💰 quantitative finance

ForesightFlow: An Information Leakage Score Framework for Prediction Markets

This paper introduces ForesightFlow, an Information Leakage Score framework for detecting informed trading in prediction markets, and proposes a deadline-anchored extension to address the methodological limitations revealed by empirical evaluations of public-event timestamps and documented insider trading cases.

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 global, digital town square where people bet on future events: "Will this politician win?" "Will this company launch a new product?" "Will a war start by Friday?" This is a prediction market. In these markets, the price of a bet acts like a weather forecast for the future. If the price of "Yes" goes up, it means the crowd thinks the event is more likely to happen.

The paper introduces a new tool called ForesightFlow, designed to spot when someone is cheating in this town square.

The Problem: The "Insider" in the Crowd

In a fair market, everyone learns about an event at the same time. But in these digital markets, because they are anonymous and operate on blockchain technology, it's very easy for someone with secret information (an "insider") to buy bets before the news hits the public.

Think of it like a horse race where a jockey knows the horse has a broken leg before the race starts. They buy a ticket on the other horses, wait for the race to start, and then cash in a massive profit when the favorite collapses. The paper notes that in recent years, insiders have made hundreds of millions of dollars this way on platforms like Polymarket.

The problem is that current ways of catching these cheaters only work after the race is over. By the time we know who cheated, the money is already gone, and the "price signal" (the public's belief) has already been distorted. We need a way to spot the cheater while they are placing their bets.

The Solution: The "Leakage Score"

The authors built a system called ForesightFlow to act like a smoke detector for these markets. The core of this system is a metric called the Information Leakage Score (ILS).

Here is the analogy for how it works:
Imagine a movie is about to be released.

The Opening Price: When the betting market opens, the price reflects what the public knows right now.
The News Event: The moment the movie trailer is officially released to the world.
The Resolution: The moment the movie actually comes out and the bet is settled.

The Information Leakage Score asks a simple question: "How much did the price move before the official trailer was released?"

Score of 0: The price stayed flat until the trailer came out. This is a fair market; no one knew anything early.
Score of 1: The price jumped all the way to the "winning" level before the trailer was even released. This suggests someone knew the outcome and bet on it early. The information "leaked" out before it was supposed to.

The "Scope" Rules: Knowing When the Score Works

The paper is very careful to say this score doesn't work for every single bet. The authors found three specific rules where the score is reliable:

The "Surprise" Rule: The score only works if the event is actually uncertain. If everyone already knows the answer (e.g., the price is 99% "Yes" from the start), a tiny price movement looks huge mathematically, but it's just noise, not cheating.
The "Deadline" vs. "Event" Rule: This is a crucial discovery in the paper.
- Event Markets: "Will X happen?" (e.g., "Will the President sign a bill?"). The score works well here.
- Deadline Markets: "Will X happen by Friday?" The authors found that most of the famous cheating cases happened in these "Deadline" markets. The original score didn't work for them because the "news" is just the clock running out.
- The Fix: The authors created a new version of the score (called Deadline-ILS) specifically for these "Will it happen by Friday?" bets. This new version compares the price to a "baseline expectation" rather than just the opening price.
The "Wallet" Check: The score isn't enough on its own. The system also checks who is betting. If a brand-new wallet (created 10 minutes ago) with no history suddenly dumps a huge amount of money on a specific bet right before news breaks, that's a red flag. If an experienced trader with a long history does it, it might just be a smart guess.

The Pilot Study: What They Learned

The authors tested their system on real data. They found some interesting things:

The "Fake" Signal: When they tried to use a simple guess for "when the news happened" (instead of finding the actual article), the system got confused and flagged innocent markets as guilty. This taught them that they need to find the exact moment the news broke to make the score accurate.
The "Deadline" Gap: They realized their original tool couldn't catch the most famous cheaters because those cheaters were betting on "Deadline" markets. This forced them to build the new "Deadline-ILS" version.
The "Barak" Case: They looked at a specific market about a political figure named Barak. Even with the perfect score, they couldn't prove it was cheating just by looking at the price. They had to look at the wallets and the timing to realize the price movement was actually people arguing about the rules of the bet, not people with secret info.

