💰 quantitative finance

Empirical Evaluation of Deadline-Resolved Information Leakage on Documented Polymarket Insider Cases

This paper empirically evaluates the Deadline-Resolved Information Leakage Score (ILS-dl) on Polymarket's documented insider trading cases, specifically within the 2026 U.S.-Iran conflict cluster, demonstrating the metric's ability to distinguish genuine signal from proxy artifacts while validating its exponential-hazard fit for military-geopolitics markets and identifying limitations in wallet traceability.

Original authors: Maksym Nechepurenko

Published 2026-05-05

📖 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 shop called Polymarket. People here place bets on future events, like "Will US troops enter Iran by April 30?" or "Will a specific company release a secret report?"

Sometimes, people win huge amounts of money on these bets. The big question is: Did they win because they were lucky, or because they knew something the rest of the world didn't? (This is called "insider trading" or "informed trading").

This paper is a detective story. The authors built a new mathematical tool (a "scorecard") to try and catch these smart bettors. They tested this tool on a specific group of bets about the 2026 US–Iran conflict.

Here is the breakdown of their investigation in simple terms:

1. The Problem: The "Wrong Clock"

The authors' new tool, called Deadline-ILS, is designed to measure how much the betting price moved before the news actually broke.

The Old Way (The Flawed Clock): Previously, researchers looked at the price of a bet just one hour before the official result was announced.
The New Way (The Real Clock): The authors realized that for many bets, the "real" news happens days or weeks before the official result.
- Analogy: Imagine a bet on "Will it rain tomorrow?" The official result is announced at midnight. But if it starts raining at 2:00 PM, the smart bettors bought their tickets at 2:00 PM, not at 11:59 PM. If you only look at the price at 11:59 PM, you miss the whole story.

2. The Investigation: Testing the Tool

The authors tried to run their new tool on 18 different bets related to the US–Iran conflict. However, they hit a few snags:

Missing Data: For most bets, they couldn't find the exact minute the news broke.
Missing Prices: For many bets, they couldn't get the full history of how the price changed every second (like a stock ticker).
The One Success: Only one bet passed all the tests: "Will US forces enter Iran by April 30?"

3. The Big Discovery: The "Fake" Signal

On that one successful bet, the authors found something shocking.

Using the Old Clock (1 hour before the end): The tool said the price was moving against the winner. It looked like the smart bettors were wrong! The score was -0.33.
Using the New Clock (When the news actually broke): The tool said the price moved with the winner. The score was +0.11.

The Analogy:
Imagine a race.

The Old Method looks at the runners 1 minute before the finish line. By then, everyone is tired and slowing down. It looks like no one was running fast.
The New Method looks at the runners when the starting gun fired. It sees a burst of speed right at the start.

The authors found that the "Old Method" was lying to them. By the time the official result was near, the market had already given up on the idea that the event would happen. The "New Method" caught the small, early hint that the event might happen. This proves that using the actual news time is crucial to finding the truth.

4. The "Wallet" Mystery

The authors also tried to track the digital wallets (the accounts) of the people making these bets to see if the same people were betting on multiple related events.

The Result: They found 332 wallets that bet on both the "US enters Iran" and "Ceasefire" bets.
The Catch: They could only see the bets made after the event happened (during the settlement window).
The Conclusion: These people weren't "insiders" betting before the news; they were just "arbitrageurs" (people who bet on the final result to make a quick profit after the news was already out).
The Limitation: The authors admit they couldn't see the "pre-event" bets because the data system they used doesn't save that history. It's like trying to solve a crime when the security camera only starts recording after the thief has left the building.

5. The Verdict: "Proof of Concept"

The authors are very honest about their findings. They say:

This is a test drive, not a final report. They only successfully tested the tool on one specific bet.
The tool works: It successfully distinguished between a "real signal" and a "fake artifact" caused by looking at the wrong time.
The tool needs upgrades: To use this on a massive scale (like checking millions of bets), they need to fix five things:
1. Make sure the bet is placed before the event starts (not after).
2. Separate different types of news (e.g., "scheduled speeches" vs. "long court cases") because they happen at different speeds.
3. Get more data on corporate news.
4. Crucially: Build a system that records the price of every bet from the very second the market opens, not just when it closes.
5. Crucially: Build a system that records every single trade made by every person from the start, not just the final settlement.

Summary

This paper is a "proof of concept." It says, "We built a new radar to spot insider trading in betting markets. We tried it on one case, and it worked much better than the old radar because it looked at the right time. However, our radar is currently blind to the most important part of the story (the early trades) because our data recording system isn't fast enough yet. We need to build better data recorders before we can catch the big fish."

1. Problem Statement

Decentralized prediction markets (e.g., Polymarket) have accumulated significant evidence of suspected insider trading, particularly in "deadline-resolved" contracts (e.g., "Will Event X occur by Date Y?"). While a companion methodology paper introduced the Information Leakage Score (ILS) and its deadline-resolved extension (ILS $_{dl}$ ), the framework lacked an end-to-end empirical validation on real-world cases.

Key challenges addressed include:

Anchor Quality: Traditional ILS calculations often rely on a proxy anchor (e.g., $T_{resolve} - 1$ hour), which fails to capture pre-event information flow if the market price collapses due to public belief that the event will not happen before the resolution.
Data Gaps: Accurate calculation requires recovered event timestamps ( $T_{event}$ ), continuous order book (CLOB) price data from market opening, and per-trade wallet history.
Methodological Rigor: Determining whether observed price movements represent genuine informed trading or merely market mispricing and resolution-criteria uncertainty.

2. Methodology

The paper implements the Deadline-ILS (ILS $_{dl}$ ) framework end-to-end on the 2026 U.S.–Iran conflict cluster from the ForesightFlow Insider Cases (FFIC) inventory.

