Market Regime Council for Dynamic Credit Assignment in Multi-Agent LLM Decision Systems

Battle Count: 69

Directly addresses core quantitative trading challenges: dynamic credit assignment across heterogeneous signal sources, regime-aware portfolio rebalancing, risk management overlays, and online adaptation to non-stationary markets. The Shapley-based framework provides principled attribution of portfolio performance to specific data modalities. Results demonstrate state-of-the-art performance among active LLM and DRL baselines on crypto portfolios. The regime-conditional multiplier system and risk-control cascade are directly applicable to live trading systems. However, the crypto-specific evaluation and zero-cost assumption limit immediate applicability to traditional equities/futures.

High complexity due to: (1) multi-stage LLM deliberation requiring 7+ LLM calls per period (3 individual + 3 pairwise debates + 1 grand coalition readout); (2) exact Shapley computation over all 2^N-1 coalitions; (3) Bayesian adaptive mixture with exponentially weighted protector; (4) regime detection and multi-stage risk-control cascade with 7 overlays; (5) five-layer causal trace for explainability; (6) asymmetric EMA smoothing; (7) portfolio constraint projection. Requires vision-language model API access, multiple data pipelines (OHLCV, on-chain, macro, sentiment), and careful hyperparameter management. Inference time ranges from ~25s (8b-thinking) to ~458s (30b-a3b-thinking) per rebalancing period.

Code and demo data included in supplementary material. Complete pseudocode for both core algorithms (Algorithm 1: Shapley Value Ledger Update; Algorithm 2: Per-Period Inference) provided. All hyperparameters explicitly listed. Five-seed evaluation with statistical significance tests. However, proprietary on-chain dataset and specific Twitter/X sentiment dataset may limit full reproduction. Backbone configuration via OpenRouter API.