acceptodds
Under review as a conference paper at ICLR 2027

Beyond Market Consensus: Modeling Trader Heterogeneity through Behavioral Residuals

Abstract

Realistic market simulation requires agents to preserve persistent behavioral differences, yet large language model (LLM) agents often converge toward average behavior when exposed to similar market conditions. We argue that trader heterogeneity can be characterized by how individuals systematically deviate from shared market consensus. Based on this insight, we propose Consensus-Guided Residual Behavior Modeling (CRBM), which decomposes trading behavior into a consensus component shared across traders and a behavioral residual capturing trader-specific deviations. CRBM derives consensus signals from local market contexts, estimates persistent residual patterns from individual transaction histories, and integrates them through a residual-conditioned expert mechanism to generate structured trading orders. Using real transaction records from Polymarket, we evaluate CRBM at three levels: individual behavioral fidelity, cross-trader heterogeneity, and emergent market dynamics. CRBM consistently improves transaction-level prediction and joint action–outcome matching over competitive baselines, while ablations demonstrate the complementary roles of consensus modeling and behavioral residuals. In Monte Carlo market replays, CRBM preserves wallet-specific behavioral differences and produces aggregate price dynamics consistent with observed market evolution. Zero-shot evaluation on Manifold further provides preliminary evidence of cross-platform generalization. These results suggest that modeling individual behavior as structured deviations from shared consensus provides an effective approach to preserving agent heterogeneity in LLM-based market simulation.

Primary area:
foundation or frontier models, including LLMs
Keywords:
Large Language Model Agents, Heterogeneous Behavior Modeling, Consensus-Guided Residual Modeling, Prediction Markets, Agent-Based Simulation

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_GWopOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This work propose to study trader behaviors by modeling heterogeneity, defined as their divergent from market consensus, in a structured way. In doing so the proposed method compute the behaviour residual of a particular trader and use it for simulating its behaviors. The proposed method uses estimation from PolyMarket and evaluated using Monte Carlo replays, showing better performance in modeling behavior aggregatively than baselines.

Critical strengths

The overall question and idea is okay. While the application may be limited, this work presents a neat, self-containing workflow that is somewhat inspiring. Especially, the idea of modeling players by considering their heterogeneity from a statistical point of view is useful.

This study is okay with selected underpinning models as well as baselines. The construction of experiments is okay, although it could benefit from having more details and more models, as well as more experiments.

Overall this work gives an interesting read. It would be, still, strengthen further.

Critical weaknesses

However there are two noticeable limitation of this work: (1) the modeling focuses on observations visible in the market, and the heterogeneity of a particular player (wallet) is from what is observed from its behaviour, not intrinsic to that player. (2) the construction of proposed method is specific to PolyMarket style binary market (only yes and no) with only buy and sell actions. It does not explicitly modeling the impact of external forces, and only see the market changes in general.

The choice of (1) excludes modeling of more complex real work market. Along this line, heterogeneity are defined as more deeply as what intrinsic to the players (to name a few works, https://aclanthology.org/2024.acl-long.829/ https://arxiv.org/abs/2505.17648). It is oaky for this work to select a scope, but this also means proper discussion is needed for correctly positioning this work and its limitation.

The modeling choice of (2) also limits the proposed work in explaining more complex market structure. For example it does not explicitly model external forces and the change of consensus over time, nor it models collective behaviors in subset of players (e.g. collusion, https://arxiv.org/abs/2404.00806 ).

Presentation Issue: Fonts size in Figure 2 is too small for reading.

Questions for authors

None

2 comments
AuthorsOpenReview ↗

Response to Reviewer GWop

We sincerely thank the reviewer for the positive assessment and constructive suggestions. In response, we have conducted additional experiments, clarified the methodological scope, and revised the manuscript and appendix accordingly. Below, we provide point-by-point responses and summarize the corresponding revisions.

Regarding observed behavior versus intrinsic trader attributes, CRBM models heterogeneity observed in transaction histories rather than traders' intrinsic preferences, beliefs, utility functions, or personalities. The residual represents a wallet's persistent observed deviation from past market consensus and should therefore be interpreted as an empirical behavioral representation rather than a causal explanation of trader motivation.

