Robust market making under persistent directional order flow

Establish robust profitability for market-making agents trained in stationary order-flow simulators when exposed to persistent directional flow, overcoming the fragility that is not resolved by simply increasing network size or training-data volume.

Background

The paper identifies robustness under non-stationary order flow as a central unresolved challenge. Stationarily trained market-making agents, including both the GLFT benchmark and the Rainbow distributional DQN controller, can accumulate inventory at their hard limits during persistent buy- or sell-pressure regimes, producing adverse-selection losses and large drawdowns. The proposed Bayesian regime-belief augmentation, queue-adjusted quote-exposure signal, and scenario-bandit fine-tuning substantially improve performance, but the authors continue to characterize robustness as an open problem rather than a fully solved one.

The unresolved scope includes behavior beyond the paper's calibrated zero-intelligence Santa Fe simulator and beyond its fixed regime width, exponentially distributed regime durations, unit-sized quotes, and single-level quoting. The paper's stress tests show that mean profitability can be preserved, but downside tail risk remains under persistent and correlated directional regimes.