The Impact of Sentiment Integration on Algorithmic Trading Performance: A Systematic Review of Multi-Modal Machine Learning Architectures

Abstract

Traditional algorithmic trading relies heavily on structured financial data, but in contemporary high-density information environments, this uni-modal approach is becoming insufficient. The purpose of this study is to evaluate the impact of integrating sentiment analysis from financial news and social media on algorithmic trading performance. I conducted a systematic literature review encompassing 80 primary empirical studies published between 2012 and 2026. The findings demonstrate that integrating textual signals with market data consistently enhances risk-adjusted returns, improving Sharpe ratios by 20–60% and reducing maximum drawdowns by up to 49.6% compared to price-only baselines. However, the informational value of sentiment is highly state-dependent, yielding significant predictive advantages primarily during high-volatility regimes and market crises. Furthermore, domain-specific transformer models (e.g., FinBERT, LLaMA) substantially outperform traditional lexicon-based methods. Despite these performance gains, the operational value of sentiment integration must be evaluated against realistic transaction costs and subjected to rigorous out-of-sample validation to guard against overfitting. I conclude that utilizing sentiment analysis as a complementary risk filter, rather than a primary signal amplifier, significantly improves strategy stability. Implementing multi-modal data fusion strongly supports building resilient trading systems.