QMML Market Making Hackathon

Competition Format

Hosted by the Queen Mary Machine Learning Society (March–April 2026), each round gave teams a training table (features + hidden price target). After fitting a model, teams submitted a bid and ask. The tightest spread won the market maker role; everyone else traded against those quotes once they were visible. The true price was then revealed and P&L updated. Negative equity meant elimination.

Strategy

Three principles structured our play:

• Never be the market maker. We posted a deliberately wide quote (bid 80, ask 400) every round so we would not provide liquidity to the entire field. The MM gets picked off when dozens of teams trade against one book.
• Bet big only with a verified edge. Stocks were bucketed by out-of-sample strength (R², RMSE, sample size). Strong linear fits (e.g. high R², few dominant features) earned larger sizes; weak or mean-only regimes were cut to minimum exposure.
• Survive first. Hard risk caps stopped any single round from wiping the bankroll — the discipline that showed up in the Sortino award.

During live finals we used t-distribution (small samples) and normal (large samples) confidence bands around predictions, adjusted for residual skew and kurtosis where relevant, and mapped probabilities into Kelly-style sizes with conservative floors and ceilings.

Modelling Notes

The nine synthetic assets behaved very differently: some were almost noiseless linear problems, one benefited from gradient boosting, several were effectively unpredictable (mean baseline), and one tiny high-R² dataset had an influential point we analysed explicitly (Cook's distance) before deciding to keep it for stability. The repo ships the per-round notebooks and the consolidated Python tooling for replication.

Repository

Open source on GitHub: training CSVs, round notebooks, strategy.py (pre-finals pipeline), kelly_based_strategy.py (interactive live sizing), and market_maker_prices.py (quote scenarios).