Quality-Diversity for Automated Market-Making Under Regime Uncertainty

Abstract

Market-making performance depends strongly on market regime. Strategies optimized for calm markets can accumulate excessive inventory when volatility rises, while strategies optimized for volatile markets quote too conservatively in calmer conditions. We apply MAP-Elites to market-making strategy generation in an Avellaneda-Stoikov simulator with five stylized regimes. Across 250 algorithm-comparison runs, MAP-Elites remains close to CMA-ES on peak fitness: 1.4–3.3% lower in the four non-illiquid regimes and 11% lower in the illiquid regime, with no statistically significant aggregate difference (p = 0.087). Unlike CMA-ES, MAP-Elites also produces 55–60 behaviorally distinct strategies per regime. MAP-Elites significantly outperforms OpenES (p < 0.001) and PPO (p = 0.006), while achieving higher mean best fitness than random search in all five regimes. Cross-regime transfer experiments show that calm-trained archives transfer positively to three of four other regimes (p < 0.007), while volatile-trained archives transfer worst. Grid-resolution ablations show that QD-score increases with resolution (ρ = 0.88, p < 10−8) without reducing peak fitness. Overall, MAP-Elites retains near-CMA-ES peak fitness while producing diverse strategy repertoires under the same evaluation budget in this stylized setting.