acceptodds
Under review as a conference paper at ICLR 2027

PosTR: Position Transition Reasoning for Multi-agent Financial Decisions

Abstract

Multi-agent financial trading systems typically improve decision quality by assigning specialized agents to analyze heterogeneous information sources, such as market signals, historical trends, and fundamental factors. However, their final decisions are often compressed into coarse-grained actions, such as buy, hold, and sell, resulting in a mismatch between rich analytical reasoning and simplified trading execution. When an aggregator receives only such action labels, distinct target-position transitions and their implied market assessments are treated as the same action, a loss we term Decision Semantic Collapse. % To address this limitation, we propose PosTR, a Position Transition Reasoning framework that reformulates multi-agent trading from action prediction into position-transition reasoning. % Instead of directly aggregating discrete trading actions, PosTR introduces an ordered position-state space, where each agent provides a target position, its transition from the previous target position, and a distribution over candidate target positions. This formulation preserves the target-position transition and its implied market assessment while representing uncertainty over the current target position. % , while enabling heterogeneous agent opinions to be fused within a unified state space rather than through action-level voting % Furthermore, PosTR develops a transition-aware credibility mechanism that dynamically evaluates agent reliability. The learned credibility scores reweight trader distributions during consensus formation, while settled branch performance controls the strength of the credibility-based correction. % Experiments with role-specialized agents on eight equities spanning eight sectors show stronger net and risk-adjusted performance than endpoint-only, static-weight, and history-aware baselines. These results highlight that effective multi-agent trading requires reasoning over position-state transitions rather than merely aggregating discrete trading actions.

Primary area:
foundation or frontier models, including LLMs
Keywords:
Multi-agent systems, Large language model agents, Financial decision-making, Position-state space, Position-transition reasoning, Credibility-aware aggregation

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_UNCmOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes PosTR, a Position Transition Reasoning framework that reformulates multi-agent trading from action prediction into reasoning over position transitions. Each trader’s decision is represented by its consecutive target positions together with a distribution over candidate positions. These representations are aggregated through online, transition-aware credibility weighting combined with an adaptive correction toward an equal-weight consensus. Experiments on eight U.S. equities show higher cumulative returns than the evaluated baselines, and matched-input ablations support the benefits of the distributional representation and context-dependent aggregation.

Critical strengths
  1. The problem is well motivated and clearly presented. The observation that compressing distinct position transitions into a single action label causes “decision semantic collapse” is a crisp and convincing framing.
  2. The paper is well written and organized, with intuitive figures that make the overall framework and each component easy to follow.
Critical weaknesses
  1. The core contribution is not isolated from simpler alternatives. The transition reparameterization appears to be invertible and therefore encodes no information beyond the two endpoint positions. The paper does not compares against a simple aggregator that operates directly on each trader’s previous and current target positions, nor against standard online expert-weighting methods (Weighted Majority, Hedge) or the closest position-aware system, FinPos (Liu and Dang, 2025). Without these baselines, it is not possible to tell whether the reported gains come from the proposed transition reasoning or simply from using position information at all. This is consequential because the paper’s central claim—that position-transition reasoning, rather than position-awareness in general, is what matters—rests on exactly this distinction.
  2. The empirical evidence is too narrow to support the claims. All results come from eight equities over a single year (September 2, 2025 – August 31, 2026). Trading strategies are notoriously regime-dependent, and a one-year window cannot show whether the credibility weighting and consensus correction generalize across bull, bear, and high-volatility conditions or around major market events. Given that the paper’s main evidence is cumulative return, this limitation directly undermines confidence in the headline result.
Questions for authors
  1. What does “reasoning” specifically refer to in “Position Transition Reasoning”? The method does not model the reasoning processes and appears closer to reasoning over structured position transitions than to capturing or verifying complete reasoning chains.
  2. How does PosTR perform across different market regimes or on a longer history? Any out-of-sample period beyond the one reported would help.
  3. Comparing against FinPos and standard online expert-weighting methods as baselines.
Reviewer_BYfoOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes PosTR for aggregating financial agents’ trading decisions. Each agent provides a target position and a distribution over Long, Flat, and Short. The previous and current targets define a transition, which is combined with market and peer information to estimate credibility using recursive Bayesian linear regression. The final position blends credibility-weighted and equal-weight consensus. Experiments on eight U.S. equities report improved net returns and risk-adjusted performance.

