A core element of the AlgoFinance research project is to develop a realistic agent-based model of algorithmic finance. The central aim of the model work is to simulate and thus get a better understanding of interactions among algorithms in markets. As our modelling objective is to create representative simulations of financial markets, our output will also look very much like that of an actual market: trades will occur and generate entries in a ‘Time & Sales’ history; orders that agents place but are not immediately matched will comprise a ‘Limit Order Book’; and as a large proportion of orders are cancelled in contemporary markets, our simulated market’s matching engine implements an audit trail to record those as well. The framework allows for agents to trade multiple financial instruments at one or more simulated exchanges.
The framework is ‘event driven’, as the simulation progresses by processing messages sent between the simulated market actors (i.e. the agents and the trading venue). The messaging system is based roughly on the FIX protocol. In the most general definition: the output of our models is the state of the whole system at each moment in time. Along with the Limit Order-Book objects and Time & Sales events generated at the matching engine, the state also includes the values of the internal variables of every agent, along with their individual positions and profit history. In line with our research objectives, this highly descriptive data allows us to explore emergent phenomena and investigate possible causes. While this is not guaranteed to lead to asset price predictions, analyzing results across many ABM configurations and parameterizations might produce testable hypotheses regarding important questions related to market structure, regulation, fragility, and dynamism. We are continuously refining the simulation framework on basis of insights generated from sociological and anthropological fieldwork as well as from order-book data.
One of the central feats of our simulation framework is that we can assign latencies to all agents, meaning that we can differentiate between faster and slower agents at all latency levels. Through collaboration with Euronext, we have obtained data about current actual latencies (and latency hierarchies) for the main types of market participants in present-day markets, something that greatly enhances the realism of our framework.
Last updated by: Administrator User 16/12/2019