Bullish, Bearish, or Just Meh? Fine Tuning LLMs to Beat Traditional ML at Financial Sentiment

In the volatile world of financial markets, understanding market sentiment from social media, whether bullish, bearish, or neutral, can make all the difference in both decision-making and algorithm design. To tackle this, I explored how fine-tuned LLMs can outperform traditional ML models in detecting nuanced financial sentiment.

What I did:

  • Built and benchmarked traditional ML models using a variety of techniques: Bag-of-Words, Word2Vec, Doc2Vec, SBERT, and open-source instruction-tuned LLMs (Qwen 2.5, Gemma 2, Phi 3).
  • Designed a Tester class for systematic evaluation with confusion matrices, accuracy, precision/recall, and error diagnostics.
  • Fine-tuned Qwen 2.5 7B Instruct using QLoRA, transforming it into a domain-specific “bullish/neutral/bearish” classifier with markedly improved precision and robustness.

Curious how the models stack up? Explore the full breakdown, methodologies, and learnings in my Medium article:
👉 Bullish, Bearish, or Just Meh? Fine-Tuning LLMs Against Traditional ML for Financial Sentiment

My code, modeling pipelines, relevant files, and data are all available in my GitHub repo:
👉 wallstreet_llm

2025

Back to top ↑