Application of news sentiment analysis algorithm based on attention mechanism in portfolio rebalancing

The impact of news events on asset prices in the financial market is becoming increasingly significant, and the existing portfolio rebalancing strategy lacks in-depth mining of news sentiment. This paper proposes a news sentiment analysis algorithm (AANSA) based on attention mechanism, which extracts sentiment features through a three-layer attention network of words, sentences and chapters, and combines event impact assessment with dynamic weight adjustment to achieve intelligent rebalancing of the portfolio. The experiment uses market news from China and the United States from 2020 to 2025 and data on four stocks such as JD and TSLA. The results show that the accuracy of AANSA sentiment analysis is 92.3%, which is 15.7% higher than that of VADER algorithm; in portfolio application, the annualized rate of return is 18.6% (benchmark 13.4%), volatility is 14.3% (reduced by 8.3%), Sharpe ratio is 1.92 (increased by 123%), and maximum drawdown is -15.8% (reduced by 29.1%). In extreme events such as the adjustment of education policies in 2024, the algorithm responded 1-2 days in advance, reducing the portfolio loss by more than 12.5%. Experiments have shown that AANSA effectively improves the quantitative accuracy of news sentiment and the timeliness of investment decisions through hierarchical attention and dynamic mapping models, providing a new technical path for quantitative investment.