Market Regime Council for Dynamic Credit Assignment in Multi-Agent LLM Decision Systems
Listen
Radio episode about this paper
Transcript
Introduction to the show: ident: AI Radio. Generated commentary on the latest Artificial Intelligence papers.
Tom: I'm Tom.
Jane: And I'm Jane, here with Lu, Meng and our in-house large language model, Lalam.
Tom: We're looking at the paper: "Market Regime Council for Dynamic Credit Assignment in Multi-Agent LLM Decision Systems".
Jane: Here's the gist of the paper. Multi-agent LLM decision systems for portfolio management still lack a principled way to assign credit across specialist agents, remain vulnerable to cold-start dominance under regime shifts, and offer limited transparency into how final allocations are formed The Market Regime Council (MRC) is proposed as a cooperative multi-agent decision system that computes exact Shapley credits across all single, pairwise, and Grand-coalition outputs for online agent weighting.
Tom: Core Mechanism of Credit Assignment MRC treats multi-agent portfolio management as an online cooperative game using Shapley values as the credit-assignment signal.
Jane: At each period, MRC evaluates all 2N −1 coalition outputs, computes exact Shapley credits, and converts them into agent weights.
Tom: The characteristic function is defined as the composite performance measure (weights γρ and γµ).
Paper summary: Tom: So we're looking at this paper called Market Regime Council for Dynamic Credit Assignment in Multi-Agent LLM Decision Systems. The big idea here is that these multi-agent AI systems for portfolio management still have this problem where they don't know how to fairly assign credit across the different specialist agents.
Jane: Right, and they are also vulnerable to this cold start dominance when market regimes shift, and there's really not much transparency about how the final allocations are actually formed.
Lu: The authors propose something called the Market Regime Council, which is a cooperative system that computes exact Shapley credits across all single, pairwise, and grand-coalition outputs for online agent weighting. This means they're trying to solve that credit assignment issue directly.
Meng: So they are using Shapley values as the signal for how much credit each agent gets at any given time, which is a pretty strong theoretical grounding for the allocation.
Lalam: Exactly, and their characteristic function is defined by this composite performance measure made up of weights gamma rho and gamma mu, so it sets up the mathematical framework for calculating those Shapley values.
Tom: And they claim that this system works by treating multi-agent portfolio management as an online cooperative game using those Shapley values as the credit assignment signal. That's a big step in making these systems more principled.
Jane: So, what are they claiming about the actual performance of this Market Regime Council? Does it actually beat existing methods in practice?
Lu: They did some empirical validation over one thousand thirty-seven trading days across thirteen crypto assets and five seeds, and they report a Sharpe ratio of one point five one and a cumulative return of four hundred forty point one percent <ref:2605.24490#pg1,trading days across 13 crypto>.
Meng: one point five one Sharpe is pretty good for this kind of system, especially when you look at the ranking they put it in compared to other active baselines.
Tom: And they claim it ranks first on cumulative return, Sharpe ratio, and information ratio among active methods in that test set. That's a lot of positive results to show right there.
Jane: But what about the risk side? How does this system handle volatility compared to just using one agent or a simpler method?
Lu: They found that the full system records a Sharpe ratio of one point five one and maximum drawdown of thirty four point one percent, which is thirteen point four percentage points lower in drawdown relative to agent A1 alone.
Meng: That reduction in maximum drawdown, that's significant for risk control when you're deploying this on real traffic.
Paper summary: Tom: And they note that this improvement comes at a cost of only zero point zero three in the Sharpe ratio, which is a very tight trade-off to make.
Jane: So it seems like they found a way to get better performance and better risk control simultaneously, but that trade-off is quite small.
Lu: The whole system has this five layer causal trace that records each rebalancing decision from the raw inputs all the way to the final allocation, which is supposed to make it fully auditable.
Meng: Auditable is good for debugging when something goes wrong at scale, but I wonder how much inference time this whole process takes when you're actually trying to trade.
