💰 quantitative finance

Information Leakage at Population Scale: An Evaluation of the Polymarket Insider-Relevant Subpopulation, 2020-2026

This paper evaluates the Information Leakage Score framework across 12,708 Polymarket markets from 2020 to 2026, revealing that its effective application is severely limited by resolution ambiguity and low anchor-sensitivity, thereby demonstrating that detecting informed flow requires methodological refinements in resolution typology and baseline correction rather than just score computation.

Original authors: Maksym Nechepurenko

Published 2026-05-04

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Maksym Nechepurenko

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine a giant, global betting hall called Polymarket. People place bets on everything from "Will the President sign this bill?" to "Will this movie win an Oscar?" The paper you're reading is an investigation into whether some people are cheating by betting on secrets they know before the rest of the public does.

The author, Maksym Nechepurenko, built a special "lie detector" tool called the Information Leakage Score (ILS). In a previous test, this tool worked perfectly on just one specific bet. This paper asks: Does this tool work when we try to use it on thousands of bets at once?

Here is the story of what they found, explained simply.

1. The Tool is Too Picky (The "Filter" Problem)

The author tried to run this tool on 12,708 different betting markets. The goal was to find the "cheaters" (people betting on inside information).

The Result: The tool only worked on 88 of those markets (less than 1%).
The Analogy: Imagine you have a metal detector that is supposed to find gold coins in a massive field. You sweep the whole field, but the machine only beeps for 88 spots. Why? Because the machine is designed to only beep for coins that are perfectly round and shiny.
The Real Issue: Most of the betting questions were too messy. For example, a bet might ask, "Will the leader be removed?" But "removed" could mean fired, resigning, or dying. The tool needs a single, clear moment in time (like a specific date a law is signed) to work. If the question is vague, the tool gives up.
The Surprise: Even when they looked at 32 famous cases where cheating was already known to have happened, the tool only worked on one of them. The other 31 were too vague for the tool to understand.

2. The "Event" is Hard to Pin Down

To catch a cheater, the tool needs to know exactly when the news broke.

The Problem: The tool tried to use AI to find the exact second a news story broke. It got the right date only about 58% of the time.
The Analogy: It's like trying to catch a thief by knowing exactly which second they walked through the front door. If you think they walked in at 2:00 PM, but they actually walked in at 2:05 PM, your camera misses them.
The Consequence: Because the tool is so sensitive to the exact time, it only works on the "cleanest" bets where the news happens at a single, clear moment (like a sports score). On messy, political, or legal bets, the tool gets confused.

3. The "Negative" Scores Were Just Math, Not Cheating

When the tool finally did work, it gave mostly negative scores.

The Misunderstanding: At first, the author thought a negative score meant people were betting against the truth (a weird kind of cheating).
The Correction: The author realized this was just a trick of the math.
The Analogy: Imagine a countdown timer for a rocket launch. As time passes and the rocket doesn't launch yet, the chance of it launching right now naturally goes down. If the rocket finally launches at the very last second, the price on the betting market looks like it crashed before rising. The tool saw this "crash" and thought it was cheating.
The Fix: The author added a "baseline correction" (like subtracting the natural countdown effect). After this fix, some of the "cheating" disappeared, but on one specific type of bet (regulatory announcements), the negative scores remained. This suggests that on those specific bets, people might actually be betting against the outcome before the news breaks.

4. The "One-Size-Fits-All" Assumption Was Wrong

The tool assumed that all events happen at a steady, predictable rate (like rain falling at a constant speed).

The Finding: The data showed that events don't happen at a steady rate. Sometimes they happen in bursts; sometimes they take a long time.
The Analogy: The tool was using a "steady rain" model, but the real world was more like a "storm" followed by a "drought." The author had to switch the tool to a more flexible model (called Weibull) to handle the storms and droughts correctly.

The Big Picture Conclusion

The paper concludes that the "lie detector" tool is real and working, but it is currently too narrow to be used on the whole betting market.

It works well on clear, discrete events (like "Will Team X win the game?").
It fails on vague, complex events (like "Will the government change its policy?").

The author says: "We found the tool works, but we need to upgrade it to handle vague questions and fix the math on how time passes." They plan to do this in a future paper (Paper 3b) by using smarter AI to understand messy questions and combining this tool with other methods that look at individual bettors' wallets.

In short: The tool is a high-precision sniper rifle. It works perfectly if you have a clear target. But the betting market is full of foggy, moving targets, so the rifle misses most of them. The author is now trying to build a scope that works in the fog.

1. Problem Statement

The paper addresses the challenge of detecting informed trading (insider trading) on decentralized prediction markets, specifically Polymarket. While prior work (e.g., Mitts & Ofir [2026], Gómez-Cram et al. [2026]) has successfully identified informed traders at the wallet or account level using aggregate trading patterns, there is a need for a market-level, real-time diagnostic that can identify information leakage before an event is resolved.

The author seeks to scale the Information Leakage Score (ILS) and its deadline-resolved extension (ILSdl), previously tested on single cases, to a population level. The core problem is determining whether the framework can effectively detect informed flow across thousands of markets, or if structural limitations (such as ambiguous event definitions) render the method inapplicable to the majority of "insider-relevant" markets.

2. Methodology

The study employs a rigorous, multi-stage pipeline to evaluate 12,708 candidate markets from the Polymarket corpus (Oct 2020 – April 2026) within three insider-relevant categories: Military/Geopolitics, Regulatory Decisions, and Corporate Disclosures.

