ForesightFlow: An Information Leakage Score Framework for Prediction Markets

ForesightFlow: An Information-Theoretic Detection Framework for Informed Trading in Decentralized Prediction Markets

Motivation and Context

Decentralized prediction markets, exemplified by platforms like Polymarket, transform diverse participant beliefs into dynamic price signals that have significant reference and coordination value for media, institutions, and the public. The transparency and pseudonymity inherent to these on-chain markets, however, create unique opportunities for market participants with material non-public information (MNPI) to extract anomalous profits, as evidenced by multi-million-dollar events around geopolitical crises, corporate disclosures, and regulatory decisions. Yet, the bulk of extant research focuses on identification of informed activity only after market resolution, a paradigm that fails to provide actionable detection during periods of live, consequential price movement.

Information Leakage Score: Formalization and Interpretation

The central methodological innovation of this framework is the Information Leakage Score (ILS), an axiomatic, information-theoretic label generator for resolved binary outcome markets. For a resolved market $M$, ILS quantifies the proportion of the market's terminal information move that was realized before the associated public news event, with the formal definition: $$ILS(M) = \frac{p(T_{\text{news}}) - p(T_{\text{open}})}{p(T_{\text{res}}) - p(T_{\text{open}})}$$ where $p(\cdot)$ denotes the mid-quote price of the YES outcome at the specified timepoints: market opening, first public mention of event-relevant information, and resolution. The score is only interpretable under explicit scope conditions: nontrivial total information movement ($\geq 0.05$), opening prices away from degenerate consensus, and robustness to anchor timing for the news event.

Crucially, the ILS is interpreted through the lens of the Murphy decomposition of the Brier score: high ILS values correspond to resolution (discriminative power) that is front-loaded in the pre-news window—indicative of price dynamics not justified by the public information set. This connects the microstructure empirics to the proper-scoring-rule literature and enables theoretically grounded, outcome-label definition without post-hoc outcome leakage.

Scope Delineation and Empirical Constraints

ForesightFlow codifies a detailed resolution typology, distinguishing between event-resolved contracts (resolution at a contemporaneous, observable event), deadline-resolved contracts ("will event X occur by deadline Y?"), and structurally unclassifiable markets. Systematic empirical analysis reveals that the most salient, documented cases of Polymarket informed trading (e.g., Iran strike, Venezuela events, major regulatory outcomes) are all deadline-resolved, not event-resolved. This mismatch required an extension of the methodology: the deadline-ILS variant, in which leakage is computed versus a baseline belief trajectory (often, a constant-hazard Bayesian prior) with the label anchored to public observation of the event itself. This closes the gap between method and validated ground truth.

Significantly, the framework identifies and corrects common sources of noise and spurious signal: false positives from edge-effect regimes (opening prices near resolution), trivial-movement markets, and anchor sensitivity in timestamp recovery. Only contracts satisfying all scope and typology criteria are retained for inference.

Data and Label Generation Pipeline

The architecture integrates heterogeneous data sources: Polymarket's CLOB and subgraph APIs (for historical trade, volume, and wallet identification), the UMA Optimistic Oracle (for evidence URLs and authoritative event timestamps), GDELT (for cross-verified news event anchoring), and on-chain wallet metadata from Polygon. The resolution evidence hierarchy is designed to recover "true" news event timestamps, with LLM-assisted matching reserved for ambiguous cases.

Empirical validation is performed both on an extensive market backfill (over 900k contracts, with category-specific filtering) and on a curated inventory of publicly documented insider trading episodes, the ForesightFlow Insider Cases (FFIC), spanning asset classes, event types, and Polymarket functional regimes.

Machine-Detectable Microstructure and On-Chain Wallet Signatures

The ForesightFlow detector incorporates a composite of real-time, causal features:

• Classical market microstructure diagnostics, adapted to the discrete-outcome, bounded-price regime:
  • Order imbalance at multiple time resolutions.
  • VPIN-style toxicity (bucketed volume imbalance).
  • Rolling estimation of Kyle’s lambda (price impact slope).
  • Variance ratio and two-sidedness (borrowed from the Signal Credibility Index) to distinguish durable, directional repricing from ephemeral or contested moves.
  • Trade-size kurtosis and Hawkes process-based self-excitation for time-clustering anomalies.
• Wallet-level features exploiting pseudonymous on-chain identity:
  • Wallet-novelty score, aggregating wallet age, prior market participation, concentration of funding sources, and entry timing.
  • Cross-market wallet analysis for abnormal patterns of coordinated positioning.

The resulting feature vector accommodates both immediate microstructure distortion and identity-dependent forensics, supporting robust, interpretable detection without reliance on post-hoc profitability signals.

Empirical Findings and Methodological Implications

An initial pilot empirical study, exploiting the ILS on event-resolved markets using proxy news timestamps, demonstrated that ILS alone does not separate informed flow from null hypotheses—primarily due to misalignment between proxy and true event timing, and spurious signal from edge-effect markets. A high-resolution, article-anchored case study (Epstein-files market) revealed that even robust ILS signals can be consistent with non-insider price dynamics, underscoring the necessity of joint evaluation with wallet-level features.

The FFIC audit established that meaningful empirical power for detecting informed trading flows is only achievable for deadline-resolved markets, explicitly motivating the deadline-ILS formalism. The empirical companion paper (cited as [nechepurenko2026foresightflow_empirical]) details per-category hazard model estimation, case-wise ILS^dl application, and cross-case consistency checks.

Limitations and Practical Impact

The framework is candid regarding both theoretical and operational limitations:

• The proper-scoring-rule connection holds only under regularity assumptions on market calibration.
• Label assignment is sensitive to news-timestamp recovery quality; some high-volume Polymarket markets remain outside indexable scope due to subgraph infrastructure limits.
• The small number of publicly documented, unambiguously labeled insider cases restricts fine-grained detector performance claims, but longitudinal production monitoring accumulates power over time.
• Current analysis is Polymarket-specific; cross-platform generalization (e.g., Kalshi) demands further harmonization and is not undertaken in this stage.

Theoretical and Practical Implications for Market Surveillance and Market Microstructure

ForesightFlow advances both the measurement and operational detection of informed trading in unprecedented ways for decentralized, pseudonymous, and globally accessible markets. The ILS bridges market microstructure and proper scoring rule approaches—enabling the quantification of anticipatory information incorporation with explicit time anchoring and category adaptation. The fusion of on-chain wallet analytics and microstructure features sets a new empirical foundation for market surveillance, regulatory assessment, and future market design.

Practically, the system provides actionable, real-time probability outputs for active markets, facilitating both transparency for the trading public and integrity monitoring for operators and regulators. Release of the entire codebase, typology, case inventories, and empirical artifacts as open infrastructure supports community-wide reproducibility and operationalization.

Conclusion

The ForesightFlow framework establishes a rigorous, scope- and category-aware methodology for detecting informed flow in decentralized prediction markets. By integrating an information-theoretic label formalized via the Murphy decomposition, category-conditional typologies, microstructure and wallet features, and multi-source timestamp anchoring, ForesightFlow pushes the frontier of real-time market forensics. The framework’s methodological discipline—mandating causal features, robust labeling, and scope-aware interpretation—ensures that future empirical and practical enhancements will rest on replicable, interpretable, and actionable foundations for the trustworthy operation of prediction markets as public information infrastructure.

All system artifacts, code, and case inventories are publicly available at https://github.com/ForesightFlow and https://foresightflow.xyz.