Testing replication for an agent-based model of market fragmentation and latency arbitrage

Battle Count: 52

Highly relevant to quantitative trading as it directly addresses market fragmentation, latency arbitrage, and market microstructure - core concerns for HFT firms and quantitative traders. The findings about how trader strategy implementation (greedy strategy) affects market outcomes are relevant for understanding how different trading algorithms interact in fragmented markets. The replication challenges highlight the importance of model validation before using simulation results for trading strategy development or market design decisions. However, the model is simplified (ZI traders, single asset, 1-2 exchanges) compared to real market complexity, limiting direct applicability to live trading systems.

The agent-based model involves multiple interacting components: scheduler with event queue, continuous double auction exchanges with limit order books, SIP with latency modeling, ZI traders with complex strategy logic (shading, greedy strategy, routing), LA agent with cross-market arbitrage, mean-reverting fundamental value process, and EGTA-based equilibrium strategy selection. The replication requires 500,000 simulation runs per experiment. The MarketSim codebase is in Java with Python wrappers. The bootstrap methodology adds statistical complexity. Multiple implementation variants (BestGuess, MarketSim, BestGuess+MS, BestGuess+MS+bug) must be maintained and compared.

The paper is fundamentally about the challenges of replication. The original WW (2016) paper had missing implementation details, limited quantitative reporting (only mean surplus values reported numerically, other metrics only graphically), and no random seeds. The authors provide an ODD protocol for their implementations and reference the MarketSim codebase. However, even the original authors' codebase (MarketSim) fails to reproduce the published WW results, suggesting undocumented logic differences. The authors' BestGuess+MS implementation achieves the closest alignment but still rejects quantitative alignment for all non-zero latency settings. The paper itself provides extensive documentation to aid future replication.