A. The ILSdl Framework

The Information Leakage Score (ILSdl) quantifies the share of a market's total price movement that occurred before the public observation of the underlying event ( $T_{event}$ ).
$ILSdl(M) = \frac{p(T_{event}^-) - p(T_{open})}{p(T_{resolve}) - p(T_{open})}$
Where:

$p(T_{event}^-)$ is the price one minute before the event.
$p(T_{open})$ is the opening price.
$p(T_{resolve})$ is the final resolution price (1 for YES, 0 for NO).

B. Scope Conditions & Filters

To compute a valid score, markets must pass four sequential filters:

Resolution Typology: Must be "deadline-resolved" (event occurs by date $D$ ) and not "unclassifiable."
Outcome: Must resolve YES (NO outcomes lack a recoverable $T_{event}$ ).
$T_{event}$ Recovery: The underlying event timestamp must be recovered via an LLM-assisted multi-source verification pipeline with confidence $\ge 0.7$ .
Scope Conditions: Positive lead time, edge-effect constraints (opening price not too extreme), and Anchor Sensitivity (score stability across short-window anchor offsets).

C. Hazard-Decay Baseline Correction

A critical methodological innovation is the introduction of a hazard-decay adjusted score ( $ILSdl_{adj}$ ).

Problem: On deadline contracts, prices mechanically decay toward zero as time passes without the event occurring (rational Bayesian updating). This creates a negative bias in raw ILSdl even without insider trading.
Solution: The authors subtract the expected price under a survival-probability baseline (derived from a Weibull hazard model) from the observed price.
$ILSdl_{adj}(M) = \frac{p(T_{event}^-) - E[p(T_{event}^-) | T_{open}, D, \hat{\lambda}]}{p(T_{resolve}) - p(T_{open})}$

D. Validation

Ground Truth: The ForesightFlow Insider Cases (FFIC) inventory (32 documented insider trading episodes) serves as a labeled validation set.
Second-Pass Validation: An independent LLM run on a 50-market sample validates the accuracy of $T_{event}$ recovery.

3. Key Contributions & Findings

Finding 1: Structural Exclusion (The "Unclassifiable" Bucket)

Result: Only 0.7% (88 of 12,708) of candidate markets yield computable ILSdl values.
Cause: The primary exclusion (43.8% of the FFIC inventory) is resolution-criterion ambiguity. Markets with predicates like "win the election," "out as Supreme Leader," or "in custody" lack a single, discrete public timestamp ( $T_{event}$ ).
Implication: The framework's effective domain is structurally narrower than initially framed. The "unclassifiable" bucket contains genuine ambiguity, not just classifier noise.

Finding 2: Anchor Sensitivity & Sample Narrowing

Result: Of the 88 computable markets, only 12 (13.6%) satisfy the anchor-sensitivity condition (stable scores across short time windows).
Validation: Independent $T_{event}$ recovery achieved only 57.8% exact-date agreement overall (vs. a $\ge 90\%$ target). Agreement was high for discrete events (Military/Corp: 78.9%) but low for multi-stage processes (Regulatory Formal: 35.7%).
Implication: The operationally meaningful sample is restricted to the 12 anchor-robust markets. The majority of computed scores are likely noise due to ambiguous event timing.

Finding 3: Negative Bias and Hazard Correction

Result: Raw ILSdl medians are negative across all sub-buckets (ranging from -0.88 to -0.16).
Interpretation: Initially, this suggests markets price against the outcome. However, after hazard-decay adjustment:
- Regulatory Formal (Post-2024): Shifts from -0.21 to -0.02 (essentially zero). The negativity was purely rational decay.
- Regulatory Announcement (Post-2024): Remains strongly negative (-0.80 $\to$ -0.84). This suggests genuine "anti-information" flow or selection bias in scheduled announcements.
- MilGeo/Corp: Partially corrected but retains some negative signal.
Implication: Raw negative scores are often artifacts of deadline mechanics; hazard-adjusted scores are required for valid interpretation.

Finding 4: Hazard Rate Specification

Result: The constant-hazard exponential model (used in prior single-case work) is rejected for the pooled post-2024 population in favor of a Weibull distribution.
Nuance: When analyzed per subcategory, the exponential model often fits better. The Weibull preference in the pooled sample is driven by category mixture (combining short-deadline announcements with long-deadline formal processes) rather than true within-category duration dependence.
Implication: Future models must either use Weibull for pooled data or re-estimate hazards per subcategory.

4. Significance & Limitations

Significance

Scope Discovery: The paper demonstrates that scaling a detection framework reveals structural limitations (semantic ambiguity) that single-case studies miss.
Methodological Refinement: It establishes that hazard-decay correction is essential for deadline contracts to distinguish between rational price decay and informed trading.
Complementary Detection: The market-level lens (ILSdl) offers a real-time, in-flight diagnostic for a narrow subset of markets, complementing wallet-level (Mitts & Ofir) and account-level (Gómez-Cram) methods which are post-hoc but cover a broader population.

Limitations & Future Work (Paper 3b)

Coverage: The current framework excludes ~99% of the candidate sample due to typology and anchor constraints.
Ambiguity: The "unclassifiable" bucket requires LLM-assisted re-classification and a multi-anchor extension (handling multiple candidate $T_{event}$ values) to include ambiguous-predicate markets.
Data: The study relies on a single pipeline run; future work will validate on independent snapshots and integrate wallet-level features (PIN, VPIN, etc.) to train a gradient-boosted detector.

Conclusion

The paper concludes that while the ILSdl framework is operational, its population-scale application is currently limited to a small, highly specific subset of markets (approx. 12 out of 12,708) where events are discrete, timestamps are recoverable, and anchors are robust. The primary obstacle to broader detection is not score computation, but resolution semantics. The authors propose a roadmap involving LLM re-classification and multi-anchor extensions to expand the framework's coverage in future work.