acceptodds
Under review as a conference paper at ICLR 2027

From Market Facts to Executable Markets: Grounding Heterogeneous Trader Populations in Empirical Market Evidence

Abstract

Data-driven financial market models can reproduce or predict observable outcomes, such as limit order book states and order flows, but they provide limited access to the participant-level mechanisms that generate them. Agent-based simulators make such mechanisms explicit, yet typically begin from prescribed trader populations rather than structures constrained by real market evidence. We address this gap by treating a heterogeneous, executable trader population as the object of calibration. It jointly specifies capital allocation across traders, asset allocation, strategy influence, and strategy-specific behavioral parameters, allowing changes in participant structure to propagate endogenously through order generation, limit-order-book execution, price formation, and resource updates. We use observed stylized facts to restrict the space of plausible populations and search for candidate markets whose simulated behavior is consistent with these observations. To make this high-dimensional problem tractable, we reparameterize constrained allocations and factorize high-dimensional residual variation, reducing the search dimension while retaining the full set of traders in every simulation. Experiments show that the resulting markets reproduce key financial stylized facts and achieve better held-out matching than alternative search strategies under comparable simulation budgets. These solutions should be viewed as evidence-consistent executable markets rather than unique reconstructions of the real trader population, providing controlled starting points for studying how changes in resources, strategies, and behavior propagate through endogenous market dynamics. This work reframes financial market simulation from reproducing market outputs toward empirically grounded experimentation on the mechanisms that generate them.

Primary area:
learning on time series and dynamical systems
Keywords:
Multi-Agent Systems, Trader Populations, Simulation-Based Calibration, Financial Market Simulation, Stylized Facts

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

The paper proposes MarketPop, a framework for calibrating executable heterogeneous trader populations to empirical market statistics. It represents endowments, asset allocations, strategy weights, and behavioral parameters through a structured low-dimensional parameterization, and introduces Progressive Hierarchical Inference to calibrate them. Experiments in an equity-market simulator compare calibration methods and population-level ablations using five stylized facts, with evaluation on a short subsequent continuation period.

Critical strengths

The framework connects population calibration to executable trading behavior through a structured parameterization. Comparisons with direct parameter search and joint search in the structured space show improved calibration, supporting the practical value of the proposed procedure within the tested simulator.

Critical weaknesses
  1. The incremental benefit of the full population hierarchy is unclear. In Table 2, endowment-only calibration achieves a held-out discrepancy of 4.85, compared with 4.89 for the full method. Although the full method improves calibration fit, this comparison does not establish a generalization benefit from the additional levels.

  2. Validation is limited to a five-day calibration horizon and three subsequent continuation days. This provides limited evidence that the calibrated population remains effective across different periods or independent simulation runs.

3.Don't talk about some related works like finevo.

Questions for authors
  1. What evidence supports the contribution of the higher population levels beyond improved calibration fit, given the endowment-only result in Table 2?
  2. Was the selected population evaluated with fresh simulation seeds independently of those used during calibration? If so, please report the resulting variability and comparison with endowment-only calibration.
4 comments
AuthorsOpenReview ↗

Q1: Role of the additional layers beyond calibration fit

We thank you for the careful reading and constructive suggestions.

After fixing the resource-scale, strategy-execution, and order-blocking issues of the original experiments, we reran the main experiments and the four-layer ablation under the revised common setup, focusing on whether each layer's role persists into subsequent simulation.

The original π1\pi_1-only comparison is retained as a main-experiment baseline; its updated calibration results, together with Full, are listed in Table 1 of the general statement (Comment 5). The layer ablation was changed from progressively enabling layers to a leave-one-layer-fixed design: from the same initial state, π1\pi_1, π2\pi_2, π3\pi_3, or θ\theta is excluded from calibration in turn, and each variant is compared with Full.

After the 10-day calibration, each selected configuration is frozen and evaluated over 60 held-out days with ten test seeds unused in calibration; we report the averages. "w/o" means the layer remains in the population but is not calibrated; the five Facts and five dynamics metrics all come from the held-out period.

Table R3-Q1: Calibration error and held-out performance of the four-layer ablation.

GroupQ1Q_1Q2Q_2Q3Q_3Q4Q_4Q5Q_5H1H_1H2H_2H3H_3H4H_4H5H_5dcald_{\mathrm{cal}} ↓dHOd_{\mathrm{HO}} ↓
w/o π1\pi_1-0.7712-0.4347-0.9287-0.8004-0.13030.29180.3118-0.33350.0023-0.09532.18143.7439
w/o π2\pi_2-1.3560-0.1389-0.78960.4814-0.13030.28110.1366-0.34000.0667-0.12183.69958.2171
w/o π3\pi_3-1.0938-0.44040.0386-0.9079-0.13030.27910.1308-0.1315-0.2908-0.29722.78965.5475
w/o θ\theta-0.8313-0.4337-0.8975-1.2422-0.13030.24240.2043-0.2098-0.4019-0.22382.91915.6391
Full-0.6555-0.5162-0.1250-0.6814-0.1303-0.0930-0.07560.06160.0564-0.03901.84743.1461

All four ablations have higher held-out losses than Full, so from the same initial state, disabling any single layer worsens overall subsequent performance. Without π1\pi_1, deviations in absolute-return autocorrelation and some order-book metrics widen; without π2\pi_2, the match in volatility and absolute-return autocorrelation worsens and the held-out loss is the highest. Without π3\pi_3, deviations in volatility, the absolute-return–volume relation, volatility persistence, spread, and drawdown increase; without θ\theta, absolute-return autocorrelation, the absolute-return–volume relation, and the five dynamics metrics all deviate more. The dynamics metrics never entered the calibration objective, so their changes provide evidence of the higher layers' contribution beyond the targeted Facts. In the revised setup, the benefits of the full hierarchy appear in both calibration fit and overall held-out loss, as well as in specific market characteristics on independent test paths.

AuthorsOpenReview ↗

Q2: Independent seeds, run variability, and the comparison with -only

Yes. After calibration, we froze the selected configuration and ran 60 trading days of held-out evaluation with ten new seeds unused in calibration. No further calibration occurs during testing, and test results play no role in configuration selection. The table reports all ten runs of Full in the January window, together with the five Facts, the five dynamics metrics, and the held-out Loss.