We had previously explored a richer four-dimensional representation by adding two past-only behavioral statistics capturing preference for trading at more certain prices and breadth and intensity of market participation. The extension slightly improves test Outcome Macro-F1 from 0.6114 to 0.6174 but does not consistently improve Action F1, Strict Accuracy, or validation performance. We therefore adopted the compact two-dimensional representation in CRBM.

Importantly, the specific choice and number of residual dimensions are not the primary focus of this work. CRBM is not restricted to two dimensions and can incorporate richer behavioral representations in future research. We have revised the terminology to emphasize observed behavioral heterogeneity and distinguish our empirical representation from intrinsic participant attributes. These clarifications appear in the Introduction and Section 5, while the exploratory four-dimensional results are reported in Appendix C.3 and Table 10.

Regarding the scope beyond binary prediction markets, the current empirical evidence is restricted to binary Yes/No Polymarket markets. This setting provides a structured environment for studying consensus and individual deviation, but the reported results should not be generalized directly to equities, futures, or continuous limit-order-book markets.

CRBM is evaluated here as a behavioral decision layer rather than a complete order-book simulation system. Extensions to richer market states and output spaces require appropriate data and remain future work. Section 5 states the current empirical scope and its limitations.

Regarding changing consensus, external forces, and coordinated behavior, CRBM dynamically recomputes cic_i for every target transaction using only earlier transactions from other wallets in the same market. The pre-trade state xix_i is constructed exclusively from information available before tit_i (Section 3 and Appendix A.1).

External forces such as news, political events, and liquidity shocks are represented only indirectly through observed market states and historical order flow. CRBM neither provides a causal model of these forces nor models coordinated behavior among subsets of wallets, such as collusion or coalition formation. The descriptive groups in Appendix C.6 summarize population structure rather than detected coalitions. Section 5 now states these limitations and identifies structured event signals, closed-loop market simulations, and coordinated multi-wallet behavior as future extensions.

Regarding experimental coverage and presentation, we added JANUS and CoPL to the common market-replay protocol, three-seed training results, a six-setting replay sensitivity analysis, and an enlarged Figure 2. CRBM outperforms JANUS on all six replay metrics and CoPL on five of six; its RMSE remains within the top two methods across all six sensitivity settings. These additions strengthen the evidence without extending the claim beyond the controlled historical-replay setting.

The expanded replay comparison appears in Table 3, the robustness analyses in Appendices B.3 and C.2, and the enlarged visualization appears in Figure 2.

Thanks for your response to my reviews. I think the response slightly strengthens points where I raise questions. Yet my concerns on the limitation and scope remains. So I'd like to keep my current Weak acceptance rating.

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

An interesting paper if LLM-based trading agents can reproduce persistent individual differences rather than converge towards average market behaviour. The paper proposes Consensus-Guided Residual Behavior Modeling which generates heterogeneous trading behaviour in prediction markets. The authors evaluate CRBM using Polymarket transaction data, comparing it against rule-based, conventional machine learning, and LLM-based baselines. The evaluation covers transaction-level prediction, preservation of wallet-specific behaviour, population-level heterogeneity, and historical market replay. The paper's strongest claim concerns individual behavioural fidelity, but much of the evidence concerns prediction of transactions.

Critical strengths

The decomposition of trading behaviour into market consensus and wallet-specific deviations. The use of historical transactions and the ablation experiments.

The paper evaluates behavioural fidelity at multiple levels, including held-out wallet behaviour, population heterogeneity, and aggregate market dynamics.

Critical weaknesses

The proposed behavioural representation relies on only two features: contrarian tendency and relative trade size. It is unclear whether these are sufficient to capture individual trading behaviour, which may also depend on risk preferences, trading strategies, and market conditions. The main performance comparisons in Table 1 are reported on a validation set of only 13 wallets, which is also used for model selection. Although the authors have a separate test set of 26 wallets, they do not report the same transaction-level comparisons on this set. This makes it difficult to assess whether the reported improvements generalise to unseen traders, particularly given the small validation sample.

