Markets as a Game - Finding Better Strategies For Automated Trading Of Cryptocurrency Using Multi Agent Systems
Budget: £750 – £1,500 GBP
Description of the assignment: (A single page, describing the assignment and indications of the techniques that will be explored)
The field of research in finance and economics has historically explored various kinds of
quantitative models for statistical inference of market data. But recent trends in Multi Agents Systems research have given it a whole new range of realism to which our knowledge is yet unexplored. Market dynamics are heavily influenced by the interest in the respective assets at any given point of time. The aim of this project is to leverage several state-of-the-art methods to find out the optimal trading strategies in any given market scenario.
The first phase of this project will be to defend the hypothesis also mentioned in [1], i.e. agents using fundamental analysis beat agents using technical market indicators more consistently. This will involve building a game simulation with virtual currency. Several populations of agents (dumb, hardcoded, stochastic, temporal, stateful and stateless) will be given a certain amount of virtual currency to begin with, all of which have a singular goal – to maximize their rewards either in the short term or long term for finding strategies optimal for varying investor risk profiles. In addition to the market data, agents will have access to the amount of trades being made at a certain time interval (which is indicative of the behavior of opponents at that time interval). For some of the agents, some degree of opponent modelling [1] will be incorporated to predict the behavior of the opponents at the next time step which in turn would be used to influence its decisions regarding
the market. Intelligent agents will learn the optimal strategies through self-play whereas agents using technical market indicators will use hardcoded strategies. A comparison of the different populations of agents will be made.
The second phase of the project will be to run real world trials of the best strategies obtained on the publicly available cryptocurrency market data, assuming sufficient quantitative overlap between stock market simulations and real world data exists. The goal is to validate whether our approach can reliably capture microstructure trends typical to a complex system like financial markets and whether multi-agent reinforcement learning could be used to implement psychological traits of decision theory and behavioral economics [2] in general. Optionally,
dynamic trading policies may be explored, with which agents will have to estimate how
fundamentalist or chartist it will have to be [2] in order to tune its behavior to real world data.
Overall, this project aims to combine breakthroughs in several recent domains (reinforcement learning, multi agent systems, game theory) to try and produce a better answer to the question "can a more optimal trading strategy be learned by using multi agent learning?"
References
1. Engle, Eric. (2008). The Stock Market as a Game: An Agent Based Approach to Trading in
Stocks. arXiv.org, Quantitative Finance Papers. 10.2139/ssrn.1270025
2. J. Lussange et al (2019). Stock market microstructure inference via multi-agent
reinforcement learning. arXiv.org, Quantitative Finance Papers, arXiv:1909.07748
The field of research in finance and economics has historically explored various kinds of
quantitative models for statistical inference of market data. But recent trends in Multi Agents Systems research have given it a whole new range of realism to which our knowledge is yet unexplored. Market dynamics are heavily influenced by the interest in the respective assets at any given point of time. The aim of this project is to leverage several state-of-the-art methods to find out the optimal trading strategies in any given market scenario.
The first phase of this project will be to defend the hypothesis also mentioned in [1], i.e. agents using fundamental analysis beat agents using technical market indicators more consistently. This will involve building a game simulation with virtual currency. Several populations of agents (dumb, hardcoded, stochastic, temporal, stateful and stateless) will be given a certain amount of virtual currency to begin with, all of which have a singular goal – to maximize their rewards either in the short term or long term for finding strategies optimal for varying investor risk profiles. In addition to the market data, agents will have access to the amount of trades being made at a certain time interval (which is indicative of the behavior of opponents at that time interval). For some of the agents, some degree of opponent modelling [1] will be incorporated to predict the behavior of the opponents at the next time step which in turn would be used to influence its decisions regarding
the market. Intelligent agents will learn the optimal strategies through self-play whereas agents using technical market indicators will use hardcoded strategies. A comparison of the different populations of agents will be made.
The second phase of the project will be to run real world trials of the best strategies obtained on the publicly available cryptocurrency market data, assuming sufficient quantitative overlap between stock market simulations and real world data exists. The goal is to validate whether our approach can reliably capture microstructure trends typical to a complex system like financial markets and whether multi-agent reinforcement learning could be used to implement psychological traits of decision theory and behavioral economics [2] in general. Optionally,
dynamic trading policies may be explored, with which agents will have to estimate how
fundamentalist or chartist it will have to be [2] in order to tune its behavior to real world data.
Overall, this project aims to combine breakthroughs in several recent domains (reinforcement learning, multi agent systems, game theory) to try and produce a better answer to the question "can a more optimal trading strategy be learned by using multi agent learning?"
References
1. Engle, Eric. (2008). The Stock Market as a Game: An Agent Based Approach to Trading in
Stocks. arXiv.org, Quantitative Finance Papers. 10.2139/ssrn.1270025
2. J. Lussange et al (2019). Stock market microstructure inference via multi-agent
reinforcement learning. arXiv.org, Quantitative Finance Papers, arXiv:1909.07748