Deterministic algorithmic execution strategies, while statistically efficient, are vulnerable to exploitation by adversaries who can infer and anticipate order flows.
This project develops a framework for stochastic execution policies to defend against such adversarial learning dynamics.
We evaluate three randomization mechanisms:
- Uniform randomization
- Ornstein–Uhlenbeck (OU) mean-reverting noise
- Pink (1/f) spectral noise
using both:
- static machine-learning binary classifiers, and
- regression-based adversaries predicting execution prices.
According to the paper’s main findings :
- OU policy reduces adversarial predictability by 27% (AUC 0.78 → 0.57)
- OU improves implementation shortfall by 14.8 bps, driven by 10.6 bps decrease in adverse selection
- Portfolio Sharpe improves 19% (0.94 → 1.12)
- OU reaches adversarial safety in 5 iterations (vs 15–20 for alternatives)
- Estimated benefit: $37M annual at $100M daily execution volume
- All policies satisfy strict exposure and turnover constraints during backtests
This establishes that adversarial robustness requires dynamic, game-theoretic adaptation, not static defense.
| Name | Role | Profiles |
|---|---|---|
| Vincenzo Della Ratta | Infrastructure, Pipeline, Backtesting | GitHub · LinkedIn |
| Preslav Georgiev | GitHub · LinkedIn | |
| Matteo Roda | GitHub · LinkedIn | |
| Rayi Makori | Project Lead | GitHub · LinkedIn |
| Hunor Csenteri | GitHub · LinkedIn | |
| Neel Roy | GitHub · LinkedIn | |
| David Livshits | GitHub · LinkedIn |
Full Paper (PDF): attached in repo (docs/BSML_final.pdf)
BSML Website: https://bsmachinelearning.com