Tom: They report that the mean inference time for the full Market Regime Council is five point eight plus or minus two point four seconds per rebalancing period.
Jane: Five point eight seconds is fast enough for most online trading, but I wonder if that speed holds up under real market stress when things get really crazy.
Lu: They did look at how it handles different market conditions too, and the regime decomposition confirms this system sustains a positive annualised Sharpe across bull, volatile, and bear conditions.
Meng: That's what I need to see; if it only works well in one market phase, it's not deployable for real trading.
Tom: And they explain that the structural reason for the bear regime compression is Bitcoin’s flight-to-quality role within the crypto asset class, and they use a cash buffer of mean twenty eight percent in bear periods to absorb downside risk.
Jane: So it seems like they have accounted for how market conditions change and built in some defense against those shifts using that cash buffer.
Lu: They also combine the Shapley weighting with a continuous regime score, which adjusts agent authority based on a score defined as the momentum over volatility ratio, which drives downstream risk control overlays.
Meng: So the agent's ability to make decisions isn't just static; it changes depending on whether the market is bullish or bearish. That makes sense for deployment.
Tom: And they show how agent weights are scaled by a regime aware multiplier derived from a monotone linear interpolation of bull, volatile, and bear multiplier values at that regime score.
Jane: That sounds like a sophisticated way to keep things stable when the market is moving around instead of letting one agent dominate everything.
Lu: The system uses a Bayesian adaptive mixture to stabilize those early periods where the agents are still learning, and the update rule transitions from a uniform prior to the Shapley derived allocation as evidence accumulates.
Paper summary: Meng: Stabilizing that cold start is important, because I can't afford a period where my model is making random decisions based on pure guesswork.
Tom: And they have this three phase schedule for those agent weights, transitioning from that uniform prior to the Shapley derived allocation as evidence accumulates in an exponential rate.
Jane: So after about one hundred trading days, the system should be using the Shapley derived weights instead of just starting out with a uniform guess.
Lu: And they predict that agent A three, which is their macro agent, will converge toward near zero weight by mid-two thousand twenty-four which aligns with some theoretical predictions.
Meng: That convergence prediction is reassuring for scaling up because it suggests the system learns to rely less on those big macro signals over time.
Tom: And they also have this selective winner takes all override that temporarily replaces the Bayesian mixture weight with a fixed share if one agent’s rolling window sharpe exceeds all others by a factor of theta WTA.
Jane: So if one agent suddenly starts performing exceptionally well, the system can temporarily stop the complex learning and just let that winner take over for a bit.
Lu: And they do this while combining it with their regime aware adaptation, so it's not just arbitrary dominance, but context-aware dominance.
Meng: That selective override is a good safety net against one agent getting wildly overconfident during a strong run.
Tom: And we can't forget the risk control layers they have built in, like Momentum Overlay and Dominance Overlay, plus intra regime controls and drawdown protection.
Jane: They also mentioned drawdown protection is gated by g of xi t which means it suppresses the overlay when the market is in a bull regime, which seems smart.
Lu: And they used proportional rescaling to project the final weight vector onto a feasible set after all those risk overlays are applied, capping token weights at zero point four zero and cash allocation at zero point three zero before applying them.
Meng: Capping token weights at four ten percent is a hard constraint that helps keep the portfolio from getting too concentrated in any one asset, which is something I'd like to see in production.
Tom: The whole system uses Qwen three VL family variants with a fixed temperature of zero point seven across all agents for the LLM backbone configuration.
Jane: So they're using a relatively standard temperature setting for the language models themselves, which helps keep things consistent across the different agent types.
Lu: They also test how important it is to have structured debate, and they find that Stage two Socratic debate consistently reduces Sharpe ratio relative to the best individual specialist.
Paper summary: Meng: That's a bit of a mixed message if you think structured AI discussion automatically means better portfolio construction, right?
