Metapopulation Differential Co-Evolution of Trading Strategies in a Model Financial Market

This paper reports results from experiments using differential evolution (DE) in a high-fidelity simulation model of a contemporary financial market in which various traders are each simultaneously trying to adapt their own trading strategy to be as profitable as possible, given the distribution of strategies currently deployed at that time by other traders in the market. In our model, each trader maintains its own private local population of trading strategies, and uses DE to adaptively improve its strategies over time. Because all traders are simultaneously trying to adapt their strategies, and because the profitability of any one strategy at time t can only be determined in reference to all other strategies also active in the market at time t, the system is _co-evolutionary_ rather than simply evolutionary. Furthermore, the existence of multiple separate DE populations in the system (i.e., one local DE population for each trader) means that technically this is a co-evolutionary _metapopulation_ system. Using DE in a co-evolutionary metapopulation context requires extension of the usual DE approaches used in less challenging applications, chief of which is the introduction of a mechanism to detect and actively prevent convergence within each local DE population. Results are presented which demonstrate that when all traders are using this nonconvergent DE, the overall economic efficiency (i.e., the sum of profitability over all traders) of the market is greatly higher than a baseline established when all traders were using a simple stochastic hill-climbing strategy optimizer instead of DE. Source-code for the experiments described in this paper has been released on GitHub as open-source, freely available for other researchers to use to replicate and extend the results presented here