PaperlayerView exact revision on arXiv

Source-linked AI summary

What Prediction-Market Contracts Reveal at Launch: Trading Volume and Dispute Risk

Ahmed Mahrous, Roberto Di Pietro

arXiv:2610.29578v1q-fin.STcs.LG

TL;DR

The paper asks whether contract information available at listing predicts later trading volume and dispute risk. It evaluates launch-time features across resolved Polymarket markets, finding strong but regime-dependent volume prediction and dispute scores useful for ranking groups rather than predicting individual cases.

  • Problem

    Most prediction-market research studies trading after it begins, leaving limited evidence on how contract information available at listing relates to later volume and dispute risk.

    Source: §1 · p. 1

  • Method

    The study predicts lifetime volume and formal dispute status from launch-time contract text, resolution rules, and metadata, using event-grouped evaluation and temporal tests.

    Source: §1 · p. 1, §3.3 · p. 4

  • Results

    Out-of-sample R^2 = 0.747 for log-transformed volume under same-period evaluation, but AUC for reaching $1,000 falls from 0.940 to 0.577 across time; dispute prediction has AUC = 0.78 and average precision = 0.012.

    Source: §1 · pp. 1–2

  • Takeaways & Limitations

    Volume models require refitting and monitoring within a stable regime, while dispute models belong in human review rather than automated accept/reject decisions.

    Source: §8 · p. 7

    Models trained on earlier contracts perform poorly on previously unseen question forms, a structural-novelty limitation of temporal deployment.

    Source: §8 · p. 8

Abstract

from arXiv · show

Using a near-census of 802,863 resolved markets listed on Polymarket between January 2023 and May 2026, and only information available at listing (the market's title and description, named resolution source, and contract metadata), we predict lifetime trading volume and whether settlement is later disputed. Contract information predicts trading volume well (out-of-sample $R^2 = 0.747$ for log volume), and most of the signal is interpretable: market topic provides the largest gain ($+0.289$ $R^2$ over a $0.220$ contract-feature baseline), wording and resolution-rule design add smaller but measurable improvements ($+0.027$ combined), and neural text models add little beyond the interpretable features. This performance does not survive a period of platform change coinciding with a shift in listing formats: accuracy in predicting whether a market reaches \$1,000 falls from AUC $0.940$ under same-period evaluation to $0.577$ when training on past quarters and testing on later ones. The loss concentrates in markets whose question forms were absent from training (76.9% of test markets; on familiar forms the model still reaches $0.778$), and a decomposition attributes more of it to a changed relationship between contract characteristics and demand than to an outdated mapping from text to features. Dispute predictors are more stable but less precise: launch-time features rank dispute risk far better than chance (AUC $0.78$), yet at a 0.31% base rate precision is too low for case-by-case prediction (average precision $0.012$), so the scores suit ranking groups of markets rather than forecasting individual disputes.

1 Introduction

The paper asks what launch-time contract information reveals about later trading volume and dispute risk, then evaluates prediction under same-period and temporal shifts. Topic carries the largest volume signal, while temporal change weakens volume prediction and dispute scores are better suited to ranking markets than forecasting individual disputes.

Motivation

  • The study asks what a contract reveals at listing, when text and terms are the market-specific information available before order flow exists.Trading volume is highly concentrated, and disputes are rare but carry a disproportionate share of volume.Source: §1 · p. 1

Study design

  • Predictors use launch-time contract information, and the design groups related contracts by event while matching dispute cases to controls from the same topic, quarter, and source type.The analysis complements prediction with a matched case-control comparison.Source: §1 · p. 1

Contract traction predictability

  • Out-of-sample R^2 = 0.747 for log-transformed volume, and AUC = 0.940 for whether lifetime volume reaches $1,000 under same-period cross-validation.Lexical topics add 0.289 R^2 over a 0.220 contract-feature baseline; Figure 1 shows topics separating high- from low-traction segments.Source: §1 · p. 1, Figure 1 · p. 2

