Shared Models, Selective Trading, and Order Flow

Battle Count: 50

Highly relevant for understanding how LLM-based trading agents behave under different information presentations. It highlights the risk of 'algorithmic homogeneity' where diverse installed models may collapse into one-sided flow due to shared reading biases, impacting market liquidity and price efficiency.

Requires setting up a multi-agent simulation environment with LLM integration (Ollama), managing stochastic seeds, and implementing complex accounting decompositions for order flow and participation. The analytical framework involves linear demand models and variance decompositions.

The paper provides detailed experimental setups, seed configurations, and code references for reproduction. However, it notes that model generation randomness was not fully controlled in the initial study (no explicit generation seeds), which is addressed in the prospective Study 2 design. The archive includes raw logs and analysis scripts.