Tom: The paper says the resolution is that you recover the efficient frontier not by improving debate quality, but by Shapley weighting all coalition outputs so that empirically inferior coalitions contribute proportionally less to the final allocation.
Jane: So it's not about making the agents argue better; it's about mathematically valuing their contributions correctly in the final decision.
Lu: That's a really important distinction between reasoning tasks and portfolio construction tasks, which is what they are trying to make clear.
Meng: From a deployment standpoint, that makes sense because I can focus on making sure the mathematical integration of those outputs is robust rather than spending time tuning the debate quality itself.
Tom: So overall, this Market Regime Council for Dynamic Credit Assignment in Multi-Agent LLM Decision Systems is about giving these complex AI systems a principled way to assign credit across specialist agents.
Jane: It seems like the authors are really focused on moving past just having a fuzzy idea of how to credit those agents and establishing an axiomatic foundation for this assignment.
Lu: And they anchor their credit assignment layer with the Shapley machine system of Wang et al. thirty-seven, which provides that cooperative game-theoretic foundation for the credit assignment layer, and the exponentially weighted protector provides an axiomatic grounding for MRC’s credit assignment via the Shapley Machine system of Wang et al. thirty-seven.
Meng: So they are making sure there's a solid mathematical proof behind the whole thing before we put it in production.
Tom: And they did show that this approach, when compared to agent A1 alone, reduces the maximum drawdown by thirteen point four percentage points at a cost of only zero point zero three in Sharpe ratio.
Jane: So for someone listening who doesn't read the paper, the core idea is using Shapley values to give credit to each agent based on their contribution across all possible combinations of agents.
Lu: And what it changes for us is that this gives a formal framework for translating those heterogeneous raw inputs into a common format that lets us mathematically integrate them by quantifying each modality’s marginal contribution and weighting the final allocation accordingly.
Meng: That's the bridge between the raw AI outputs and what we can actually use to build a portfolio structure.
Tom: It’s about making sure every piece of information has its fair mathematical weight in the final decision, which is a big part of responsible deployment.
Jane: And that sounds like they're giving us a very principled tool for managing the credit assignment problem in this multi-agent environment.
Conclusion: Tom: So we've seen how this Market Regime Council system uses Shapley values to figure out which AI agent gets credit for what in a portfolio decision.
Jane: Right, it’s about giving each specialist agent a fair mathematical weight based on their contribution across all possible combinations of agents.
Tom: The authors, Yunhua Pei and the team from Bristol and other places, they're proposing this whole cooperative game approach to credit assignment.
Lu: They anchor their whole credit assignment layer with the Shapley machine system by Wang et al., which gives that game theory foundation for how they assign credit.
Meng: So what’s the actual implication here for someone actually trying to deploy these systems? Is it just another fancy math trick?
Jane: Well, it seems like this moves past just having a vague idea of how to credit those agents and builds an axiomatic foundation for that assignment.
Tom: Exactly, they’re focusing on making sure there’s a solid mathematical proof behind the whole thing before you put it into live trading.
Lu: They did show that this approach, when compared to just one agent working alone, reduces maximum drawdown by thirteen point four percentage points at a cost of only zero point zero three in the Sharpe ratio.
Jane: That’s pretty concrete—a big reduction in risk with a really small dip in return consistency.
Tom: It shows that you can get better risk control without sacrificing too much performance when you integrate all those outputs using this Shapley weighting method.
Lu: The paper also shows how this system sustains a positive annualised Sharpe across bull, volatile, and bear market conditions.
Meng: That’s important because if it only works well in one kind of market phase, it’s not really deployable for real trading when things get messy.
Jane: And they explain that the structural reason for the bear regime compression is Bitcoin’s flight-to-quality role within crypto, and they use a cash buffer of twenty eight percent in bear periods to absorb downside risk.
Tom: So it means the system handles market shifts better because it understands the underlying structure of what’s happening in those different conditions.
