Representation Signatures and Risk-Feedback Alignment in LLM Trading Agents

Battle Count: 58

Highly relevant to quantitative trading as it directly addresses LLM-based trading agents, execution realism, risk management, portfolio construction, and failure detection. The paper provides a framework for auditing trading-agent decisions, detecting pre-failure representation signatures, evaluating risk-feedback alignment, and identifying correlation-blindness in high-dimensional portfolio reasoning. However, it explicitly states results are benchmark diagnostics rather than investment evidence, and does not claim profitability. The execution modeling (commissions, slippage, latency, partial fills) and risk-gate mechanisms are directly applicable to production trading systems.

TradeArena involves a complex modular architecture with data providers, agent lifecycle (observe-plan-propose-revise-reflect), risk layer, execution simulator, memory/tools, trajectory logging, and evaluation outputs. The representation analysis requires multiple embedding views (hash, LSA, BGE-M3, white-box hidden states), rolling failure anchors, noise injection, CoT-free ablations, and manifold diagnostics. The experimental matrix spans 25+ experiment families with 500+ Poe-mediated calls. However, the modular plugin-like interfaces and deterministic synthetic data make the core benchmark reproducible. The frontier model integration via Poe adds external dependency complexity.

TradeArena records full replayable trajectories with deterministic synthetic data, cached LLM responses (JSONL), commit-hash-command provenance, and a 98-run trajectory collection with 1.0 coverage on risk lifecycle, trajectory reproducibility, and agent trace. Code and data artifacts are available through the TradeArena repository. However, frontier model results are routed through Poe (cached), and commercial model aliases are version-sensitive. The BGE-M3 and Qwen2.5-0.5B probes are local and reproducible. The 50-step human annotation file for hallucination calibration is exported but not completed in this submission.