Table R3-Q2: Ten independent held-out runs of Full.

RunQ1Q_1Q2Q_2Q3Q_3Q4Q_4Q5Q_5H1H_1H2H_2H3H_3H4H_4H5H_5dHOd_{\mathrm{HO}} ↓
1-0.7180-0.51270.4409-0.5546-0.1303-0.5713-0.4666-0.0293-0.4030-0.30785.0827
2-0.6819-0.51950.2798-0.7721-0.13030.3560-0.1907-0.29590.21800.21084.9028
3-0.7207-0.4936-0.4803-1.0212-0.13030.4010-0.44740.23480.2754-0.49343.6900
4-0.6364-0.53750.0487-0.6224-0.1303-0.4846-0.5489-0.56360.79340.89282.9319
5-0.7538-0.5075-0.6247-0.4629-0.1303-0.5549-0.04790.7758-0.2265-0.66332.2908
6-0.6667-0.51570.1008-0.4333-0.1303-0.4711-0.83020.5960-0.6321-0.65642.5802
7-0.6324-0.5146-0.1350-0.7130-0.13030.01730.0445-0.27800.40090.14712.5961
8-0.4773-0.5099-0.2164-0.8553-0.13030.20040.7649-0.16980.32500.34702.8211
9-0.6703-0.5352-0.2500-0.6684-0.1303-0.13710.17020.45200.11000.08092.3540
10-0.5971-0.5163-0.4139-0.7113-0.13030.31460.7963-0.1064-0.29740.05252.2117
Mean-0.6555-0.5162-0.1250-0.6814-0.1303-0.0930-0.07560.06160.0564-0.03903.1461
Min-0.7538-0.5375-0.6247-1.0212-0.1303-0.5713-0.8302-0.5636-0.6321-0.66332.2117
Max-0.4773-0.49360.4409-0.4333-0.13030.40100.79630.77580.79340.89285.0827

On these calibration-independent paths, the five Facts lie essentially within one standard deviation of their reference means, with only isolated runs slightly outside. The five dynamics metrics show more pronounced variation across runs, and some deviations change sign; nevertheless, all lie within one reference standard deviation in every run, although they were not part of the calibration objective. Held-out loss ranges from 2.2117 to 5.0827, with the first two runs relatively high and the rest closer together. These results indicate some repeatability of the selected configuration's held-out performance alongside path-to-path variability; we report the ten-run average together with the per-run range so that the evaluation does not hinge on any single run. (Min/Max are computed per column and do not correspond to any single run.)

For the comparison with π1\pi_1-only, we included it as a baseline in the updated main experiments. Under the common setup and budget, endowment-only calibration still exhibits extensive price-limit pinning and a much poorer calibration fit than Full; the calibration-period results are given in Table 1 of the general statement (Comment 5). The ten independent tests above show Full's subsequent performance and run-to-run variability after calibration.

AuthorsOpenReview ↗

Weakness: short validation window and cross-period performance

We ran the same calibration procedure separately in two windows starting in January and July 2024, each extended to 10 trading days of calibration followed by 60 of held-out evaluation:

  • January window: Jan 2–Apr 18, 2024 — calibration Jan 2–15, held-out Jan 16–Apr 18.
  • July window: Jul 1–Oct 15, 2024 — calibration Jul 1–12, held-out Jul 15–Oct 15.

These spans count only open trading days on the Shanghai and Shenzhen markets, excluding weekends and holiday closures. Each window searches and selects its configuration on its own calibration period; the configuration is then frozen, with no further calibration during the held-out period.

Table R3-W: Calibration error and held-out performance in the two historical windows.

WindowQ1Q_1Q2Q_2Q3Q_3Q4Q_4Q5Q_5H1H_1H2H_2H3H_3H4H_4H5H_5dcald_{\mathrm{cal}} ↓dHOd_{\mathrm{HO}} ↓
Jan 2024-0.6555-0.5162-0.1250-0.6814-0.1303-0.0930-0.07560.06160.0564-0.03901.84743.1461
Jul 2024-0.5621-0.4258-0.1877-0.7014-0.13030.1091-0.03570.1182-0.1758-0.38751.76113.7535

Both windows achieve comparable calibration and held-out performance, with overall errors of similar magnitudes; the five Facts deviate similarly, and the five dynamics metrics all lie within one standard deviation of their reference means. The same procedure, calibrated separately in the two periods, thus achieves a good subsequent market match in both. Together with the independent-seed results in our reply to Q2 above, this round of evaluation both broadens the time coverage and examines the repeatability of a fixed configuration across random paths.

AuthorsOpenReview ↗

Weakness: related work and the relationship to FinEvo

Thank you for pointing out this omission. We agree that FinEvo should be discussed more explicitly.

FinEvo introduces a closed-loop market framework for evaluating financial trading agents, capturing how their interactions, strategy competition, and capital reallocation shape market outcomes. MarketPop addresses a complementary problem: constructing and calibrating complete heterogeneous trader populations using empirical market evidence. Specifically, MarketPop jointly calibrates resource ownership, asset allocation, strategy composition, and behavioral parameters through structured population representations and hierarchical search. While both works connect participant-level mechanisms with market dynamics, FinEvo focuses on agent evaluation and evolutionary dynamics, whereas MarketPop focuses on the empirical grounding of executable trader populations.

We will add a discussion of FinEvo in the revised related-work section and clarify its relationship to MarketPop, particularly the differences in their research objectives and methodological approaches.

Reviewer_h8hQOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces MarketPop, a framework for calibrating executable heterogeneous trader populations against empirical market statistics. The population is represented through four coupled levels: initial endowment allocation, asset allocation, strategy influence, and behavioral parameters. To make calibration tractable, the framework combines constraint-preserving reparameterizations with low-rank residual representations, while decoding every proposal into a complete trader population before simulation. The proposed Progressive Hierarchical Inference procedure alternates level-specific and cross-level updates and allows representation ranks to adapt during search. Experiments in an A-share market simulator show improved matching of five targeted stylized facts relative to the reported population templates and internal search baselines, with an additional three-day continuation evaluation.