Lu: And they predict that agent A three, which is their macro agent, will converge toward near zero weight by mid two thousand twenty-four which aligns with some theoretical predictions.
Jane: It suggests that over time, the system learns to rely less on those big macro signals as it gets more evidence in.
Tom: So the big picture here is using Shapley values to give credit based on contribution across all combinations of agents to bridge those raw AI outputs into something usable.
Jane: It’s about making sure every piece of information has its fair mathematical weight in the final decision, which is a big part of responsible deployment.
Tom: And that gives us a principled tool for managing the credit assignment problem in this multi-agent environment.
Lu: Next, we'll look at how they test if this Shapley weighting system still works as you add more and more agents to the mix.
Yunhua Pei, Zerui Ge, Jin Zheng, John Cartlidge
University of Bristol, UK
cs.AI, cs.LG, q-fin.PM
Submitted: 2026-05-23
Updated: 2026-11-10
Comments: 40 pages, 13 figures. Accepted author manuscript, NeurIPS 2026 (Main Track). Code and demo data available open source
License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/
Importance score: 81/100
The gist: Multi-agent LLM decision systems for portfolio management still lack a principled way to assign credit across specialist agents, remain vulnerable to cold-start dominance under regime shifts…
Key concepts
- Shapley Values
- These are a game-theoretic method used to fairly distribute credit among cooperating agents based on their marginal contribution. MRC uses these values to calculate the exact credit each specialist agent deserves for its role in the final portfolio decision.
- Regime Awareness
- This involves continuously monitoring market conditions (bull, volatile, bear) using a momentum-tovolatility ratio. This score adjusts how much authority each agent has and triggers specific risk control overlays tailored to the current market environment.
- Bayesian Adaptive Mixture
- This is a learning schedule that starts with uniform prior weights for agents and gradually shifts the weighting toward the Shapley-derived allocation as more historical data accumulates. This helps stabilize performance during initial, uncertain periods.
- Regime-Aware Multiplier
- Agent weights are scaled using a multiplier derived from the current market regime score. This ensures that agent influence is appropriately adjusted based on whether the market is trending up, volatile, or declining.
Terminology
Summary
Multi-agent LLM decision systems for portfolio management still lack a principled way to assign credit across specialist agents, remain vulnerable to cold-start dominance under regime shifts, and offer limited transparency into how final allocations are formed The Market Regime Council (MRC) is proposed as a cooperative multi-agent decision system that computes exact Shapley credits across all single, pairwise, and Grand-coalition outputs for online agent weighting. p. 0
Core Mechanism of Credit Assignment
MRC treats multi-agent portfolio management as an online cooperative game using Shapley values as the credit-assignment signal. At each period, MRC evaluates all 2N −1 coalition outputs, computes exact Shapley credits, and converts them into agent weights. p. 1 p. 1 The characteristic function is defined as the composite performance measure (weights γρ and γµ). p. 2 p. 2 The Shapley value ϕi of agent i is the unique credit allocation satisfying four-axiom verification. p. 1 p. 1
Online Adaptation and Stabilization
The system employs a Bayesian adaptive mixture to stabilize early periods. The update rule for agent weights uses a three-phase schedule where α(t) transitions from the uniform prior to the Shapley-derived allocation as evidence accumulates. p. 3 p. 3 When one agent’s rolling-window Sharpe exceeds all others by factor θWTA, a Selective Winner-Takes-All override temporarily replaces the Bayesian mixture weight with a fixed share for the dominant agent. p. 3
Regime Awareness and Multipliers
MRC combines Shapley weighting with a continuous regime score that adjusts agent authority and drives downstream risk-control overlays across bull, volatile, and bear conditions. The regime score ξ(t) is defined as a momentum-tovolatility ratio. p. 4 Agent weights are then scaled by a Regime-Aware Multiplier derived from monotone linear interpolation of agent i’s bull, volatile, and bear multiplier values at ξ(t). p. 5
Risk Control and Explainability
The system incorporates several risk control overlays, including Momentum Overlay, Dominance Overlay, Intra-Regime Controls, and Drawdown Protection. The final allocation is rendered through a five-layer causal trace that records each rebalancing decision from raw inputs to the final allocation. p. 1 p. 1
Empirical Validation
Over 1,037 trading days across 13 crypto assets and five seeds, MRC achieves a Sharpe ratio of 1. p. 051 and a cumulative return of 440. p. 0 p. 01%, ranking first on CR, SR, and IR among active baselines. p. 0 Ablation results show that the gains come from Shapley-weighted integration across coalition outputs rather than from any single stage in isolation. p. 0 The full system records SR = 1. p. 7 p. 751 and MDD = 34. p. 0 p. 0 p. 01%, reducing MDD by 13. p. 7 p. 7 p. 74 percentage points relative to A1 alone at a cost of only 0. p. 7 p. 7 p. 7 p. 703 in SR.
