Investment decisions driven by fine-tuned large language models and uniform manifold approximation and projection-supported clustering and hierarchical density-based spatial clustering

This paper examines the investment signals generated from a combination of large language models (LLMs). The algorithm is applied to two sources of text: YouTube social media posts, and the business and economic news from more than 100 mainstream news sources such as CBS News, CNBC and Forbes. Investment signals based on news outlets outperform the Standard & Poor’s 500 (S&P 500) index, chosen as a benchmark, in relative and risk-adjusted terms, from September 2020 to April 2023. The fine-tuned LLM model clearly outperforms the base LLM model as well as a bidirectional encoder representations from transformers (BERT) model combination, which was chosen as a benchmark and comprises entity recognition from the BERT base model and sentiment analysis from FinBERT. Classification of LLM sentence embeddings with a novel approach using uniform manifold approximation and projection (UMAP) dimensionality reduction and hierarchical density-based spatial clustering of applications with noise (HDBSCAN) also generates investment signals that outperform the S&P 500 index on a risk-adjusted basis over the period from June 2018 to June 2023. All investment strategies are demonstrated to be unique by means of a standard regression against the Fama–French investment factors.