Questions for authors

Are the improvements over the strongest baselines statistically significant and stable across different training seeds and wallet splits? To what extent do the historical replay results depend on the predetermined market contexts, price-update rules?

3 comments
AuthorsOpenReview ↗

Response to Reviewer arnu 1/2

We sincerely thank the reviewer for the constructive questions and valuable suggestions. In response, we have conducted additional experiments, clarified the methodological scope, and revised the manuscript and appendix accordingly. Below, we provide point-by-point responses and summarize the corresponding revisions.

Regarding whether two residual dimensions are sufficient, we had previously explored alternative behavioral representations with additional dimensions and found no consistent performance improvement.

Specifically, we extended the two-dimensional residual with two additional past-only behavioral statistics capturing preference for trading at more certain prices and breadth and intensity of market participation, while keeping all other settings unchanged. The results are reported below.

SplitModelParseAct. Acc.Act. F1Out. Acc.Out. F1StrictSize Log-MAE
ValidationCRBM-2D1.00000.87220.54770.57690.55210.47071.1140
ValidationCRBM-4D1.00000.83150.51120.50770.50000.42511.2874
TestCRBM-2D1.00000.64510.45270.61180.61140.42661.5764
TestCRBM-4D1.00000.59070.38930.60550.61740.42231.5801

CRBM-4D improves test Outcome Macro-F1 from 0.6114 to 0.6174 but does not consistently improve the remaining metrics. We therefore adopted the compact two-dimensional representation in CRBM. The specific choice and number of residual dimensions are not the primary focus of this work; the framework can be extended to incorporate richer behavioral representations. See Appendix C.3 and Table 10.

Regarding generalization and statistical significance, we evaluated the fixed submitted checkpoint against P-Soups on 26 unseen wallets (2,316 transactions) and performed 10,000 paired wallet-clustered bootstrap replicates. The test set was not used for training or checkpoint selection.

MethodAct. Acc.Act. F1Out. Acc.Out. F1StrictSize Log-MAE
P-Soups0.64510.40230.55100.54830.37821.4577
CRBM-2D0.64510.45270.61180.61140.42661.5764

CRBM obtains higher point estimates for Action F1, Outcome Accuracy, Outcome Macro-F1, and Strict Accuracy, while P-Soups retains a lower Size Log-MAE (1.4577 versus 1.5764). Its Outcome Macro-F1 and Strict Accuracy improvements are 0.0631 and 0.0484, with 95% confidence intervals [−0.0028,0.1743][-0.0028,0.1743] and [−0.0263,0.1633][-0.0263,0.1633]. Both intervals include zero, so we interpret CRBM and P-Soups as statistically comparable on the current 26 test wallets rather than claiming a significant advantage.

The results indicate competitive generalization to unseen wallets, while the limited number of wallet clusters prevents a claim of statistically significant superiority over P-Soups. The complete comparison and bootstrap are reported in Appendix C.1, Tables 7 and 8.

AuthorsOpenReview ↗

Response to Reviewer arnu 2/2

We next address training-seed stability and dependence on the historical-replay mechanism.

Regarding training-seed stability, we independently retrained CRBM with seeds 42, 43, and 44 and found low variation in the classification and joint-order metrics. Table 1 reports the submitted checkpoint on the 13-wallet validation set, whereas the retrained models are evaluated on the separate 26-wallet test set. The seed-42 entry below is a newly retrained model rather than the submitted checkpoint reported in Appendix C.1.

MetricSeed 42Seed 43Seed 44Mean ± Std.
Action F10.43970.44890.45360.4474 ± 0.0071
Outcome Macro-F10.59430.58820.57860.5870 ± 0.0079
Strict0.40880.41750.41280.4130 ± 0.0044
Size Log-MAE1.50511.52601.44111.4907 ± 0.0442

The standard deviations are at most 0.0079 for Action F1, Outcome Macro-F1, and Strict Accuracy; Size Log-MAE shows greater variation (0.0442) but remains within a similar range. The retrained seed-42 model and submitted fixed checkpoint are separately trained artifacts. See Appendix B.3 and Table 6. Full retraining under new wallet partitions remains outside the current experiment.