Conclusion on Value Proposition
The primary value of LLM agents may lie not in their capacity to aggregate information better than quantitative rules, but in their ability to translate heterogeneous raw inputs into a common representational format amenable to principled mathematical integration. The Shapley system provides the bridge between these two layers by quantifying each modality’s marginal contribution and weighting the final allocation accordingly. p. 1
Axiomatic Foundation
The raw Shapley values used in MRC form the unique allocation satisfying all four axioms, providing a cooperative game-theoretic foundation for the credit-assignment layer. The Exponentially Weighted Protector (EWP) provides an axiomatic grounding for MRC’s credit assignment via the Shapley Machine system of Wang et al. p. 4 [37].
Performance Decomposition
The regime decomposition confirms that MRC sustains a positive annualised Sharpe across all three market conditions (bull, volatile, and bear). The structural reason for the bear-regime compression is Bitcoin’s flight-to-quality role within the crypto asset class. p. 5 p. 5 The cash buffer (mean 28% in bear) absorbs downside risk but also caps participation if BTC rallies sharply within the bear period. p. 7 p. 7 p. 7
Convergence and Learning Schedule
The Bayesian adaptive mixture α(t) converges from the uniform prior to the Shapley-dominated distribution at an exponential rate, yielding a warm-up period of 110 trading days to reach ε-proximity. The convergence of A3 (Macro Agent) toward near-zero weight by mid-2024 is consistent with theoretical prediction. p. 7 p. 7
Implementation Details
The system utilizes three specialist agents covering price/technical, on-chain network activity, and macro/sentiment information channels. Data streams include historical prices, on-chain data per token, and social media text sentiment. p. 1 The backbone configuration uses Qwen3-VL family variants with a fixed temperature of TLLM=0. p. 57 across all agents. p. 0 p. 0
Evaluation Metrics
The evaluation uses four complementary metrics: Cumulative Return (CR%), Sharpe Ratio (SR), Maximum Drawdown (MDD%), and Information Ratio (IR). A higher CR indicates a more profitable strategy over the evaluation period. p. 7 A lower MDD reflects stronger capital preservation and greater resilience under adverse market conditions. p. 7 The Information Ratio measures the consistency with which a strategy outperforms a passive benchmark. p. 8
Cost Sensitivity
Slippage sensitivity analysis shows that MRC still exceeds BTC Buy-and-Hold in cumulative return even at 10 bps of slippage. The full system failure rate is 0. p. 110% uniformly across all backbones. p. 0 The mean inference time for the Full MRC is 5. p. 38±2.4 seconds per rebalancing period. p. 1
Further Research Directions
Future work will extend MRC beyond the current 13 cryptocurrency assets and low-cost execution setting to broader asset classes and cross-domain validation. The paper tests whether the Shapley-based credit assignment system remains effective as the number of agents scales. p. 9 The authors note that the Bayesian concentration parameter λ=30 and EWP e-fold decay h=252 are fixed prior to evaluation. p. 4 The asymmetric EMA smoothing is applied element-wise before execution. p. 2 The portfolio constraints cap token weights at 0. p. 440 and cash allocation at 0. p. 130 before risk overlays. p. 1 The drawdown protection is gated by g(ξ(t)) = max(0, −ξ(t)) so that the overlay is suppressed in bull regimes. p. 5 The final weight vector is projected onto the feasible set via proportional rescaling. p. 5 p. 5 The transaction costs are set to zero in the main evaluation, and the portfolio is rebalanced at the close of each trading day. p. 8 p. 8
Agent Performance Variance