Critical strengths
  1. Calibrating explicit trader populations enables controlled analysis of how participant structure affects market dynamics through an executable trading process.
  2. The framework separates population and search representations, clearly specifies constraints and decoding, and appropriately acknowledges non-identifiability.
  3. Table 1 shows successive improvements from Direct-X to Joint-Z to MarketPop, supporting the structured representation and search procedure in the evaluated setting.
Critical weaknesses

W1. The experimental evaluation is limited, particularly in validating the framework’s use for mechanism analysis. The paper emphasizes using calibrated markets to study changes in participant structure, but the main experiments focus on matching five stylized facts. It would be useful to apply the same changes in resource allocation or strategy composition to populations with similar calibration quality but different internal configurations, and compare their market responses. This would clarify whether the resulting mechanistic conclusions depend on the particular population selected. W2. The added value of the core design is not sufficiently established. In Table 2, calibrating only the initial endowment allocation achieves a held-out discrepancy of 4.85, compared with 4.89 for the full four-level model. Moreover, no rank changes were accepted during the main experiment. Thus, although the full model improves calibration fit, the practical benefits of the richer population hierarchy and adaptive representation capacity remain unclear.

Questions for authors

The appendices contain considerable repetition, frequently following the same structure of defining an object, stating a property, and explaining what it does not guarantee. More direct and concise statements would improve readability, and material essential to understanding the core method should be moved into the main text. Furthermore, the appendices extensively document standard approximation properties, decoder stability, and the unbiasedness of the fixed-candidate score. However, their connection to PHI’s actual optimization performance and post-selection generalization remains weak, leaving the independent theoretical contribution unclear. The authors should clarify how these analyses support the core method and streamline derivations and repeated explanations that have limited relevance to the main conclusions.

4 comments
AuthorsOpenReview ↗

Question: theoretical positioning, empirical linkage, and paper organization

We thank you for the careful reading and constructive suggestions.

Our main contributions are the population representation, the calibration method, and its empirical validation. The theoretical properties support the construction of the method.

(1) Approximation and decoding analyses. The full population contains traders' resource shares, asset allocations, strategy weights, and behavioral parameters. Low-rank factors represent this high-dimensional heterogeneity through shared directions while retaining individual differences. The approximation analysis relates finite-capacity representation error to the discarded singular-value tail, clarifying the capacity–compression trade-off; the required capacity is assessed through simulation. Decoding uses normalization, range restoration, and type checks to produce an executable population with shares summing to one and valid behavioral parameters. Its stability analysis explains how small coordinate changes propagate to actual parameters, clarifying the space used by local search. Market behavior emerges through strategy execution and order matching, so candidates are evaluated by simulation. Joint-Z (structured) fits clearly better than Direct-X (dense) under the same budget, demonstrating the representation's practical value.

(2) PHI's role under a common planned budget. Even with the structured representation, joint optimization over all coordinates must reconcile the interactions among resources, strategies, and behavior. PHI interleaves per-layer specialized updates with cross-layer joint search, and cross-stage inheritance lets later stages continue from previously found configurations, reusing compatible historical evaluations and search directions. Joint-Z and Full share the representation and budget, with the main difference being joint optimization over all coordinates versus PHI's layered search organization; π1\pi_1-only calibrates one layer under the same budget.

Table R2-Q: Calibration loss across representations and search schemes.

MethodRepresentationSearch schemedcald_{\mathrm{cal}} ↓
Direct-XDense full parametersDirect joint search228.7481
Joint-ZStructured coordinatesJoint over all coordinates25.3443
π1\pi_1-onlyStructured coordinatesOnly π1\pi_1; other layers frozen212.1733
FullStructured coordinatesPHI layer-wise + cross-layer search1.8474

Despite searching fewer parameters, π1\pi_1-only still exhibits price-limit pinning and fits much worse than Full, showing the need for cross-layer coordination in this setting. Joint-Z improves on Direct-X, demonstrating the practical value of the structured representation; Full improves further with PHI's specialized updates, cross-layer coordination, and reuse of search progress. The approximation and decoding analyses explain how the representation is constructed, while the main experiments show the calibration results obtained by combining it with the search procedure under the common planned budget.

(3) Scoring analysis and independent evaluation. A candidate's metrics fluctuate across simulation paths. For a candidate fixed independently of the evaluation randomness, the cross-product of two independent paths' deviations cancels fluctuation-induced cross terms in expectation. Under the paper's assumptions, it targets the error between expected market metrics and the real-data target, explaining the scoring objective and two-path construction. Since calibration repeatedly searches and selects configurations using this score, we test the selected population with fresh seeds unused in calibration. Main experiments and ablations report averages over ten independent 60-day held-out runs. Full matches both the five Facts and additional dynamics metrics well across these paths, with run-to-run variation. Disabling any layer increases held-out error, supporting four-layer coordination on independent paths. Both historical windows yield similar held-out errors; the dynamics metrics, excluded from calibration, check behavior beyond the target Facts.

(4) Organization. We will streamline derivations and repeated explanations and move material essential to the method into the main text. We apologize for the difficulty our presentation caused in distinguishing the early exploratory runs from the final experiments. We will make that distinction clear. We will rewrite the experimental sections with the new setup and results, and update the parameter configurations, evaluation protocol, and reproducibility notes accordingly; the method part will focus on the population representation and the core PHI procedure, and the appendix will merge duplicated derivations and reorganize the supplementary material. The adaptive rank will be described according to the actual records.

AuthorsOpenReview ↗

Weakness: mechanism analysis across compatible populations (1/2): fit and internal composition

To examine how much the choice of population matters, we first compared the strategy allocations of the five configurations with the best calibration scores across the full search history. We froze each configuration before evaluating it on held-out paths, so selection depended only on calibration performance.

(1) Fit of the five candidates. Calibration and held-out results:

Table R2-Ma: Held-out metrics and the two losses of the five candidates.

