This research explores a cutting-edge framework that integrates financial news sentiment with machine learning-based Learning to Rank (LTR) algorithms to optimise stock portfolio selection. By using pre-trained NLP models like FinBERT and DistilRoBERTa, the study analyses daily, weekly, and quarterly sentiment data to rank stocks more effectively. Tested on 614 S&P 500 stocks, the approach demonstrates that long-only strategies outperform benchmarks like the S&P 500, while long-short strategies often return reduced risk at the cost of lower returns. Weekly rebalancing is shown to adapt best to market changes, striking a balance between responsiveness and stability. Concentrated portfolios yield the highest cumulative returns, emphasising the importance of focused stock selection in portfolio construction.
At the core of this framework is the application of LTR, a machine learning method commonly used in search engines and recommender systems, adapted here for financial markets. By leveraging the LambdaMART algorithm (a listwise ranking approach) and financial news sentiment indicators, the study captures complex stock interactions, leading to more precise predictions. This innovative integration of sentiment analysis and advanced ranking methodologies highlights a powerful new way for investors to achieve better risk-adjusted returns in dynamic market conditions.
Institutional