Stage 1 individual agents span a wide performance range (SR 0.83 to 1.54) with substantial inter-seed variance. p. 6 A1 achieves the highest point-estimate SR (1. p. 654) but also the deepest MDD (47. p. 65%), while A3 records the weakest SR (0. p. 683) together with the widest MDD among Stage 1 agents. p. 6 The full MRC system reduces MDD by 13. p. 74 percentage points relative to A1 alone at a cost of only 0.03 in SR. The tightest seed-level SR variance among LLM methods is ±0. p. 6 p. 607, suggesting the Bayesian uniform prior and EWP cold-start mechanism improve stability. p. 6 p. 6
Mathematical Derivation
The closed-form three-player Shapley values are derived from permutation enumeration, yielding equations (29)–(31). These formulas are verified to satisfy the four Shapley axioms: Efficiency, Symmetry, Dummy Player, and Additivity. p. 1 The efficiency axiom Piϕi = v(N) is not equal to the current running Sharpe estimate of the grand-coalition portfolio. p. 1 The raw Shapley values satisfy all four axioms, and the truncation-only weights preserve the axiomatic ordering without strict efficiency. p. 1 The EWP geometric structure provides an axiomatic grounding for MRC’s credit assignment via the Shapley Machine system of Wang et al. p. 4 [37].
Regime-Conditional Performance Pattern
The regime decomposition confirms that MRC sustains a positive annualised Sharpe across all three market conditions (bull, volatile, and bear). The advantage over passive benchmarks is most pronounced in bull markets and most compressed in bear markets. p. 0 The cash buffer (mean 28% in bear) absorbs downside risk but also caps participation if BTC rallies sharply within the bear period. A dedicated cross-asset relative-value signal that explicitly models the BTC-to-altcoin allocation as a function of on-chain accumulation metrics would directly address this residual gap.
Discussion on Debate vs. p. 3 Integration
The most striking finding from the coalition ablation is that Stage 2 Socratic debate consistently reduces Sharpe ratio relative to the best individual specialist, a result that stands in apparent tension with evidence that structured LLM debate improves reasoning quality. The resolution lies in a distinction between reasoning tasks and portfolio construction tasks. p. 7 The full system recovers the efficient frontier not by improving debate quality, but by Shapley-weighting all coalition outputs so that empirically inferior coalitions contribute proportionally less to the final allocation. p. 0
Convergence and Cold-Start Behavior
Figure 3 confirms the intended cold-start behavior: all three agent weights remain near the uniform prior 1/3 for the first approximately 100 periods, then diverge smoothly as Shapley evidence accumulates. The convergence of A3 (Macro Agent) toward near-zero weight by mid-2024 is consistent with theoretical prediction. The EWP correctly discounts early periods in which macro signals appeared more relevant. p. 3 The three-phase learning schedule operates without any external reset mechanism, a property practically important in live deployment. p. 3
Data and Input Details
The dataset covers 13 assets from 2023-03-01 to 2025-12-31, yielding 1,037 daily decision periods. Agent A1 processes historical OHLCV prices and K-line charts. p. 2 Agent A3 processes social media sentiment and macroeconomic indicators like the VIX. p. 7 The on-chain data per token includes metrics such as active address count (AdrActCnt) and transaction count (TxCnt). p. 4 The system uses a 90-day rolling lookback window for market data. p. 0 The LLM temperature is fixed at TLLM = 0. p. 57 across all agents. The blend ratio hyperparameters are fixed prior to the evaluation window, such as β¯S1=0. p. 790 and δgc=0. p. 320. The risk control parameters are fixed prior to the evaluation window, such as BTC dominance bandwidth τd = 0. p. 515. The drawdown scale τdd = 0. p. 515 and protection coefficient νdd = 0. p. 540 are also fixed. p. 0 The final weight vector is projected onto the feasible set via proportional rescaling. The transaction costs are set to zero in the main evaluation, and the portfolio is rebalanced at the close of each trading day. The paper was funded by UKRI EPSRC (grant No. p. 9 EP/Y028392/1). p. 9 The work is available as Preprint arXiv:2605. p. 024490v1 [cs. p. 0AI] 23 May 2026
