Trading industry taking increasingly quantitative approach to market testing

Dutch Authority for Financial Markets and London Stock Exchange Group are both developing agent-based simulation environments for testing.

The trading industry is taking an increasingly quantitative approach to testing trading strategies and maintaining regulatory oversight, with several initiatives expected to hit the market in the coming months.

Both the Dutch Authority for Financial Markets (AFM) and the London Stock Exchange Group are in the process of creating and launching agent-based testing models, The TRADE can reveal – aimed at bringing the current testing process and available testing environments more in line with the actual markets.

Agent-based models use real time data to simulate the markets as closely as possible, using next generation and artificial intelligence “agents” to react to strategies and hypothesis in a realistic way.

Speaking to The TRADE, capital markets data scientist at AFM, Rob Graumans, explained why the agent-based model was particularly important for regulators to evidence market manipulation.

“Those kind of insights are very valuable for us as a supervisor especially with regards to market manipulation because here you have to show that the actions of one agent affected the actions of other agents,” he said. “You have to show that other agents started trading differently because this agent acts in a certain way.”

The Dutch markets watchdog is in the development stage of its simulator as part of a partnership with The Alan Turing Institute and plans to have a prototype ready for the end of the year. AFM intends to open source the code and the modelling approaches it uses for the simulator meaning participants will be able to use their own data to create their own simulators.

“We think that it’s better to do so because if you open source you create transparency which creates trust. We think that is the way, that we as a supervisor at least, want to go,” added Graumans.

The London Stock Exchange Group (LSEG) is also in the process of trialling its own agent-based model for its clients. The exchange operator has partnered with fintech Simudyne to develop a next generation intelligence-based simulation environment.

LSEG said the agent-based approach to testing would allow participants to test algos and trading hypothesis and strategies by using real time data at a much faster pace and at a lower cost.

“The historic way of building and then testing a strategy in the most life-like conditions possible was to actually deploy that strategy into the market live,” Scott Bradley, head of securities trading sales and platform distribution, capital markets at London Stock Exchange Group, told The TRADE. “Using agent-based models allows for the move from what was conformance testing in CDS to now enable performance testing in the current CDS Plus trial offering.”

LSEG’s CDS Plus simulator is currently in an early-stage trial for member firms with 12 agent-based models available including for low latency trading, fundamental and momentum trading, and arbitrage. There are a sample of securities available to be tested using these and the exchange plans to include additional ones as the product develops.

“We see this becoming a gold standard of testing environment and process whereby it could be considered the destination of choice for this type of [algo] certification process and regulatory testing, aligned with RTS 6,” added Bradley.

Bradley also highlighted how the agent-based model could be used by the buy-side to evaluate algorithms within broker’s algo wheels and to simulate unprecedented market events like those seen in March 2020 at the height of the global pandemic to stress test institutions and safeguard for the future.

“How do buy-side firms evaluate the algorithms available for their algo wheel? It’s going to be through anecdotal evidence, it’s going to be through TCA data supplied by the brokers,” said Bradley. “CDS Plus could provide that level playing field opportunity within a high-fidelity environment for algorithmic strategies to be tested.”