Dispute predictability

  • At a 0.31% dispute base rate, AUC = 0.78 but average precision is 0.012, making dispute scores suitable for grouping markets by risk rather than individual prediction.Adjusted analyses associate disputes with reading difficulty, unidentifiable or unparsable sources, and edge-case drafting patterns; these are not causal estimates.Source: §1 · p. 2

Temporal change and regime dependence

  • AUC falls from 0.940 to 0.577 when training on past quarters and testing on later ones, with 76.9% of test markets using question forms absent from training.Familiar forms retain AUC 0.778; re-estimating the feature–volume relationship recovers more loss than relearning text representations.Source: §1 · p. 2

2 Related Work

Prior prediction-market research mainly examines price accuracy, liquidity, trading, and settlement after markets begin. This paper addresses the gap in evidence on how information available at listing relates to subsequent market outcomes.

Prediction markets and market quality

  • Research on prediction-market quality focuses on whether trading aggregates information into forecasts and on the roles of securities, trading mechanisms, liquidity, and participation.Recent studies examine price discovery and price behavior as participation expands, rather than launch-time predictors of later outcomes.Source: §2 · p. 2

Contract language and financial NLP

  • Financial NLP and crowdfunding studies find that pre-outcome language can predict economic behavior, while prediction-market contract wording also defines the claim and its settlement conditions.Prior contract-design work examines expressible uncertainties, cross-venue ambiguity, and manipulation risks.Source: §2 · p. 2

Temporal generalization and concept drift

  • Dataset-shift research motivates testing whether contract-language models generalize as topics, formats, or trader populations change; this study separates text-representation drift from relationship drift.The paper reports that strong within-period performance generalizes poorly to later markets.Source: §2 · p. 2

Prediction-market data resources

  • The dataset pairs full contract text and design measures with lifetime volume and formal dispute status, complementing resources that observe other parts of the market lifecycle.Prior work reproduces dispute rulings after disputes occur but does not predict beforehand which markets will be disputed.Source: §2 · p. 3

3 Data and Empirical Design

The study addresses dependence, template leakage, platform change, and rare disputes with event-grouped evaluation, template exclusions, temporal cutoffs, and matched case-control comparisons. It analyzes resolved markets using launch-time contract predictors and evaluates volume and dispute outcomes with designs suited to their different data challenges.

3 Data and Empirical Design

  • Four empirical challenges shape the design: dependent related contracts, recurring title templates, rapid platform change, and rare disputes concentrated by subject.The study pairs these with event-grouped evaluation, template exclusion, temporal cutoffs, and matched case-control comparisons, respectively.Source: §3 · p. 3

3.1 Sample and Outcomes

  • The analysis sample contains 802,863 markets in 391,955 events from January 2023 to May 2026; dispute analyses are restricted to the 65.9% with observable oracle logs.Markets are the unit of analysis and events are the clusters for splits and standard errors; 1,634 of 529,408 dispute-observable markets were disputed (0.31%).Source: §3.1 · p. 3, Table 1 · p. 3

3.2 Launch-Time Features

  • Predictors use fields set at listing: structured metadata, topic, wording, and resolution-rule design, excluding realized prices, trader counts, and outcomes.Topics derive from NMF and embedding clusters; wording and rule measures capture features such as text length, specificity, edge-case coverage, readability, and source verifiability.Source: §3.2 · p. 3

3.3 Models and Evaluation

  • A nested model ladder compares basic contract features, topic representations, boosted trees, and neural text benchmarks, with event-grouped folds and performance scored on held-out events.Volume uses out-of-sample R^2 and high-volume classification uses AUC; dispute risk uses matched controls from the same topic, quarter, and source-type stratum.Source: §3.3 · p. 4

4 Predicting Trading Volume

Contract information predicts trading volume well within the same period, with topic contributing more than wording or rule design. Interpretable features capture nearly all of the launch-time signal tested, while feature interactions account for much of the boosted model’s advantage.