CandidateQ1Q_1Q2Q_2Q3Q_3Q4Q_4Q5Q_5H1H_1H2H_2H3H_3H4H_4H5H_5dcald_{\mathrm{cal}} ↓dHOd_{\mathrm{HO}} ↓
1-0.6555-0.5162-0.1250-0.6814-0.1303-0.0930-0.07560.06160.0564-0.03901.84743.1461
2-0.6796-0.5235-0.1291-0.6782-0.1303-0.3075-0.22600.24560.0565-0.13131.89743.1081
3-0.5519-0.5145-0.1236-0.6160-0.13030.14870.1572-0.0882-0.2936-0.17441.89862.9743
4-0.6416-0.5094-0.1182-0.6690-0.13030.33700.1184-0.0494-0.0874-0.09281.92433.5537
5-0.6439-0.4923-0.1206-0.6280-0.13030.07440.1482-0.1284-0.3800-0.02021.94743.3535

The five candidates maintain similar overall errors in both the calibration period and held-out evaluation, so the good fit and subsequent performance do not rely on the single final selection. Per metric, lag-1 absolute-return autocorrelation and the price-limit frequency are close across candidates, whereas volatility persistence, order-book dynamics, and spread differ noticeably. Similar overall fit can correspond to distinct market dynamics, which makes the cross-candidate mechanism comparison you suggested practically meaningful.

(2) Internal strategy composition. We computed the fraction of capital controlled by each of the seven strategy types—execution, fundamental, market-making, momentum, noise, graph, and LLM. Rows sum to 1 up to rounding.

Table R2-Mb: Strategy capital composition of the five candidates before perturbation.

CandidateExecutionFundamentalMarket-makingMomentumNoiseGraphLLM
10.11490.22650.10150.15950.31580.06000.0218
20.11220.22270.10670.16060.31990.05670.0212
30.11370.22590.10230.15940.31780.06080.0200
40.11360.22550.10220.15930.31770.06080.0210
50.11510.22660.10150.15950.31560.06000.0218

From an equal-weight start (1/71/7 per strategy type), calibration shifts capital in the same directions in all five candidates: fundamental up markedly, noise up the most, momentum slightly up; execution and market-making down; graph and LLM down substantially. Their final capital shares are also close to one another, reflecting structural features shared by multiple well-fitting candidates.

  • Fundamental and momentum. The fundamental strategy trades on the price–value deviation; its rising share allocates more resources to value reversion, while the slight rise in momentum allocates more capital to trend following. These increases suggest that a good market match in the current model may involve both behaviors.
  • Noise. The noise share rises the most: more capital goes to random order generation, possibly absorbing dispersed demand and behavior not captured by other strategies. This describes allocation inside the model, not the white-noise intensity or the retail share of the real market.
  • Execution and market-making. Both shares fall below the equal-weight start but remain meaningful: calibration rebalances scheduled execution and two-sided liquidity provision against other strategies. We therefore take the market-making share as the perturbation target in the next comment, to observe how the market responds when its relative allocation is increased.
  • Graph and LLM. These two custom strategies lose the most capital, so under the current implementation and objective the search tends to lower their weights—possibly because their simplified information inputs or decision mechanisms limit their potential. Richer graph relationships, contextual information, and decision mechanisms are possible directions for improvement; the share decline alone does not identify the cause.

These shared shifts show that calibration forms strategy portfolios with common tendencies from the equal-weight start. Since the dynamics metrics above still differ across candidates, we further applied the same market-making perturbation to check whether these structurally similar populations also respond similarly.

AuthorsOpenReview ↗

Weakness: mechanism analysis across compatible populations (2/2): unified market-making perturbation

Across the five candidates, calibration produced similar strategy allocations, yet some held-out market dynamics differed. We therefore tested how each population responded to the same increase in the share of capital allocated to market making.

(3) Perturbation design. Our population configuration is represented by matrices and decoded into an actual strategy composition; calibration does not search each strategy's capital share directly. A perturbation with a specific economic meaning therefore requires first deriving an interpretable capital allocation from the configuration and then setting the adjustment accordingly. Given the experiment runtime and the response period, we focus on one change: increasing the market-making share.

Concretely, for each candidate we double the market-making capital share and scale down the other six shares by a common factor, keeping total capital unchanged and preserving the relative proportions among the other strategies. The same rule is applied to every candidate from its own original composition; the market-making share rises from about one tenth to about one fifth.

Table R2-Mc: Strategy capital composition of the five candidates after perturbation.

CandidateExecutionFundamentalMarket-makingMomentumNoiseGraphLLM
10.10190.20090.20300.14150.28020.05320.0193
20.09880.19610.21330.14140.28170.05000.0187
30.10080.20010.20460.14130.28160.05390.0177
40.10060.19980.20440.14110.28150.05390.0186
50.10210.20100.20290.14150.28000.05320.0193

We then compare held-out metrics before and after this intervention without recalibration.

Table R2-Md: Held-out performance of the five candidates before perturbation.

CandidateQ1Q_1Q2Q_2Q3Q_3Q4Q_4Q5Q_5dHOd_{\mathrm{HO}} ↓
1-0.6555-0.5162-0.1250-0.6814-0.13033.1461
2-0.6796-0.5235-0.1291-0.6782-0.13033.1081
3-0.5519-0.5145-0.1236-0.6160-0.13032.9743
4-0.6416-0.5094-0.1182-0.6690-0.13033.5537
5-0.6439-0.4923-0.1206-0.6280-0.13033.3535

Table R2-Me: Held-out performance of the five candidates after perturbation.

CandidateQ1Q_1Q2Q_2Q3Q_3Q4Q_4Q5Q_5dHOd_{\mathrm{HO}} ↓
1-0.8031-0.4864-0.0891-0.6834-0.13033.3245
2-0.7393-0.6167-0.1006-0.6446-0.13033.3640
3-0.4993-0.6298-0.1266-0.7633-0.13033.2499
4-0.7527-0.6369-0.1216-0.6845-0.13033.5547
5-0.7914-0.6221-0.0847-0.6300-0.13033.5354

In Tables R2-Md and R2-Me, Q1Q_1–Q5Q_5 are standardized deviations from the real-data reference (closer to 0 is closer to the reference mean); the loss keeps its original scale. Both use the same reporting reference, so the direction and magnitude of each change can be read directly.

(4) Market response. For the return distribution, return volatility and excess kurtosis decrease in most candidates, although the directions and magnitudes of the changes vary across configurations. The effect of the same adjustment on the return distribution therefore depends on the original configuration.

For volatility persistence and the price–volume relation, lag-1 absolute-return autocorrelation changes in both directions, the absolute-return–volume correlation decreases in most candidates, and price-limit frequency is unchanged. The common strategy-composition adjustment therefore produces partly consistent market responses while preserving dynamic differences across candidates.

