From DeepStack to market microstructure
Martin Schmid, Matej Moravcik, and Rudolf Kadlec wrote DeepStack during a visiting PhD stint at DeepMind's Edmonton office. The system solved incomplete-information decision-making in real time via deep reinforcement learning, training across billions of simulated hands. The technical pivot from poker to markets rests on a shared mathematical structure: both domains penalize rigid heuristics and reward agents that reason under sequential uncertainty with an adversarial counterparty. EquiLibre's algorithms now execute across US equities and crypto, reportedly beginning in crypto in 2025 before expanding to stocks.
Supervised fitting vs. reinforcement learning
The architectural distinction matters. Most ML in trading remains supervised: fit to historical price sequences, predict the next return, iterate. DeepStack's lineage operates differently — the agent learns by acting, not by fitting. The reported zero-negative-months track record, if it survives independent audit, would suggest the approach holds in regime shifts where supervised models degrade. Creandum's commitment, unusual in scale for a European VC into quant infrastructure, signals institutional appetite for RL-native execution stacks rather than retrofitted signal libraries.
Verification checklist
Three variables determine whether the thesis is tradable intelligence or marketing:
- Attribution. "Billions in daily volume" is not Sharpe. Confirm notional vs. alpha-generating volume, and whether the figure represents gross turnover or net P&L.
- Survivorship in the claim. A zero-negative-months record over 18+ months is unusual but possible under aggressive risk caps. Audit the drawdown distribution, not just the monthly sign.
- Capacity constraint. RL strategies trained on self-play often degrade as AUM scales past a liquidity threshold. The $500 million raise increases pressure on the strategy to absorb capital without signal decay.
Watch for independent execution-quality disclosures, TCA reports, or third-party backtest audits. Until then, the edge remains unverified.