Order-Book Foundation Model for Slippage & Fill-Probability Prediction

Customer

Challenge

• Asynchronous feeds: Depth and trade updates arrive at different cadences; signals must be robust to partial or delayed book views.
• Label construction: True fill likelihood depends on queue position and cancels; labels must be derived from reconstructed order-book events.
• Cross-symbol generalization: Learn patterns that transfer across instruments and regimes, with drift detection to avoid stale models.
• Confidentiality: Provide decision support without codifying client-specific order templates or thresholds.

Solution

Approach: A two-stage ML pipeline—(1) self-supervised representation learning on depth/prints sequences, then (2) supervised prediction heads for fill-probability and slippage—wrapped in a lightweight real-time service with guardrails.

Representation learning (foundation model)

Trained a sequence model to capture microstructure dynamics using tasks such as next-step prediction and masked-span reconstruction on order-book streams. Applied robust augmentations (e.g., simulated latencies, partial-book views) to improve out-of-distribution stability.

Task heads & calibration (generalized)

Fill-probability head: Estimates likelihood of execution for a proposed price offset and size over short horizons. Slippage head: Produces quantile estimates of expected short-horizon price impact. Uncertainty & abstention: Calibrated uncertainty scores drive an abstain-from-recommendation policy in ambiguous or stressed conditions.

Real-time service & controls

Low-latency inference service exposes a simple API to request predictions for candidate orders; outputs include confidence, risk flags, and rationale features. Guardrails: Venue/auction detectors, stale-data checks, and conservative fallbacks protect against anomalous book states or connectivity issues. Learning loop: High-error cases are automatically queued for review/retraining.

Technologies & tools

Process

1. Data contracts & rebuild — unify depth/prints; define labeling and QA checks.
2. Pretraining — train the representation on multi-month sequences with robustness augmentations.
3. Supervised heads — add fill-probability and slippage predictors; perform time-aware validation and calibration.
4. Service & guardrails — deploy the inference API with drift monitors, abstention, and venue/auction safeguards.
5. Pilot & iterate — A/B against baseline heuristics; integrate feedback; schedule retraining and release gating.

Results

• Decision support: Provided probabilistic guidance for offset/size selection, improving consistency over static rules.
• Operational safety: Abstention and guardrails reduced recommendations during unreliable data states or auctions.
• Scalability: A single representation generalized across symbols/venues, with retraining workflows aligned to regime changes.

(All numerical improvements, parameter values, and venue-specific outcomes are intentionally omitted.)