For the overall fit, held-out loss does not improve after perturbation, and the errors remain of a similar magnitude. Since volatility is already below the reference mean, further dampening can worsen the empirical match. Increasing the market-making share thus alters market dynamics without necessarily improving the empirical match; its effect should be considered together with the original strategy portfolio and the simultaneous reduction of the other shares. Comparing the five candidates reveals both common trends and configuration differences. We will include this experiment in the revision if needed.

AuthorsOpenReview ↗

Weakness: benefits of the full hierarchy and the adaptive rank

We address the two parts separately: how the updated experiments demonstrate the role of each layer, and the actual usage record of the adaptive rank.

(1) Benefit of the full hierarchy. To examine the practical role of the additional layers, we fixed the resource-scale, strategy-execution, and order-blocking issues and improved the asset-budget initialization and the search procedure, so that adjustments at each layer are more fully transmitted into actual trading. Under these common fixes, we extended the experiments to 10 days of calibration and 60 days of held-out evaluation and reran the main experiments and the four-layer ablation.

π1\pi_1-only is retained but moved into the main experiments as a baseline. As shown in Table 1 of the general statement (Comment 5), full calibration fits much better; although π1\pi_1-only searches fewer parameters, the other three layers remain frozen, and adjusting the endowment alone still cannot eliminate price-limit pinning. Under the current setup, the population therefore needs the capacity for coordinated multi-layer adjustment.

We further ran four leave-one-layer-fixed ablations: from the same initial state, π1\pi_1, π2\pi_2, π3\pi_3, or θ\theta is excluded from calibration, and each variant is evaluated by the average of ten independent 60-day held-out runs. "w/o" means the layer remains in the population but is not calibrated; the five Facts and five dynamics metrics are all held-out results, showing whether each layer's role persists beyond the calibration period.

Table R2-H: Calibration error and held-out performance of the full model and the leave-one-layer-fixed ablations.

GroupQ1Q_1Q2Q_2Q3Q_3Q4Q_4Q5Q_5H1H_1H2H_2H3H_3H4H_4H5H_5dcald_{\mathrm{cal}} ↓dHOd_{\mathrm{HO}} ↓
w/o π1\pi_1-0.7712-0.4347-0.9287-0.8004-0.13030.29180.3118-0.33350.0023-0.09532.18143.7439
w/o π2\pi_2-1.3560-0.1389-0.78960.4814-0.13030.28110.1366-0.34000.0667-0.12183.69958.2171
w/o π3\pi_3-1.0938-0.44040.0386-0.9079-0.13030.27910.1308-0.1315-0.2908-0.29722.78965.5475
w/o θ\theta-0.8313-0.4337-0.8975-1.2422-0.13030.24240.2043-0.2098-0.4019-0.22382.91915.6391
Full-0.6555-0.5162-0.1250-0.6814-0.1303-0.0930-0.07560.06160.0564-0.03901.84743.1461

Disabling any single layer increases both calibration and held-out loss, supporting a contribution from each layer in this setting; Full has the lowest overall loss among these variants. In detail: without π1\pi_1, deviations in absolute-return autocorrelation and some order-book dynamics metrics widen; without π2\pi_2, the match in return volatility and absolute-return autocorrelation worsens and the held-out error is the highest; without π3\pi_3, deviations in volatility, the absolute-return–volume relation, volatility persistence, spread, and drawdown increase; without θ\theta, absolute-return autocorrelation, the absolute-return–volume relation, and the five dynamics metrics all show larger deviations. Different layers act on different market behaviors, and full calibration coordinates them into a better overall match.

(2) Adaptive-rank mechanism and its stated contribution. Your skepticism about the practical benefit of the adaptive representation capacity is justified. Across all experiments, only one automatic rank adjustment, from (r3,rθ)=(8,8)(r_3,r_\theta)=(8,8) to (8,4)(8,4), was ever accepted, in early exploratory runs; as the calibration procedure, parameters, and experimental configurations evolved, no further rank change was accepted. The submission did not adequately distinguish this early record from the later experiments and thus emphasized the adaptive-rank contribution beyond what the results support.

In the revision, we will remove the main-text statements presenting the adaptive rank as a main contribution, and move the mechanism description, the early adjustment, and the record of no further accepted adjustments to the appendix. The procedure still retains the conditional structure check (item 16 of Schedule Table B in the general statement, Comment 3). Subsequent experiments continue to execute it without accepting new rank changes; the appendix will document this from the actual records.

Reviewer_3vNzOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The authors propose "MarketPop," which decouples "market representations" from "search representations." It represents high-dimensional populations using precise reparameterization and factorized coordinates (ZZ), which are then mapped to complete, valid populations via a deterministic decoder. A "Progressive Hierarchical Inference" (PHI) mechanism is employed to conduct multi-stage, cross-level searches while adaptively determining factor ranks. For the calibration objective, the method utilizes a cross-product derived from two independent simulation paths to estimate the expected factual bias, thereby avoiding the simulation variance bias that arises when squaring the output of a single path.

Critical strengths

The design of a method for representing a full population—paired with a valid decoder—is of substantial value. Rather than replacing a high-dimensional population with a low-dimensional economic model, the paper allows the search coordinates ZZ to be decoded back into a complete, valid population for simulation; this effectively renders the "full heterogeneous population" a searchable object. The two-path cross-product scoring method provides an unbiased expectation for fixed candidates, preventing the reordering of candidates that might otherwise occur with single-path squared scoring due to candidate-specific noise—a feature critical for simulation-based calibration.

The experimental implementation and auditability are notably comprehensive. The paper achieves full population calibration within the context of the A-share Limit Order Book (LOB) regime, providing detailed analyses such as hierarchical and SEIS-based ablations, robustness checks, coordinate/path accounting, and a replay manifest. Compared to studies that calibrate only at the category level or focus on low-dimensional shared parameters, this "executable market + full population" approach is significantly more ambitious and establishes a practical pipeline for future mechanism analysis.