Core Definitions

ILS $_{dl}$ Formula:
$ILS_{dl}(M) = \frac{\Delta_{pre}^{dl}}{\Delta_{total}^{dl}} = \frac{p(T_{event}^-) - p(T_{open})}{p_{T_{resolve}} - p(T_{open})}$
Where $p(T_{event}^-)$ is the price one minute before the public event, and $p_{T_{resolve}}$ is the final resolution price.
Scope Conditions: The score is valid only if $|\Delta_{total}| \geq \epsilon$ (market moved significantly), the market opens near 0.5 (edge-effect check), and the result is robust to anchor offsets.
$T_{event}$ Recovery: A Tier 3 LLM (Claude Haiku 4.5) with web-search tools recovers the timestamp of the first public occurrence of the underlying event, cross-verified against $\geq 3$ independent news sources.
Hazard-Rate Estimation: The framework assumes an exponential time-to-event distribution $S(\tau) = \exp(-\lambda\tau)$ . The hazard rate $\lambda$ is fitted via Maximum Likelihood Estimation (MLE) for specific categories (military, corporate, regulatory).

Evaluation Pipeline

Hazard-Rate Fitting: Estimate $\lambda$ per category using 20 resolved markets per category.
Single-Case Computation: Apply the pipeline to the largest applicable market ("US forces enter Iran by April 30").
Cross-Market Wallet Analysis: Analyze wallet overlap between correlated contracts to detect coordination.
Methodological Refinement: Identify infrastructure and classification gaps preventing scaling.

3. Key Contributions

First End-to-End Empirical Validation: The paper moves the ILS $_{dl}$ framework from theoretical definition to quantitative application on a documented insider trading cluster.
Demonstration of Anchor Sensitivity: It empirically proves that using an article-derived $T_{event}$ yields a fundamentally different (and more accurate) signal than a resolution-anchored proxy.
Hazard-Rate Characterization: It provides the first empirical estimates of time-to-event distributions for prediction markets, distinguishing between military/geopolitical and regulatory events.
Infrastructure Gap Identification: It explicitly identifies the lack of continuous per-trade data collection as the primary bottleneck for detecting pre-event coordination.

4. Key Results

A. Hazard-Rate Estimation

The study fitted exponential distributions for three categories (sample size $n=20$ per category where possible):

Military/Geopolitics: Adequate fit ( $KS\ p=0.609$ ). Half-life: 2.3 days. Median lead time: 2.2 days.
Corporate Disclosure: Preliminary adequate fit ( $n=5$ , $KS\ p=0.616$ ). Half-life: 4.5 days.
Regulatory Decision: Rejected ( $KS\ p=0.013$ ). The distribution is bimodal, mixing short-announcement markets with long-deliberation markets. This category requires sub-categorization before the model can be used.

B. Single-Case Analysis: "US forces enter Iran by April 30"

This was the only market in the cluster satisfying all four requirements (recovered $T_{event}$ , positive $\tau$ , CLOB coverage, valid $\Delta_{total}$ ).

Market Context: Opened at $p=0.25$ , peaked at $0.46$, then collapsed to near zero by April 8 (day before resolution), before resolving YES on April 9.
ILS $_{dl}$ Comparison:
- Article-Derived Anchor ( $T_{event}$ = April 3): $ILS_{dl} = \mathbf{+0.113}$ . Indicates mild positive front-loading (approx. 11% of the total move happened before the event).
- Legacy Proxy ( $T_{resolve} - 1h$ ): $ILS_{dl} = \mathbf{-0.331}$ . Indicates price moved against the outcome.
- Significance: The proxy failed because it captured the market's collapse in belief (thinking the event wouldn't happen) rather than the pre-event information flow. The values differ by 0.444 and are on opposite sides of zero.
Drift Analysis: Short-window variants (30-min, 2-h) were exactly zero, ruling out a last-minute "informed spike." The signal was a mild, early drift, not a concentrated spike.

C. Cross-Market Wallet Analysis

Data Limitation: Available trade history only covered the resolution-settlement window (April 8–11), not the pre-event period.
Findings: 332 wallets were active in both major Iran-cluster markets. Top wallets traded between \ $170k and \$ 1.97M.
Interpretation: Because the data is post-event, this signal represents settlement arbitrage (trading on the resolution outcome), not pre-event informed coordination. The methodology cannot currently detect the "canonical" insider trading signature due to data gaps.

5. Significance and Future Work

Methodological Refinements Required

The evaluation surfaced five critical constraints for scaling the methodology:

Positive- $\tau$ Requirement: Markets must open before the event occurs ( $T_{event} > T_{open}$ ). Duration markets created after an event starts are out of scope.
Regulatory Sub-categorization: Regulatory markets must be split into "announcement" vs. "formal deliberation" to fix the bimodal hazard distribution.
Sample Expansion: The corporate-disclosure sample ( $n=5$ ) is too small for stable hazard estimation.
Continuous CLOB Collection: Current infrastructure misses 82% of relevant markets due to indexing latency. Continuous collection from $T_{open}$ is required.
Continuous Per-Trade Collection: To detect coordination, the system must capture individual trades from market opening, not just settlement data.

Conclusion

The paper serves as a methodological proof of concept. It demonstrates that the Deadline-ILS extension successfully distinguishes signal from proxy artifacts when using accurate event timestamps. However, the current inability to access pre-event trade history prevents a full population-level evaluation of insider trading. The authors propose a "work programme" focused on building continuous data pipelines and refining classification taxonomies to enable future large-scale detection.

Data Availability: The FFIC inventory and the resolution-typology classification of 911,237 Polymarket markets are released as open datasets on GitHub.