BRAIN-Agent

This project kills two birds with one stone: It helped me strengthen my AI agent orchestration skills while teaching me about Quantitative Trading fundamentals from the ground up.

I created a fully autonomous agent for researching algorithmic trading alphas and developing new algorithms to WorldQuant BRAIN. I selectively combined classical Machine Learning with LLMs to ensure the output did not drift, and is connected to reality. Specifically, I used an LLM to generate hypotheses and synthesize reasonings behind results, then populate a Knowledge Graph in the form of a Context Tree. I used Monte-Carlo Tree Search to guide future research direction in a deterministic and tunable way. I used Lopez de Prado scoring to weight each tree node based on various aspects of the submission results. In the end, I reached full autonomy, requiring zero human interventions to run a loop over 6 days, submitting high-quality alphas that landed me in the top tier of the WorldQuant BRAIN competition platform in the minimum amount of time.