LLM-Driven Market Reaction Prediction
Using financial markets as a natural labeling machine to train reasoning-based LLMs.
Built a full-stack pipeline that scrapes and structures corporate press releases from thousands of U.S.-listed company websites.
Used the market's own reaction — abnormal returns following each release — as a free, high-quality labeling signal, eliminating the need for manual annotation.
Queried an 8B-parameter LLM to predict market reaction (positive, negative, neutral) and provide an accompanying reasoning chain.
Selected reasoning chains that matched actual market outcomes and fine-tuned the model on those tokens to improve predictive accuracy.
Demonstrates a scalable ML paradigm: turning market behavior into labeled data for continuous model improvement.