Critical weaknesses

The core empirical claims are not supported by the ablation results. While the paper’s central contribution is the use of a "full, heterogeneous population of traders" as an empirical constraint, Table 2 shows that calibrating only π1\pi_1 is sufficient to bring the five stylized facts within the empirical ranges (dcal=2.98d_{cal}=2.98). In fact, the performance on the held-out set is slightly better when using only π1\pi_1 compared to using all four layers (4.85 vs. 4.89). Furthermore, Appendix L reveals that calibration deviations are almost entirely concentrated in the initial endowment distribution, while θ\theta shifts only marginally. This suggests that, given the five stylized facts currently employed, the model primarily identifies the skew in initial endowments rather than the composition of strategies or behavioral parameters. Unless the authors can provide specific empirical targets that necessitate the inclusion of π1,π2,π3,\pi_1, \pi_2, \pi_3, and θ\theta, the paper’s claims should be narrowed to "endowment-level calibration suffices for these facts"; otherwise, the core motivation is inconsistent with the evidence.

There are also concerns regarding external validity and the fairness of the methodological comparisons. The validation is limited to a single market (China A-shares) and five assets, using a 5-day calibration period and a 3-day held-out period. Moreover, the empirical ranges are quite broad (e.g., Q2=4.562±10.835Q_2 = 4.562 \pm 10.835 and Q5=0.0129±0.099Q_5 = 0.0129 \pm 0.099); thus, the fact that the results fall within one empirical standard deviation does not necessarily imply high market fidelity. It is also unclear whether Direct-X, Joint-Z, and MarketPop utilize identical simulation budgets, candidate positions, proposal schedules, and optimizer settings. Since the parameter space for Direct-X is significantly denser than that of the structured representation, any discrepancy in computational budgets would severely confound the comparison. The absence of multi-window and multi-market analyses, as well as a lack of assessment regarding the uncertainty of the calibration and held-out results, renders the conclusions—that the proposed approach outperforms the baselines and is consistent with empirical evidence—insufficiently robust.

Questions for authors

In Table 2, the held-out results for the "π\pi-only" configuration are slightly better than those for the full four-layer model (4.85 vs. 4.89). Is this difference statistically significant, or could it simply be due to sampling uncertainty inherent in the simulated paths/held-out data? Furthermore, which specific empirical facts actually require π3\pi_3 or θ\theta to fall within the observed empirical range? If none do, would the authors consider narrowing the scope of their core contribution—shifting from "full-population calibration" to the claim that "endowment-level calibration is sufficient to capture current empirical facts"?

Did Direct-X, Joint-Z, and MarketPop all utilize the exact same simulation budget, candidate positions, launched paths, and evaluation schedule? Please specify the actual number of candidates/paths evaluated for each method. If Direct-X—due to its higher dimensionality—undergoes less thorough exploration under the same budget, then the performance comparisons in Table 1 cannot be definitively attributed to the use of structured representations or PHI.

Could the authors provide results across multiple calibration windows, different asset pools, at least one additional market regime, and longer held-out periods? The current setup—relying on a single 5-day calibration window and a 3-day held-out period—is highly susceptible to specific market conditions; results across multiple windows are essential to substantiate the claim of an "evidence-consistent executable market."

The paper notes that empirical moments identify an "equivalence set" rather than a unique micro-state, and that claims are restricted to structural features that remain stable across this set. Could the authors provide an approximation of this equivalence set and identify the population features that remain stable within it? Currently, only a single X^\hat{X} is reported, making it difficult to determine whether the concentrated adjustments in π1\pi_1 represent a robust, stable identification result or merely the outcome of a single, specific search trajectory.

7 comments
AuthorsOpenReview ↗

Q1: Role of the full hierarchy and scope of the contribution

We thank you for the careful reading and constructive suggestions.

To examine whether endowment-only calibration is sufficient and what the additional layers contribute, we reran the main experiments and ablations under the revised common setup after fixing the resource-scale, strategy-execution, and order-blocking issues.

The design also changed: π1\pi_1-only is now a main-experiment baseline. As shown in Table 1 of the general statement (Comment 5), calibrating π1\pi_1 alone still cannot eliminate price-limit pinning, whereas the full model achieves a much better calibration fit. The ablation was changed from progressively enabling layers to a leave-one-layer-fixed design: starting from the same initial state, calibration of π1\pi_1, π2\pi_2, π3\pi_3, or θ\theta is disabled in turn, and each variant is compared with Full (all four layers calibrated).

"w/o" means the layer remains in the population but is excluded from calibration. After calibration, each selected configuration is frozen and evaluated by the average of ten independent 60-day held-out runs.

Table R1-Q1: Calibration error and held-out performance of the four-layer ablation.

GroupQ1Q_1Q2Q_2Q3Q_3Q4Q_4Q5Q_5H1H_1H2H_2H3H_3H4H_4H5H_5dcald_{\mathrm{cal}} ↓dHOd_{\mathrm{HO}} ↓
w/o π1\pi_1-0.7712-0.4347-0.9287-0.8004-0.13030.29180.3118-0.33350.0023-0.09532.18143.7439
w/o π2\pi_2-1.3560-0.1389-0.78960.4814-0.13030.28110.1366-0.34000.0667-0.12183.69958.2171
w/o π3\pi_3-1.0938-0.44040.0386-0.9079-0.13030.27910.1308-0.1315-0.2908-0.29722.78965.5475
w/o θ\theta-0.8313-0.4337-0.8975-1.2422-0.13030.24240.2043-0.2098-0.4019-0.22382.91915.6391
Full-0.6555-0.5162-0.1250-0.6814-0.1303-0.0930-0.07560.06160.0564-0.03901.84743.1461

From the same initial state, disabling any single layer increases both calibration and held-out loss, supporting a contribution from each layer in this setting. The roles of π3\pi_3 and θ\theta, which you asked about specifically, are visible in concrete market characteristics: without π3\pi_3 calibration, the deviations in volatility, the absolute-return–volume relation, volatility persistence, spread, and drawdown widen; without θ\theta calibration, the deviations in absolute-return autocorrelation, the absolute-return–volume relation, and the five dynamics metrics widen. These results show how strategy composition and behavioral parameters affect the market characteristics, with all four layers together achieving a better overall match.

AuthorsOpenReview ↗

Q2: Common budget, evaluation schedule, and interpretation

The original comparison already used a unified budget, and in this round of re-runs we placed the newly added π1\pi_1-only under the same control: Direct-X, Joint-Z, π1\pi_1-only, and Full share the same candidate-batch schedule, simulation quota, and total budget cap.