Market Topic

  • Adding lexical topics contributes +0.289 R^2 over the basic contract-feature baseline, making topic the largest measured source of predictive gain.Topic alone separates high- from low-traction segments, and its predictive strength holds across topics.Source: §4 · pp. 4–5

Launch-Time Predictability

  • M5 reaches out-of-sample R^2 = 0.747 and AUC = 0.940, compared with the M1 baseline’s R^2 = 0.220 and AUC 0.752.Figure 2 shows the out-of-sample fit as feature blocks are added.Source: Figure 2 · p. 4, §4 · p. 4

Neural Text Benchmarks

  • Neural benchmarks add little beyond interpretable features: end-to-end fine-tuning reaches R^2 0.754 versus 0.747 for M5, a +0.007 gain.Frozen embeddings match M5 to three decimals, and the larger BGE-M3 encoder adds nothing over e5-small.Source: §4 · p. 5

Wording and Rule Design

  • Conditional on topic, wording adds +0.021 R^2 and rule-design measures add +0.006, about one-fiftieth of the topic contribution.Specificity scores of 0 and 3 correspond to 50.5% and 83.0% of markets reaching $1,000, respectively.Source: §4 · p. 5

Feature Interactions

  • Pairwise interactions close 57% of the M4→M5 R^2 gap, versus 25% for depth-one trees and 32% for a linear model with named interactions.Duration’s estimated slope is +2.92 for Bitcoin Up/Down Micro-Markets and −0.03 for Basketball Game Lines and Totals; no launch feature has a universal effect.Source: §4 · p. 5

5 Robustness Checks

Robustness checks preserve the main finding that contract information predicts volume, including after recurring title templates are removed. One within-event comparison nevertheless shows how market role can be mistaken for a wording effect.

Recurring Title Template Exclusion

  • After removing 83.1% of markets with recurring title templates, M5 still reaches R^2 = 0.652 and AUC = 0.891.The feature hierarchy persists, though performance is lower, indicating that repeated question forms contribute to predictability.Source: §5 · p. 5

Additional Robustness

  • Robustness variants usually move the reference specification by less than 0.01; event-level aggregation yields R^2 = 0.674 and AUC 0.956.The headline performance is not carried by one topic family; removing sports raises R^2 to 0.785.Source: §5 · p. 5, Table 3 · p. 6

Within-Event Comparison Caveat

  • Within-event comparisons can confound wording with a market’s role: the vague-language coefficient falls from +6.29 to +0.21 (p = 0.29) among structurally identical siblings.The original association reflected boilerplate concentrated on flagship, highest-volume markets.Source: §5 · p. 6

6 Temporal Generalization

Strong same-period performance does not transfer reliably to later quarters as the platform’s listing formats change. The loss is concentrated in novel question forms, and re-estimating the feature–volume relationship recovers more performance than relearning text representations.

Platform Regime Change

  • AUC falls from 0.940 in same-period evaluation to 0.577 prospectively; 2025 cutoffs yield AUC of 0.52–0.58, near chance.Only the 2026Q1 cutoff, whose training window includes the modern regime, recovers usable performance (R^2 = 0.627, AUC = 0.906).Source: Table 4 · p. 6, §6 · p. 6

Structural Novelty

  • At the 2025Q4 cutoff, 76.9% of test markets have unseen question forms; AUC is 0.487–0.575 on novel forms versus 0.68–0.78 on familiar forms.Novel forms heighten but do not fully explain the failure; they are disproportionately high-volume markets.Source: §6 · p. 6

Representation vs. Relationship Drift

  • Relearning the representation raises R^2 from 0.011 to 0.249, while re-estimating the feature–volume relationship raises it to 0.772; representation adds almost nothing afterward (0.772 vs. 0.773).A uniform prediction shift leaves AUC unchanged because the broken ranking, not just the overall prediction level, is at issue.Source: §6 · p. 6

Practical Response

  • Quarterly retraining yields AUC 0.64–0.71 during the 2025 shift and 0.907 by 2026Q2; the paper recommends tracking novel forms and treating their predictions as low-confidence.Using only the two most recent quarters works about as well as using the entire history.Source: §6 · pp. 6–7

7 Disputes: Prediction and Risk Factors

Launch-time features rank dispute risk, but the 0.31% base rate makes individual predictions imprecise. Matched analyses identify stable associations with readability, source verifiability, and edge-case drafting, rather than causal effects of design.

