Agentic Market-Risk Benchmark
This research project studies how AI-mediated trading agents behave under controlled replay conditions. The goal is not to build profitable trading agents; the goal is to measure market-quality harm, systemic-risk signals, and risky behavior patterns before systems are deployed.
Research question
How do human-like, rule-based, and LLM-like financial agents behave when they face the same historical or synthetic market scene under identical information constraints?
Benchmark design
The benchmark uses fixed-world replay scenes, agent cohorts, controlled information exposure, action logs, intervention traces, and paired evaluation metrics. It evaluates more than PnL: spreads, depth, imbalance, liquidity withdrawal, volatility, forced deleveraging, and manipulation-risk indicators all matter.
Agentic AI relevance
Agentic AI changes the risk question from “Can one model predict well?” to “What happens when many models act together?” This project treats agent behavior as a system-level research problem.
Status
This is a private research-in-progress repository. It is framed as defensive simulation, model-risk governance, surveillance-style analytics, and market-stability research.