Traders that deliberately ignore some data beat those that don't: in high-fidelity market simulations, reinforcement-learning agents that mask noisy microstructure signals produce higher risk-adjusted returns and temper destabilizing herding, suggesting selective blindness can be a practical tool for improving alpha and market stability.

Main Finding

Agents that deliberately limit or mask incoming information—strategic ignorance—outperform full-information agents in noisy, high-frequency market environments. Under finite Shannon-capacity constraints, selectively ignoring/transitively masking microstructure noise yields higher risk-adjusted returns (higher Sharpe ratios), lower maximum drawdowns, and reduced overreaction and herding, with the benefits concentrated in high-volatility regimes. Reinforcement-learning agents trained to dynamically mask noisy features via a Kalman-enhanced framework learn to prioritize persistent fundamental signals over transient fluctuations, revealing “less is more” as both a performance and market-stability mechanism.

Key Points

• Strategic ignorance is modeled as deliberate filtering, delay, or omission of information streams under a processing-capacity constraint (finite Shannon capacity).
• Combining rational inattention, information design, and multi-agent RL produces agents that learn when to ignore microstructure noise to improve decision quality.
• Selective-ignorance agents reduce feedback-driven phenomena: overreaction, herding, and volatility amplification.
• Reinforcement-learning agents with a Kalman-enhanced masking module learn dynamic, state-dependent feature masks that emphasize persistent fundamentals.
• Benefits are most pronounced in high-volatility regimes and fragmented markets where transient signals dominate microstructure noise.
• Strategic blindness is a source of orthogonal alpha (returns not explained by standard signals), and also contributes to systemic market stability.
• The paper supplies a practical system architecture for implementing selective-ignorance filters in institutional trading stacks.

Data & Methods

• Environment: high-fidelity simulated limit order book (LOB) market used to emulate microstructure, order flow, execution, and feedback among heterogeneous agents.
• Agent setup: multi-agent reinforcement learning (MARL) framework where agents interact, post orders, and update policies under realistic latency and liquidity constraints.
• Information constraint: agents are subject to a finite Shannon-capacity model that limits information throughput/processing; this formalizes rational inattention costs.
• Selective-ignorance implementations compared:
  • Full-information baseline (no masking).
  • Static filters (hand-crafted feature thresholds).
  • Learned dynamic masks (RL agents that choose which features to observe or weight).
  • Kalman-enhanced masking: a state estimator (Kalman filter) augments the RL agent to separate persistent fundamental signals from transient noise before masking decisions.
• Training & evaluation:
  • Policies trained in multi-agent LOB simulations across multiple volatility regimes and noise levels.
  • Performance metrics: Sharpe ratio, maximum drawdown, realized volatility, measures of herding/market-impact, and return attribution.
  • Robustness checks: parameter sweeps over channel capacity, noise intensity, agent population composition, and alternative masking baselines.
• Principal empirical findings:
  • Learned masking agents consistently outperform full-information agents on risk-adjusted metrics.
  • Kalman-enhanced RL yields superior feature-selection of persistent signals and accelerates convergence to stable policies.
  • System-level metrics indicate lower propensity for flash events and destabilizing feedback loops.

Implications for AI Economics

• Strategy design: Trading systems should consider intentional information constraints (feature selection, input-rate limits, or deliberate delays) as a tool to improve performance and robustness, not merely as a cost or limitation.
• Information theory + markets: Modeling agent processing capacity with Shannon-style constraints provides a tractable, normative framework for rational inattention in algorithmic settings and links micro-level agent design to macro-level market outcomes.
• Market stability: Widespread adoption of selective-ignorance heuristics can dampen harmful feedback loops (overreaction, herding), suggesting regulators and exchanges might favor or even design infrastructure that reduces microstructure noise amplification (e.g., batching, minimum latencies, curated data feeds).
• Alpha generation: Selective ignorance creates orthogonal sources of alpha—filtering-induced returns that are not captured by standard factor models—implying new avenues for portfolio construction and risk management.
• Implementation guidance: Institutional trading stacks can integrate Kalman-type state estimation with RL-based masking modules as a modular layer before decision policies; this can reduce compute/latency costs and improve out-of-sample robustness.
• Risks and trade-offs: Deliberate blindness risks missing rare but important signals; adversaries might exploit predictable masking rules; coordination of ignorance across agents could create new externalities. These trade-offs motivate continued research into adaptive masking, adversarial robustness, and policy design.
• Research directions: Empirical validation on historical LOB data, theoretical analysis linking capacity bounds to equilibrium market properties, and exploration of market-design interventions that leverage selective-ignorance principles.