Simulating Artificial Stock Markets with Efficiency

Stock markets have been studied for years in economy and finance academic departments by relying on the idea of a general homo economicus that makes rational choices. Classical approaches use equational representation to enable a global markets characterization, but they fail to explain the link between individual behaviours and the global market dynamic and trends that emerge. Schelling seminal work on segregation models [12] has initiated a novel approach relying on an individual behaviour characterization that leads to global observations. This approach enables a finer grain behaviour modelling that allows more detailed simulations. Within this context, it is possible to have several kind of agents, with heterogenous cooperative or concurrent behaviours, and to study their actions aggregation and feedback loops which produces emergent macroscopic behaviours. The trade-off between the expressivity available with an agent-based approach and its execution speed relies on the complexity and information volume of involved processes. Our study is focused on order driven markets and describes why scalability issues appear when using an agent-based approach. After introducing our existing simulation platform ATOM, we study the impact of the number of orders and transactions and the cost of notification on execution time. These aspects lead us to search how we could optimize our platform to be able to handle higher volumes that are close to those generated by High Frequency Trading (HFT) algorithms. First section studies problematics linked to order-driven market simulators in a centralized setting. Second section identifies problematics that have to be tackled to scale agent-based simulations. Third section presents the two main approaches to efficiently distribute stock market simulations: network-based or GPGPU-based. Last section concludes and identifies future works.