How a Reinforcement Learning Agent learns and decides in Financial Markets

Abstract

This project asked what a learning agent in financial markets should be looking at. Working with an agent that allocates across twelve funds and makes one decision each day, early experiments showed that information about how assets move together contributed more to performance than any change to the learning algorithm, roughly tripling a baseline agent's risk adjusted return. That result motivated building a better source of structure: a map of how companies depend on one another, drawn from public filings by a language model and covering 169 firms. Six tests were run on the map. Linked companies moved together beyond what their sector membership explained, shocks propagated to named partners, and the map improved correlation estimates in fifteen out of fifteen years tested. Transmission was strongest when a documented cause was available, while unexplained moves behaved more like noise. The result is a practical and auditable way to give a learning agent better information about the economy.