. The highest portfolio weights were assigned to CRM, PG, KO, and AAPL, indicating their superior risk–return balance. The results indicated that integrating machine learning algorithms with investor sentiment analysis and classical financial optimization frameworks can substantially enhance stock return prediction accuracy and investment decision quality. The proposed model successfully reduced the gap between theoretical financial analysis and real market behavior by generating a balanced, executable, and high-performing portfolio in terms of risk and return. These findings highlight the strong potential of behavioral data and text analytics in the development of intelligent investment management systems.