Model Predictive Control for Trade Execution

In March 2026, Bayforest Technologies published a research paper, “Model Predictive Control for Trade Execution” in collaboration with Professor Dimitri P. Bertsekas (MIT / Arizona State University).
The paper can be found on https://arxiv.org/pdf/2603.28898.

The research focuses on a key issue for institutional investors: execution costs.
Inefficient execution can materially erode alpha before it reaches the portfolio, making execution quality a direct driver of performance. Our approach introduces a real-time execution framework that determines how to route and place orders dynamically, rather than relying on schedules such as VWAP or TWAP. It balances market impact, timing, and opportunity cost within defined risk parameters.

Unique approach:
– Speed and scalability. Each optimisation step takes approximately 1 millisecond, allowing the system to manage large numbers of simultaneous orders without latency-driven degradation.
– Modular architecture. The system is composed of distinct components — scheduling, cost modelling, fill probabilities, and risk controls — enabling transparency, testing, and continuous improvement.
– Explicit risk control. Two intuitive parameters govern behaviour: one controls adherence to the execution schedule, and the other controls the level of execution risk. Both are directly tunable and empirically calibrated.
– Integration of predictive signals. The framework is designed to incorporate short-horizon price signals. When directional information is available, the system can adjust execution to reflect it.
– Academic and practical foundation. The work was developed in collaboration with Professor Dimitri Bertsekas, a leading authority in dynamic programming and reinforcement learning, ensuring both theoretical rigour and practical relevance.

Demonstrated results:
The framework was validated using six months of NASDAQ Level 3 order book data, simulating approximately 170,000 parent orders across 1,200 instruments.
Relative to a standard spread-crossing benchmark, the results show the following improvements:
– Schedule shortfall: ~40–50% reduction
– VWAP slippage: ~25–40% reduction
– Arrival slippage: ~3–13% reduction
These improvements are consistent across multiple execution profiles (TWAP, VWAP, and Almgren–Chriss), demonstrating robustness across different trading styles. The framework is designed for live deployment across venues and brokers, integrating with standard execution workflows.

This work reflects Bayforest’s focus on delivering measurable improvements in execution outcomes and alpha preservation.