Building upon the findings of our previous research, this study further explores the role of Acuity’s lexicon-based sentiment scores in enhancing stock portfolio selection through machine learning. In Part I, we demonstrated how alternative data and Learning-to-Rank (LTR) models could effectively predict stock performance using LambdaMART, a powerful ranking algorithm optimized for financial data. This second phase expands on that work by incorporating sentiment scores from financial news, offering a deeper analysis of how sentiment-driven models can refine stock ranking and trading strategies. By integrating returns-related metrics and sentiment-based signals, we aim to enhance the predictive power of LTR models and improve decision-making in portfolio management.
Our study evaluates 614 stocks from the S&P 500, spanning 2009 to 2024, and applies long-only and long-short trading strategies to test the effectiveness of sentiment-based ranking models. Results show that lexicon-based sentiment models outperform transformer-based models in downtrend periods, while transformer models excel in longer-term, quarterly strategies. Additionally, weekly rebalancing adapts best to market fluctuations, whereas monthly and quarterly strategies provide greater stability but slower recovery during downturns. These insights reinforce the potential of sentiment-driven machine learning models in optimizing stock selection and risk-adjusted returns.
Hedge Funds