Pilot with authority for the financial markets

Developing an agent-based market simulator that can be used to test policy hypotheses

Project status



This project is a partnership between The Alan Turing Institute and the Dutch Authority for the Financial Markets (AFM) and aims to improve the stability and integrity of financial markets using modern data science tools.

Explaining the science

Generally, a simulation environment is needed to create agent-identifiable data which is lost in the flow of real-world orders. This project provides a unique opportunity to work with a real-world agent dataset, which in combination with modern technologies opens a range of new possible research questions.

Using this dataset, we will build models of individual-level trading activity, to form the basis of a market simulator. This simulator will then be used to understand the risks to market stability of changing trading patterns, to provide a testing environment for market participants, and to inform regulatory activity in financial markets.

From a research perspective, this is an exceptional opportunity to work with high-sensitivity data from financial markets, which allows us to go beyond the usual agent-based calibration patterns. There are challenges to be addressed in how to build models which preserve sufficient privacy of agents’ actions, which accurately reproduce market phenomena at both the individual trade and aggregate levels, and which can be simulated sufficiently quickly to be practically relevant for regulation.

Project aims

The project aims to:

  • Build a first draft simulator with interacting agents and test the accuracy against a small subset of data and known stylised facts.
  • Evaluate the changes of agents during periods of market stress. See how agents change their behaviour at that time and consider the implications for policy during market stress.
  • Develop a long-term project plan to address outstanding experimental questions, e.g., advanced stress tests, unique market conditions, and unintended algorithmic consequences.



There are many potential research questions that such a framework could be applied to: What is the effect of short-sellers on a market? Can a machine-learning algorithm (unintentionally) learn to manipulate the markets? What happens in the market when a large market maker leaves? How does the market react to agents becoming more “risk-averse”? In what scenarios can we expect to see abnormal behaviour by market participants? This project connects closely with existing work at The Turing on time series modelling, simulation, anomaly detection and privacy preservation.


Contact info

Monica Vakil-Dewar, [email protected]