Regarding dependence on replay assumptions, we evaluated six configurations across Monte Carlo seeds 40–44 and found that CRBM remains among the top two methods in RMSE under every configuration.

ConfigurationPrice impactActivation probabilitiesUpdate rule
Reported setting0.370.10 / 0.50 / 0.85Logit
Low impact0.200.10 / 0.50 / 0.85Logit
High impact0.500.10 / 0.50 / 0.85Logit
Low activity0.370.05 / 0.25 / 0.60Logit
High activity0.370.20 / 0.70 / 1.00Logit
Probability update0.370.10 / 0.50 / 0.85Probability

The three activation probabilities correspond to the pre-election, vote-counting, and result-largely-certain phases. CRBM achieves the lowest RMSE in three configurations, the lowest MAE and DTW in four, and the highest price-change correlation in five. Thus, its replay performance is not driven by a single price-impact coefficient, activation schedule, or update rule.

The historical market sequence is fixed to isolate replay-mechanism sensitivity, so this experiment does not establish generalization across markets or political events. The six configurations are listed in Table 9, while their aggregate outcomes and scope are summarized in Appendix C.2.

AuthorsOpenReview ↗

Brief Update on the Revised Manuscript

We would like to briefly follow up and note that the revised manuscript now incorporates the additional evaluations described in our response. Specifically, we added the controlled two- versus four-dimensional residual comparison (Appendix C.3, Table 10), the fixed-checkpoint evaluation on 26 unseen wallets and wallet-clustered bootstrap analysis (Appendix C.1, Tables 7–8), the three-seed stability experiment (Appendix B.3, Table 6), and the six-configuration replay sensitivity analysis summarized in Appendix C.2, with the configurations listed in Table 9. We also clarify that the bootstrap intervals include zero and therefore do not claim statistically significant superiority over P-Soups on the current 26-wallet test set. We hope this brief update helps locate the corresponding revisions in the updated manuscript.

Reviewer_5KAXOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces a residual-decomposition-based approach, CRBM, to characterize trader heterogeneity in order flow simulation. The experimental results suggest that this modeling approach can relatively effectively capture real traders’ trading actions and reconstruct historical order flow trajectories.

Critical strengths

Compared with commonly used approaches based on assumptions about various economic preferences, trader heterogeneity characterized through this decomposition is more closely connected to real markets. Because it is not derived entirely from particular empirical assumptions, but is grounded in the observed behavior of real traders.

Critical weaknesses

The most serious issue with this paper is that, although inferring trader behavior from real transactions is a good idea, its implementation still relies on highly simplified modeling: a simple summary of the current market and simple ratios calculated from the historical transactions of selected wallets—the behavioral residuals described in the paper—are used to characterize heterogeneous traders and the basis of their decisions. This raises a fundamental concern: when making decisions, the LLM still has a vague representation of the trader it is supposed to imitate, consisting only of a historical decision ratio. Such a reconstruction is not rigorous and does not align with the behavioral logic of market participants, because it is unclear what drives the generated trading orders. At the level of individual orders, I find that the effectiveness of these behavioral residuals in characterizing trader behavior lacks interpretability: the model is told “how often the trader went against the consensus in the past,” but not “under what circumstances the trader went against the consensus.”

Questions for authors

1.Section 4.4 states that CRBM achieves an Outcome Accuracy of 0.6101 and an Outcome F1 of 0.6078. However, the corresponding figures in the appendix table are 0.5101 and 0.5278, respectively. Why do these results differ? Please provide a detailed explanation. 2.Is this approach to modeling heterogeneity compatible with existing order book systems? If incorporated into their architectures—for example, by replacing preference-based trader representations in some order book simulation systems—could it improve their market simulation performance? Table 1 includes JANUS and CoPL, but neither is included in the market replay results in Table 3. Could you provide their results under the same replay protocol to assess the market-level advantages of CRBM’s heterogeneity modeling approach over these methods?

2 comments
AuthorsOpenReview ↗

Response to Reviewer 5KAX

We sincerely thank the reviewer for the constructive comments and valuable suggestions. In response, we have conducted additional experiments, clarified the methodological scope, and revised the manuscript and appendix accordingly. Below, we provide point-by-point responses and summarize the corresponding revisions.