Concretely, every method is budgeted for two search starting points and two rounds, with two scoring paths per candidate. The regular search comprises 158 waves and 1,664 candidate slots (3,328 scoring paths); adding one initialization wave (17 slots) and six waves reserved for the conditional structure check (at most 146 slots), the full calibration plan comprises 165 waves, at most 1,827 candidate slots, and at most 3,654 scoring paths. The per-stage allocation is given in Schedule Table B of the general statement (Comment 3); held-out evaluation is counted separately.

These planned quotas apply uniformly to all compared methods. Actual launched-path counts depend on candidate deduplication, historical-score reuse, and conditional branches, as recorded in the run logs.

All methods share the same market setup, calibration window, and scoring protocol, and this round's fixes to resource scale, strategies, and implementation apply to every group. Direct-X searches the full parameters directly, Joint-Z jointly searches the structured coordinates, π1\pi_1-only calibrates one layer, and Full uses PHI to organize the four-layer search. Direct-X's slow progress in a high-dimensional space illustrates the practical difficulty of calibration with limited computation; the comparison shows what each scheme achieves under the same budget constraints. Joint-Z and Full share the structured representation, so their comparison examines PHI's search organization under a common planned budget.

AuthorsOpenReview ↗

Q3: Cross-period, asset coverage, and longer follow-up evaluation

We report two historical windows, independent seeds, asset subsets, and market rules.

(1) Two historical windows. Each window uses 10 trading days of calibration and 60 of held-out evaluation (open trading days only): January—calibration Jan 2–15, held-out Jan 16–Apr 18; July—calibration Jul 1–12, held-out Jul 15–Oct 15.

Table R1-Q3a: Results for the two historical windows.

WindowQ1Q_1Q2Q_2Q3Q_3Q4Q_4Q5Q_5H1H_1H2H_2H3H_3H4H_4H5H_5dcald_{\mathrm{cal}} ↓dHOd_{\mathrm{HO}} ↓
Starting Jan 2024-0.6555-0.5162-0.1250-0.6814-0.1303-0.0930-0.07560.06160.0564-0.03901.84743.1461
Starting Jul 2024-0.5621-0.4258-0.1877-0.7014-0.13030.1091-0.03570.1182-0.1758-0.38751.76113.7535

Both windows yield comparable errors of similar magnitude.

(2) Run variability. We froze the Full configuration selected in the January window and ran 60-day held-out evaluation with ten fresh seeds unused in calibration.

Table R1-Q3b: Ten independent held-out runs of Full.

RunQ1Q_1Q2Q_2Q3Q_3Q4Q_4Q5Q_5H1H_1H2H_2H3H_3H4H_4H5H_5dHOd_{\mathrm{HO}} ↓
1-0.7180-0.51270.4409-0.5546-0.1303-0.5713-0.4666-0.0293-0.4030-0.30785.0827
2-0.6819-0.51950.2798-0.7721-0.13030.3560-0.1907-0.29590.21800.21084.9028
3-0.7207-0.4936-0.4803-1.0212-0.13030.4010-0.44740.23480.2754-0.49343.6900
4-0.6364-0.53750.0487-0.6224-0.1303-0.4846-0.5489-0.56360.79340.89282.9319
5-0.7538-0.5075-0.6247-0.4629-0.1303-0.5549-0.04790.7758-0.2265-0.66332.2908
6-0.6667-0.51570.1008-0.4333-0.1303-0.4711-0.83020.5960-0.6321-0.65642.5802
7-0.6324-0.5146-0.1350-0.7130-0.13030.01730.0445-0.27800.40090.14712.5961
8-0.4773-0.5099-0.2164-0.8553-0.13030.20040.7649-0.16980.32500.34702.8211
9-0.6703-0.5352-0.2500-0.6684-0.1303-0.13710.17020.45200.11000.08092.3540
10-0.5971-0.5163-0.4139-0.7113-0.13030.31460.7963-0.1064-0.29740.05252.2117
Mean-0.6555-0.5162-0.1250-0.6814-0.1303-0.0930-0.07560.06160.0564-0.03903.1461
Min-0.7538-0.5375-0.6247-1.0212-0.1303-0.5713-0.8302-0.5636-0.6321-0.66332.2117
Max-0.4773-0.49360.4409-0.4333-0.13030.40100.79630.77580.79340.89285.0827

Across ten runs, the five Facts mostly lie within one reference standard deviation. The dynamics metrics vary more across paths, sometimes changing sign, but all stay within one reference standard deviation despite exclusion from calibration. Held-out loss ranges from 2.2117 to 5.0827, with the first two runs relatively high. Min/Max are per-column statistics.

(3) Asset coverage. The five stocks—Sinopec, Kweichow Moutai, Ping An (Shanghai); Vanke A, ZTE (Shenzhen)—span energy, beverages (baijiu), insurance and diversified financials, real estate, and telecom equipment. Historical tradable A-share market-cap estimates used for initialization are about 10¹¹ CNY for Vanke A and ZTE, several times 10¹¹ for Ping An and Sinopec, and over 2×10¹² for Kweichow Moutai (roughly a 20-fold range). Additional-stock experiments have some promising preliminary results, but full evaluation cannot be completed within the response period. We therefore report the existing three Shanghai and two Shenzhen stocks separately below.

Table R1-Q3c: Held-out results by exchange subset.

SubsetQ1Q_1Q2Q_2Q3Q_3Q4Q_4Q5Q_5H1H_1H2H_2H3H_3H4H_4H5H_5dHOd_{\mathrm{HO}} ↓
Shanghai (3)-0.6372-0.5081-0.1271-0.6669-0.1303-0.0836-0.07210.06240.43250.65003.2431
Shenzhen (2)-0.6828-0.5284-0.1218-0.7033-0.1303-0.1071-0.08090.0603-0.5078-1.07243.0007

The subsets have similar overall errors; spread and drawdown deviate in opposite directions.

(4) Market rules. The system's configurable rule layer supports T+N position-unlocking and sellability restrictions and adjustable price-limit widths. The original paper's appendix describes the rule layer and current configuration. Preliminary T+0 and no-price-limit runs have shown promising results. U.S.-equity experiments are under consideration, but formal validation has not been completed within the response period. The revision will explain the rule configuration and completed validations, and acknowledge the limited evaluation of new asset pools, other markets, and additional market regimes.

