BMLL’s 30 September agreement with Simudyne addresses a central problem in testing execution algorithms: the market should react when the algorithm trades. The companies are combining historical order-book datasets with Simudyne’s Pulse simulator. Simudyne said training of generative Large Market Models had begun, alongside its existing models of individual market participants.[1]
In a fixed historical replay, the recorded sequence remains the same even after a test strategy submits a large order. That is a useful control, but it cannot by itself show how other participants might respond to the additional trade. A reactive simulation changes the market state and generates subsequent behaviour. Its value depends on whether that response is credible, not merely on producing a realistic-looking price chart.
What the training data can describe
| Dataset | Information retained |
|---|---|
| Level 2 | Depth aggregated at each price level. |
| Level 3 | Individual orders and their insertions, changes, cancellations and executions. |
Order-level history provides a way to study queue behaviour and changes in displayed liquidity that a closing price cannot capture. BMLL retains exchange sequence numbers and order identifiers so an order’s lifecycle can be reconstructed. These features are directly relevant to modelling how a strategy reaches, changes or leaves the book.[2]
Pulse’s current FIX documentation describes a simulated venue receiving order messages and returning execution reports. A user starts a simulation, subscribes to its book and submits orders through a FIX engine, using a Pro or demo account with a configured session. A material testing limit is that cancel and amend requests are rejected: the original order remains working. This describes the existing interface, rather than completion of every element of the new partnership.[3]
Simudyne also publishes a separate validation interface, Pulse-Check, comparing simulations with historical data for the same instrument and date. It evaluates distributions, market-behaviour patterns and price responses to order events. Pro access returns aggregate comparison scores, while withholding the underlying historical series. Buyers therefore need to distinguish a validation result from a licence to inspect or redistribute the source data.[4]
The partnership uses BMLL’s Activate data-credit programme, which reduces or removes upfront data-licence costs during development through a defined allowance. That helps explain the commercial model: subsidised research access supports building a product before a broader customer deployment.[1]
An execution team can compare order schedules across reproducible scenarios and measure their effects on the book. The interface reports costs against a matching simulation that did not receive the test orders. Teams should also retain their own order and fill records: Pulse says its session-usage history stores parameters, not strategies or executions. Validation against subsequent trading experience then tests how useful the simulated results are.[3]
Sources & further reading
- BMLL and Simudyne partner to power AI market simulation with historical data BMLL · 30 Sep 2026 · Accessed 29 Oct 2026
- BMLL data and analytics — Level 1, 2 and 3 definitions BMLL · Accessed 29 Oct 2026
- Pulse FIX 4.4 documentation Simudyne · Accessed 29 Oct 2026
- Pulse-Check validation API documentation Simudyne · Accessed 29 Oct 2026