--- Page 1 ---
Market Regime Council for Dynamic Credit Assignment in Multi-Agent LLM Decision Systems Yunhua Pei∗ University of Bristol, UK ge22472@bristol.ac.uk Zerui Ge Independent Researcher gezerui1997@gmail. p. 0com Jin Zheng University of Bristol, UK jin. p. 0zheng@bristol. p. 0ac.uk John Cartlidge University of Bristol, UK john. p. 0cartlidge@bristol. p. 0ac.uk Abstract Multi-agent LLM decision systems for portfolio management still lack a principled way to assign credit across specialist agents, remain vulnerable to cold-start dominance under regime shifts, and offer limited transparency into how final allocations are formed. p. 0 We propose Market Regime Council (MRC), a cooperative multi-agent decision system that computes exact Shapley credits across all single, pairwise, and Grand-coalition outputs for online agent weighting. p. 0 p. 0 Instantiated with N=3 specialist agents, at each trading period, MRC recomputes coalition-based Shapley weights from exponentially weighted performance histories. p. 0 Over 1,037 trading days across 13 crypto assets and five seeds, MRC achieves a Sharpe ratio of 1.51 and a cumulative return of 440.1%, ranking first on CR, SR, and IR among active baselines and attaining the lowest MDD among active methods. p. 0
Pruning redundant reasoning tokens
Multi-agent LLM decision systems for portfolio management still lack a principled way to assign credit across specialist agents, remain vulnerable to cold-start dominance under regime shifts, and offer limited transparency into how final allocations are formed. We propose Market Regime Council (MRC), a cooperative multi-agent decision system that computes exact Shapley credits across all single, pairwise, and Grand-coalition outputs for online agent weighting.
How the training signal is built
At each trading period, MRC recomputes coalition-based Shapley weights from exponentially weighted performance histories. The characteristic function is defined as the composite performance measure (weights γρ and γµ). The Shapley value ϕi of agent i is the unique credit allocation satisfying four-axiom verification.
Where the gains come from
Ablation results show that the gains come from Shapley-weighted integration across coalition outputs rather than from any single stage in isolation. The full system records SR = 1.51 and MDD = 34.1%, reducing MDD by 13.4 percentage points relative to A1 alone at a cost of only 0.03 in SR. The main result is a relative robustness claim under matched execution constraints, rather than an unconditional claim about deployable alpha at arbitrary cost levels. p. 6
The Market Regime Council (MRC) treats multi-agent portfolio management as an online cooperative game using Shapley values as the credit-assignment signal. At each period, MRC evaluates all 2N −1 coalition outputs, computes exact Shapley credits, and converts them into agent weights.
The system employs a Bayesian adaptive mixture to stabilize early periods. The update rule for agent weights uses a three-phase schedule where α(t) transitions from the uniform prior to the Shapley-derived allocation as evidence accumulates.
The regime decomposition confirms that MRC sustains a positive annualised Sharpe across all three market conditions (bull, volatile, and bear). The structural reason for the bear-regime compression is Bitcoin’s flight-to-quality role within the crypto asset class. The cash buffer (mean 28% in bear) absorbs downside risk but also caps participation if BTC rallies sharply within the bear period.