Regarding whether two residual dimensions are sufficient, we had previously explored alternative behavioral representations with additional dimensions and found no consistent performance improvement.

CRBM also conditions generation on the pre-trade market state xix_i and past-only market consensus cic_i; therefore, decisions are not generated from the two residual features alone. Component ablations in Table 2 show that using either dimension alone reduces Outcome Macro-F1 to 0.5093 or 0.4829, compared with 0.5521 for the complete model.

Specifically, we extended the two-dimensional residual with two additional past-only behavioral statistics capturing preference for trading at more certain prices and breadth and intensity of market participation, while keeping all other settings unchanged. The results are reported below.

SplitModelParseAct. Acc.Act. F1Out. Acc.Out. F1StrictSize Log-MAE
ValidationCRBM-2D1.00000.87220.54770.57690.55210.47071.1140
ValidationCRBM-4D1.00000.83150.51120.50770.50000.42511.2874
TestCRBM-2D1.00000.64510.45270.61180.61140.42661.5764
TestCRBM-4D1.00000.59070.38930.60550.61740.42231.5801

The four-dimensional variant provides a small test Outcome Macro-F1 gain but no consistent improvement in complete-order generation. We therefore adopted the compact two-dimensional representation in CRBM. Importantly, the specific choice and number of residual dimensions are not the primary focus of this work; the framework is not restricted to two dimensions and can incorporate richer behavioral representations in future research. See Appendix C.3 and Table 10.

Regarding the inconsistent Outcome Accuracy and Outcome Macro-F1 values, the Section 4.4 results—0.6101 Outcome Accuracy and 0.6078 Outcome Macro-F1—are correct; the appendix contained transcription errors.

HistoryNOutcome Acc.Outcome Macro-F1Size Log-MAE
Overall9490.61010.60781.2133
Long7120.62220.61851.1980
Medium1640.59760.60451.2375
Limited730.52050.52251.3079

The subgroup-weighted Outcome Accuracy reproduces 0.6101 after rounding; overall Macro-F1 was recomputed from all 949 predictions. Training and predictions were unaffected. We have corrected the appendix values and audited the corresponding numerical references. The corrected cross-platform results are reported in Section 4.4 and Appendix C.7, Table 14.

Regarding compatibility with order-book systems, CRBM can in principle serve as an agent decision layer without replacing the external matching engine. CRBM is evaluated here as a behavioral decision layer rather than a complete order-book simulation system. Extensions to richer market states and output spaces require appropriate data and remain future work. Section 5 states the current empirical scope and its limitations.

Regarding the omitted market-replay baselines, we evaluated JANUS and the CoPL baseline under the same replay protocol as CRBM. All methods use the same 105 wallets and 134 decision scenarios under identical wallet-activation, price-impact, and price-update mechanisms (Section 4.3, Table 3). Appendix C.2 additionally varies these mechanisms across six configurations using Monte Carlo seeds 40-44.

MethodRMSEMAEUp/DownFlow Corr.Change Corr.DTW
JANUS0.13840.05880.5078-0.4945-0.49220.0307
CoPL0.11430.07040.50780.20930.16480.0437
CRBM0.04410.03620.52340.14990.21330.0095

Lower is better for RMSE, MAE, and DTW. CRBM outperforms JANUS on all six metrics and CoPL on five of six, while CoPL has the higher flow correlation. The expanded comparison and common replay protocol are now reported in Section 4.3 and Table 3.

AuthorsOpenReview ↗

Brief Update on the Revised Manuscript

We would like to briefly follow up and note that the revised manuscript now incorporates the corrections and additional evaluations described in our response. Specifically, we added the controlled two- versus four-dimensional residual comparison (Appendix C.3, Table 10), corrected the Manifold results and audited the corresponding numerical references (Section 4.4; Appendix C.7, Table 14), clarified CRBM’s compatibility with order-book systems and its empirical scope (Section 5), and added JANUS and CoPL to the common market-replay comparison (Section 4.3, Table 3). We hope this brief update helps locate the corresponding revisions in the updated manuscript.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.