license.txt

Abstract

Trading in financial markets represents a large scale game where agents with varying levels of information interact. It has been shown that over time, in sufficiently liquid markets, information is disseminated through bid-ask spreads such that the traded value of a security converges to incorporate all available information. Informed traders make the greatest profit by hiding the information they know (preventing leakage) as long as they can through deceptive/camouflaging trades. Uninformed traders lose the least by learning information from the market as quickly as possible. This paper presents a model of sequential auctions that replicates the information flow in financial markets. The model is then used to train agents via reinforcement learning towards optimal policies. The experiment serves as proof-of-concept for trading as reinforcement learning and the ability of deep Q networks (DQN) to capture value from non-public information. This study further aims to answer the questions: 1) How best to learn information via market prices and 2) How best to hide information from the market.