Comments on the paper

8 comments
AuthorsOpenReview ↗

General statement

We thank all reviewers for their careful reading and constructive comments. The reviews center on three issues: the practical benefit of full-population calibration, the sufficiency and robustness of the experimental validation, and the fairness of the method comparison. We had begun improving the experimental setup, the framework implementation, and the calibration procedure before receiving the reviews, and over the past month we have systematically revised them and rerun nearly all major experiments. The updated results provide a stronger basis for addressing these concerns.

Because these changes affect the shared foundation and the main results of every experiment, we first describe the revisions and the updated main results (Comments 2–5 below), and then respond to each reviewer's points individually. As the experiments are time-consuming, we report the completed work and results here; the paper revision is still in progress, and an updated PDF will follow.

AuthorsOpenReview ↗

General statement-2: Revisions to the experimental setup and implementation

To clarify how this round of results differs from the original submission, we first describe the changes shared by all experiments.

Evaluation horizon. The main experiments were extended from 5 trading days of calibration + 3 days of held-out evaluation to 10 + 60 trading days. In the main window, calibration covers Jan 2–15, 2024 and the held-out period covers Jan 16–Apr 18, 2024, i.e., the 10 and 60 trading days above. The calibration period is used to search for and select configurations; the selected configuration is then frozen for the longer held-out evaluation, whose results are never fed back into the search or the candidate selection.

Initial asset scale. We raised the total initial assets of the simulated market from 10⁸ to 10¹⁰ CNY. The selected stocks all have market capitalizations of at least 10¹⁰ CNY; the original scale was too small and produced insufficient liquidity in our experiments. The larger pool eases this artificial constraint on trading activity.

Asset-budget initialization. Approximate market-capitalization proportions now serve as the initialization prior for the asset budget π2\pi_2, giving the search a more reasonable starting point and reducing the cost of early asset-allocation adjustments; calibration still refines π2\pi_2 afterward.

Strategy fixes. We corrected strategy execution so that order behavior is tied to the capital controlled by each strategy. The original implementation did not fully reflect this link, which weakened the transmission of changes in strategy composition and behavioral parameters into actual trading.

Bug fix. We fixed a spurious check that incorrectly blocked orders, so legitimate trading intentions now reach the order book.

Since these changes alter the shared foundation of all experiments, we reran nearly all major experiments under the revised setup. The updated results below re-examine the benefit of full-population calibration and the reliability of the findings.

AuthorsOpenReview ↗

General statement-3: Revisions to the calibration procedure

For the calibration procedure, we improved the coordination among per-layer updates and the reuse of prior search progress.

Procedure organization. The old schedule had eight specialized or joint stages per round, visited in different orders in the two rounds. The new schedule adds a leading Cross at the start of each round and inserts six joint refinements between the major specialized updates, giving 15 regular search nodes plus one conditional structure node. Both rounds now use the same order, with the conditional structure check between them, and the π1\pi_1 update receives more consecutive feedback waves.

Joint search. The old schedule used Cross only at the end of each round. The new one also performs cross-layer exploration at round start and interleaves joint refinements between specialized updates: after one update changes a layer, the next continues from an already cross-layer-coordinated configuration rather than waiting for the end-of-round adjustment.

Cross-stage inheritance. Each stage inherits the complete candidates found so far and reuses eligible historical scores, successful directions, and search-radius statistics, retaining the current configuration when no improvement is found. Scores are reused only when the evaluation conditions match and the stage reproduces the same full parameter vector; compatible direction and radius statistics preserve the local search structure. The next stage uses this information in its reconstructed optimizer, reducing information loss and repeated exploration at stage transitions.

Both tables give planned budgets under two starting points (Starts), two rounds, and two scoring paths per candidate. The two Starts advance in the same logical wave; each batch of candidate evaluations is completed before the next search feedback step.

Schedule Table A: Original eight-stage schedule and budget.

StageR1 pos.R2 pos.WavesPaths/wave (2 Starts)Planned paths
π1\pi_1 endowment allocation11732, then 16×6256
π2\pi_2 local asset allocation27732, then 16×6256
π3\pi_3 strategy-weight core33732, then 16×6256
θ\theta behavioral-parameter core42732, then 16×6256
π2\pi_2 global asset allocation54732, then 16×6256
π3\pi_3 trader loadings66732, then 16×6256
θ\theta trader loadings75732, then 16×6256
Cross-layer joint search (Cross)88732, then 16×6256
Regular search subtotal8/round—7 per visit; 112 total—2,048
Initializationbefore R1—one pass—34
Conditional structure checkafter R1before R2between rounds, conditional—≤388
Totalinit + 2 rounds + check—112 regular waves + init + check—≤2,470

Schedule Table B: New schedule and budget.

Pos.Node and roleWavesPaths/wave (2 Starts)Paths (2 rounds)
1Leading Cross: cross-layer exploration532, then 16×4192
2π1\pi_1 endowment allocation964, then 32×8640
3Joint refinement: to local asset allocation532, then 16×4192
4π2\pi_2 local asset allocation532, then 16×4192
5Joint refinement: to strategy-weight core532, then 16×4192
6π3\pi_3 strategy-weight core532, then 16×4192
7Joint refinement: to behavioral-parameter core532, then 16×4192
8θ\theta behavioral-parameter core532, then 16×4192
9Joint refinement: to global asset allocation532, then 16×4192
10π2\pi_2 global asset allocation532, then 16×4192
11Joint refinement: to strategy-weight loadings532, then 16×4192
12π3\pi_3 trader loadings532, then 16×4192
13Joint refinement: to behavioral-parameter loadings532, then 16×4192
14θ\theta trader loadings532, then 16×4192
15Closing Cross: cross-layer correction532, then 16×4192
Regular search subtotal15 nodes/round79/round, 158 total—3,328
16 (conditional)Keep/Grow/Prune check between rounds6 total≤4 anchors; branch search ≤96, then ≤48×4≤292
Initializationshared baseline and initial candidates13434
Totalinit + regular search + check165 waves—≤3,654

The new regular search has 1,664 candidate slots and the full calibration at most 1,827 (3,328 and ≤3,654 paths); the extra budget goes to more frequent joint search and a stronger π1\pi_1 update. The tables report planned quotas; actual new simulations depend on cache reuse, candidate deduplication, and conditional branches. Held-out evaluation is counted separately.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.