← All work

Quant · Systems · 2026

Microexec

A fast, deterministic market-microstructure simulator: a C++ matching engine exposed to Python through pybind11, wrapped in an event-driven loop that benchmarks RL execution against TWAP, VWAP, and Almgren-Chriss baselines.

Order-book depth, throughput trace, and execution-cost comparison from the Microexec simulator.
Order-book depth, throughput, and execution-cost comparison from the simulator.

Problem

Execution research needs a simulator that is quick enough to run constantly but precise enough that comparing two policies actually means something.

Approach

I wrote the matching engine in C++, exposed it to Python with zero-copy bindings, and built an event-driven loop around it running both baseline and reinforcement-learning agents on the same order flow.

Outcome

A setup where the speed, the determinism, and the policy results are all easy to check instead of buried. PPO execution came in about 12% under the baselines on shortfall in adverse liquidity.