Dispute Predictability

  • AUC 0.78 ranks disputed markets above undisputed ones, but average precision is 0.012 against a 0.0031 random baseline at a 0.31% base rate.Reviewing the riskiest 5% raises the observed dispute rate to approximately 1.7% while capturing 27.4% of disputes; the authors treat scores as triage, not individual prediction.Source: §7 · p. 7

Matched Risk-Factor Analysis

  • The primary matched design compares each disputed market with up to four controls in the same topic × quarter × source-type stratum, estimating non-source characteristics.Source presence, type, and accessibility are estimated separately; Figure 5 reports matched conditional-logit odds ratios with 95% event-cluster bootstrap intervals for unstarred rows.Source: §7 · p. 7, Figure 5 · p. 7
  • Each additional Flesch–Kincaid grade level is associated with 23% higher dispute odds; no identifiable or parsed source versus a mechanical feed has odds ratio 4.93.Any edge-case clause is associated with 3.8× higher odds, while denser coverage among clauses is associated with lower odds (odds ratio 0.25); these are associations, not causal effects.Source: §7 · p. 7

8 Discussion

Launch-time contract information supports different workflows for volume and dispute risk, but deployment is bounded by regime shifts, the tested neural models, single-platform data, and associational estimates. Volume models require refitting and monitoring, while dispute scores belong in human review.

Implications for deployment

  • Volume models can rank listings only within a stable regime and require refitting and monitoring; dispute models belong in human review rather than automated accept/reject decisions.At a 0.31% dispute base rate, even a well-ranked risk list is mostly false positives.Source: §8 · p. 7

Scope, limitations, and future work

  • The neural comparison covers two open-weight encoders; BGE-M3 adds nothing over the smaller encoder, while e5-small fine-tuning improves R^2 by only 0.007.The conclusion is limited to the tested model class and the textual signal those models recover.Source: §8 · pp. 7–8
  • Because the analysis covers only Polymarket, transfer to venues such as Kalshi remains untested, and listing processes, trader populations, and mechanisms may differ.Cross-platform analysis could test which predictors transfer.Source: §8 · p. 8
  • Estimated relationships between contract design and outcomes are associational, not causal, because creators jointly choose remaining contract characteristics.The within-event audit shows that an apparent wording effect can instead reflect a market’s role or subject.Source: §8 · p. 8
  • Models trained on earlier contracts perform poorly on previously unseen question forms, making structural novelty a deployment limitation.Future work could detect regime changes, report greater uncertainty on unseen forms, and update feature–outcome relationships online.Source: §8 · p. 8

9 Conclusion

Using a near-census of resolved Polymarket markets, the paper finds that launch-time information predicts volume strongly but unevenly, while dispute signals support risk ranking rather than individual forecasts. These findings also identify limits on what launch-time information reveals, including regime-dependent volume prediction and individual dispute outcomes.

9 Conclusion

  • The study uses a near-census of 802,863 resolved markets listed on Polymarket between January 2023 and May 2026.The investigation concerns what a contract reveals at launch, before trading, about later volume and dispute risk.Source: §9 · p. 8
  • Launch-time text predicts volume strongly and interpretably: topic carries most of the signal, wording and rule design add smaller topic-conditional increments, and handcrafted features match the tested neural encoders.For disputes, launch text supports ranking markets by risk, not predicting individual disputes; adjusted analyses identify readability, source verifiability, and edge-case drafting as stable design correlates.Source: §9 · p. 8
  • The findings identify what launch-time information can reveal and what it cannot: which listings attract volume and which designs are associated with disputes, but not individual disputes or volume after a regime shift.The paper also highlights dependence among related contracts within events, repeated templates, temporal regime change, and evaluation of extremely rare disputes.Source: §9 · p. 8
Loading 2610.29578v1…