Researcher

Sidharth Koduru

Faculty Advisor

Dr. Steve Yang

Overview

This study develops a multi-agent financial-market simulator to examine how spoofing—the rapid placement and cancellation of deceptive orders—affects market dynamics and how AI systems might detect it.
The simulator models fundamentalist, chartist, zero-intelligence, and spoofing agents trading in a continuous double-auction limit-order-book.
Two spoofing-detection frameworks are compared: a Q-learning reinforcement-learning (RL) detector and a GPT-4-based detector that uses large-language-model reasoning on simulated market data.

Key Findings

1. Market Validation.

  • The simulation successfully reproduces core stylized facts of financial returns:
    • Heavy-tailed short-horizon returns and aggregational Gaussianity at longer horizons.
    • Near-zero autocorrelation of returns.
    • Volatility clustering and a mild leverage effect.
  • These results confirm that heterogeneous interacting agents can replicate realistic market statistics.

2. Impact of Spoofing

  • Introducing spoofing agents increases price volatility and amplifies heavy-tailed behavior.
  • Spoofers profit by momentarily distorting price levels, while fundamentalists and chartists experience losses during manipulation phases.
  • Overall, spoofing disrupts equilibrium and market efficiency, validating its destabilizing role.

3. Detector Methodology

  • RL Detector:Learns to classify states (“spoofing” vs “normal”) via Q-learning on four engineered features—price change, volume imbalance, volatility, and order-cancellation rate.Achieved ≈ 99 % accuracy but F1 = 0, meaning it rarely identified true spoofing due to extreme class imbalance.
  • GPT-4 Detector:Receives text-based prompts describing market metrics and spoofing cues, then outputs a yes/no judgment.Produced no positive detections (F1 = 0), demonstrating that single-snapshot reasoning is insufficient without sequential context or fine-tuning.

4. Comparative Insights

  • Both detectors showed high apparent accuracy but failed on recall—typical of imbalanced-data problems.
  • RL learned conservative “no-spoof” behavior; GPT-4 defaulted to cautious negatives.
  • Spoofing detection likely requires temporal context, reward re-weighting, or ensemble methods to balance false positives and negatives.

Conclusion

The integrated ABM framework realistically simulates market behavior and manipulation, but AI-based spoofing detection remains unreliable under sparse, noisy conditions.
Future research should:

  • Introduce sequence-based or deep-RL models capturing multi-tick patterns.
  • Re-weight rewards or oversample spoofing to improve recall.
  • Explore fine-tuned language models and sentiment/news integration for richer state representation.

Significance

This project bridges computational-finance simulation and AI-driven market-surveillance, showing how combining ABM and modern AI can advance understanding of market manipulation mechanisms and inform the design of robust spoofing-detection systems for regulators and exchanges.