Applying Generative Language Models for Portfolio Construction
Artificial Intelligence & Machine Learning
Francesco Fabozzi and Dr. Ionut Florescu’s recent paper, “Cut the Chit-Chat: A New Framework for the Application of Generative Language Models for Portfolio Construction,” provides an innovative take on sentiment analysis for financial applications, especially for those leveraging generative language models (GLMs) like ChatGPT. Traditional sentiment analysis methods typically rely on discrete sentiment labels (positive, neutral, or negative) to guide portfolio construction, assigning numerical values (1, 0, -1) to predict market movements. However, the authors argue that this approach oversimplifies sentiment, as it fails to capture its nuanced, continuous nature.
Fabozzi and Florescu introduce “Logit Extraction” (LE), a method that analyzes the underlying token probabilities generated by GLMs. LE allows for the creation of a more refined, continuous sentiment score, which significantly enhances the model’s predictive capabilities in portfolio construction. Their findings reveal that even less advanced models, like Mistral and Llama-2, outperformed more sophisticated models like GPT-4 when LE was applied. The LE method, implemented in the Python package “TokenProbs,” presents a cost-effective, open-source alternative that could revolutionize sentiment-based portfolio strategies by using granular, probability-based sentiment rankings.
The authors also explore LE’s potential beyond finance, such as in healthcare for patient outcome predictions or marketing for recommendation systems. This flexible framework could be a game-changer for categorical predictions in data-scarce domains, marking a shift from closed-source proprietary models to open-source tools that use probability-driven sentiment extraction.
