BMLL and Simudyne have partnered to combine granular historical order book data with generative AI and agent-based simulation, creating simulated markets that can respond dynamically to trading activity rather than simply replaying historical events.
The collaboration brings BMLL’s harmonised Level 2 and Level 3 historical data into Simudyne’s Pulse intraday market simulator. Simudyne is using the data to pre-train what it calls Large Market Models: generative models of intraday order flow designed to reproduce market behaviour at the order-book level.
The companies see applications across algorithmic trading, transaction cost analysis (TCA), trading technology testing, exchanges and clearing and risk management. For trading firms, the immediate application is in back-testing execution strategies against a market that reacts to their activity, rather than replaying historical conditions that remain unchanged regardless of the orders being tested.“Historical backtests are static, the market will not move due to your backtest. Reactive simulations respond dynamically to orders injected into the market using a market-impact model that captures both transient and permanent effects,” Justin Lyon, CEO of Simudyne, explains to TradingTech Insight.
“Our reactive simulations offer two features: (a) testing execution strategies in market environments that respond directionally as if it were a live market while matching market microstructure features such as intra-book correlations and far-from-touch dynamics, and (b) providing a straightforward way to test stressed scenarios such as liquidity thinning, flash-crashes or other changes to market conditions.”
From rules to data-driven simulation
Market simulation has traditionally depended heavily on assumptions about the behaviour of different market participants. Simudyne itself has spent more than five years developing market simulators using conventional techniques, including agent-based modelling. The new approach adds deep generative models trained directly on granular order-book data.
The partnership – established through BMLL Activate Data Credits Programme, which provides selected technology partners with data credits and access to BMLL’s Data Lab and Data Feed during development – gives Simudyne access to BMLL’s normalised historical datasets across Level 2 and Level 3. At Level 3, that includes individual orders and their position and movement within the book, providing the models with a much more detailed representation of market microstructure than aggregated price-level data alone.
“These models depend on expert-specified assumptions about how participants behave, which can be difficult to validate against historical data,” says Lyon. “In the last ten years, generative AI has taken a purely data-driven approach to a number of different domains, including language, computer vision and even time series. Our deep generative models use granular L2 and L3 order-book data to make domain-specific foundation models for financial exchanges, without the need for expert-specified assumptions. We find these models are able to capture queue dynamics, conditional order-flow dependencies, and regime behaviour better than our previous models.”
BMLL will host Simudyne’s simulation models within its Data Lab, giving clients an environment in which they can run simulations against its historical order-book data.
Training algorithms inside simulated markets
The combination also opens up another potential application: using synthetic markets as environments in which execution algorithms can be trained rather than simply tested.
That changes the range of conditions available during algorithm development. Historical datasets inevitably contain only market regimes and events that actually occurred. Generative simulation can instead vary characteristics of the market and observe how an algorithm behaves as those conditions change.
“We have early internal results showing that our simulated markets can more effectively explore the cost-versus-risk frontier for execution,” says Lyon. “We will be extending this work to use our foundation models to train optimal execution strategies within a range of tuned conditions. One of the benefits of the foundation models is that it enables fine-grained control of market conditions such as the order book imbalance, liquidity, trend and volatility.”
For execution and TCA teams, the companies envisage using the environment to simulate large orders before they reach the market, estimating slippage, transaction costs and market impact while observing how simulated participants respond. Quants could also use the models as benchmarks when developing their own machine-learning execution strategies.
A test environment for trading infrastructure
The same simulated market can be connected to staging and testing systems through FIX or APIs, allowing technology teams to test end-to-end trading environments against a market that responds to their activity. The companies also envisage exchanges using the technology to test changes to matching engines, trading protocols, order types, auctions, circuit breakers and fee structures before production deployment.
Risk and clearing applications include stress-testing margin and default-fund models under simulated volatility and liquidity conditions. Simudyne says its market simulation models operate with sub-millisecond precision across equities, derivatives, options and futures.
“We believe that all these use cases are well suited to our models and, more importantly, represent a natural evolution of current methods towards a framework built around domain-specific AI where foundation models of financial exchanges unlock a new level of control and efficiency in capital markets,” says Lyon. “We are working closely with major exchanges to understand how these foundation models can be used to assess the impact of market-structure changes and to forecast liquidity risk under stress scenarios.”
The partnership puts historical market data into a different role from its conventional use in research, analytics and back-testing. Rather than functioning solely as a record of past market behaviour, granular order-book histories become training material for models designed to generate new market environments – environments in which orders, algorithms and trading systems can themselves affect what happens next.
Subscribe to our newsletter