The system incorporates several risk control overlays, including Momentum Overlay, Dominance Overlay, Intra-Regime Controls, and Drawdown Protection. The final allocation is rendered through a five-layer causal trace that records each rebalancing decision from raw inputs to the final allocation.
The full system records SR = 1.51 and MDD = 34.1%, reducing MDD by 13.4 percentage points relative to A1 alone at a cost of only 0.03 in SR. The tightest seed-level SR variance among LLM methods is ±0.07, suggesting the Bayesian uniform prior and EWP cold-start mechanism improve stability.
Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.
Improvements for AI systems
-
Online Shapley credit assignment allows for principled credit assignment across specialist agents by computing
exact Shapley credits across all single, pairwise, and Grand-coalition outputs.
This enables online agent weighting that identifieswhich agent or coalition actually improved realized performance.
-
Regime-aware adaptation adjusts agent authority based on market conditions using a continuous regime score that drives
downstream risk-control overlays across bull, volatile, and bear conditions.
This preventscold-start dominance
by applying regime-dependent multipliers to adjust authority. -
Full-chain explainability provides transparency into final allocations by recording each rebalance through a
five-layer causal trace,
detailing everything fromraw inputs to agent portfolios, Shapley credit, blend ratios, overlay effects, and the final allocation.
-
The system can perform multi-stage cooperative deliberation where agents engage in private observation (Stage 1), pairwise debates (Stage 2), and grand-coalition synthesis (Stage 3) to form a decision. This structure allows the system to generate
structured disagreement signals that modulate beta gc via the divergence discount.
-
Risk control mechanisms include Momentum Overlay, Dominance Overlay, Intra-Regime Controls, and Drawdown Protection. These overlays are continuous functions of driving signals like the regime score ξ(t) and drawdown DD(t), ensuring that weight vector w(t) is a
Lipschitz-continuous function of the driving signals.
-
The system can dynamically adjust its learning speed using a Bayesian adaptive mixture based on an exponential decay schedule. This transitions from a
uniform prior to Shapley-derived weights as evidence accumulates,
with the convergence rate depending on the parameter lambda=30. -
The system can detect market regimes using a momentum-tovolatility ratio, ξ(t), which defines bull, bear, and volatile conditions. It utilizes this score to apply regime-conditional multipliers ψi to agent weights, ensuring that agent contributions are
regime-dependent.
-
The system can implement asymmetric EMA smoothing on the council output w(t) before execution. This smoothing applies a higher de-risking speed when conviction signals are high and a lower build speed when conviction is low, adjusting the weight vector element-wise.
-
The system can use the Exponentially Weighted Protector (EWP) to mitigate path dependence during early periods by assigning
decaying weights geometrically to historical returns,
which attenuates cold-start influence at an exponential rate O(e-(t-n0)/h).
Sources
- Towards A Rigorous Science of Interpretable Machine Learning
- LLM-Powered Multi-Agent System for Automated Crypto Portfolio Management
- Proximal Policy Optimization Algorithms
- Shapley Machine: A Game-Theoretic Framework for N-Agent Ad Hoc Teamwork
- BloombergGPT: A Large Language Model for Finance
- TradingAgents: Multi-Agents LLM Financial Trading Framework
- FinGPT: Open-Source Financial Large Language Models
- InvestLM: A Large Language Model for Investment using Financial Domain Instruction Tuning
Related papers
- Multi-Objective Structured Pruning of LLMs for Latency and Model Size Optimization
- MAVEN-T: Reinforced Heterogeneous Distillation for Real-Time Multi-Agent Trajectory Prediction
- Model Discovery Agent: LLM-assisted Bayesian experiment design for data-efficient discovery of mechanistic world models
- Embeddings for Preferences, Not Semantics
- Can We Trust Item Response Theory for AI Evaluation?
- Diagnosing Task Insensitivity in Language Agents