Critical strengths
  1. Full state improves on Semantics by about 8.7 percentage points in mean cumulative return, with gains on all eight equities (Table S28). This supports the use of transition features in addition to current opinion distributions.
  2. The experimental protocol is carefully specified, with separate development and test periods, frozen-input comparisons, repeated sampling, and paired block-bootstrap intervals. The 16.7-point gain over Equal consensus has a 95% interval of [8.9, 24.8].
Critical weaknesses

The benefit of the proposed representation over a standard history-aware aggregator is unclear. Raw state in Table S16 represents the previous/current target pair through two derived transition features. For the same current distribution, these features cannot distinguish Short-to-Long from Long-to-Long. The reported 4.8-point gain therefore does not establish whether the proposed representation improves on a baseline that retains both previous and current targets.

Questions for authors

Does the gain persist when a baseline receives complete previous/current target pairs and the same probabilities and market/peer features, using the same learner and development-period tuning budget?

Reviewer_6X7JOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper explores how to aggregate recommendations from multiple LLM trading agents without losing information through coarse buy/hold/sell labels. PosTR retains each trader's previous and current target positions, along with a distribution over long, flat, and short. Based on these inputs and market context, it learns reliability weights, and then adaptively blends the weighted consensus with equal pooling to produce a continuous trading exposure. The method is evaluated on eight U.S. equities over one year, and the authors report improvements over both trading baselines and simpler aggregation rules, supported by matched ablations, uncertainty estimates, and robustness checks in the supplement.

Critical strengths

1/ The aggregation problem is well motivated. Closing a long position and opening a short position convey different views, even though both may be described as "sell." Preserving target transitions and uncertainty therefore provides the aggregator with useful information for resolving disagreement. The distinction between individual recommendations and the system's executed exposure is also clearly formulated (Sections 3–4).

2/ The matched evaluation provides meaningful evidence for learned aggregation. Replaying the same frozen trader outputs makes comparisons with equal pooling and other aggregation rules informative. The reported improvement over equal consensus is 16.7 percentage points in mean cumulative return, with a joint date-block bootstrap 95% interval of [8.9, 24.8]. This evidence is further strengthened by the separate development period, transaction-cost sensitivity, and repeated trader generation.

Critical weaknesses

While the matched experimental results are promising, I would need the execution discrepancy to be resolved, and the distinct contribution of the representation to be demonstrated more clearly.

1/ I cannot reconcile the execution example with the stated method. Figure 5(c) attributes a 3.68-percentage-point advantage to Equal consensus being Short and PosTR being Long on a day with a +1.84% return. However, Eqs. (14)–(15) define continuous exposures as convex mixtures of the same three trader distributions. The exposure difference between equal pooling and such a reweighting is at most 4/3, which permits only about 2.45 points of gross advantage on that move. This suggests that a different execution rule may be involved. Although this may be confined to the illustrative replay, the discrepancy needs to be resolved before the example can support the claimed mechanism.

2/ The representation comparison does not clearly establish the main methodological contribution. As the paper acknowledges, (m,v) is an invertible transformation of the previous and current targets. The "Raw state" baseline changes two handcrafted features, rather than directly supplying those endpoints to the same learner (Table S16). In particular, its previous target enters only through its absolute magnitude, and thus loses its sign. The reported improvement therefore supports the selected features, but it remains unclear how much the proposed representation adds over a straightforward contextual learner that retains both signed targets.

3/ The comparison with established multi-agent trading systems is limited. TradingAgents is the only external multi-agent trading baseline, while FinMem is a single-agent system; other relevant frameworks discussed in the paper, such as FinCon, are not evaluated (Table 1; Section 2). Moreover, the LLM baselines execute categorical positions, whereas PosTR uses continuous position sizing. The matched ablations provide useful evidence for the aggregation mechanism, but the external comparisons leave unclear whether its advantage would persist against other established multi-agent systems with comparable position-sizing capabilities.

Questions for authors

1/ How does PosTR compare with the same credibility learner given the previous and current signed targets directly, or an encoding of the nine transitions, while keeping the remaining context, distributions, and execution rule fixed?

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.