The Big Picture

The paper concludes that catching cheaters in real-time is possible, but it requires a very specific toolkit:

A Score: To measure how much the price moved too early.
A Typology: To know if the bet is an "Event" or a "Deadline" (and use the right math).
Wallet Forensics: To see if the people betting look like insiders (new wallets, huge bets) or just regular people.

The authors have released their code and a list of known cheating cases (the "FFIC inventory") so others can use and improve this system. Their goal isn't just to catch cheaters for profit, but to make sure these prediction markets remain trustworthy tools for the public, rather than playgrounds for people with secret advantages.

1. Problem Statement

Decentralized prediction markets (e.g., Polymarket) aggregate dispersed private information into price signals but operate in a regulatory gray area with pseudonymous participation. This environment facilitates informed trading on material non-public information (MNPI).

The Gap: Existing detection methods (e.g., Mitts & Ofir, 2026) are almost exclusively post-hoc, relying on realized profits and resolution outcomes. They offer no actionable signal during the trading window when informed flow is moving prices.
The Challenge: Real-time detection requires features that are causal (observable before resolution) and labels that do not leak future information. Furthermore, the specific structure of many insider trading cases (deadline-based contracts) does not fit standard event-resolved models.

2. Methodology: The ForesightFlow Framework

The paper proposes a system to detect informed flow in real-time using a combination of information theory, market microstructure, and on-chain forensics.

A. The Information Leakage Score (ILS)

The core innovation is the Information Leakage Score (ILS), a scalar label generator for resolved binary markets.

Definition: $ILS(M) = \frac{\Delta_{pre}}{\Delta_{total}} = \frac{p(T_{news}) - p(T_{open})}{p(T_{resolve}) - p(T_{open})}$ $I L S (M) = \frac{Δ _{p r e}}{Δ _{t o t a l}} = \frac{p ( T _{n e w s} ) - p ( T _{o p e n} )}{p ( T _{r eso l v e} ) - p ( T _{o p e n} )}$
- $T_{open}$ : Market creation.
- $T_{news}$ : Timestamp of the first public mention of the resolution-relevant event (recovered via UMA evidence URLs, GDELT, or LLM-assisted search).
- $T_{resolve}$ : Formal resolution time.
- $p(t)$ : Mid-price of the "YES" token.
Interpretation: ILS quantifies the fraction of the total price movement that occurred before the public news event.
- $ILS \approx 1$ : Perfect anticipatory pricing (strong leakage).
- $ILS \approx 0$ : Market reacted only after public news.
Theoretical Basis: The authors connect ILS to the Murphy decomposition of the Brier score. High ILS indicates that the "Resolution" component (discriminative power) was front-loaded before the public information set caught up.

B. Resolution Typology & Scope Conditions

The authors identify that ILS is only interpretable under strict conditions, leading to a three-way classification of markets:

Event-Resolved: Specific observable event triggers resolution (e.g., election results). ILS is valid here.
Deadline-Resolved: Market resolves NO if an event fails to occur by a date (e.g., "Will X happen by Friday?"). Standard ILS is invalid here because $T_{news}$ is undefined or trivial.
Unclassifiable: Sports, counts, etc.

Scope Conditions: To avoid spurious signals, ILS is only reported if:
- Edge-Effect Check: Opening price is not near consensus ( $p_{open} \in [0.1, 0.9]$ ).
- Trivial-Resolution Check: Total price move $|\Delta_{total}| > 0.05$ .
- Anchor-Sensitivity: The ILS value must be robust across different $T_{news}$ anchor specifications.

C. Real-Time Feature Pipeline

To detect flow before resolution, the system computes a feature vector $x(M, t)$ combining:

Adapted Microstructure Measures:
- Order Imbalance (OI): Buy/sell volume ratios.
- VPIN-style Toxicity: Volume-synchronized order flow imbalance.
- Kyle's Lambda: Estimated price impact per unit of order flow.
- Variance Ratio & Two-Sidedness: Diagnostics for persistent drift vs. temporary noise (adapted from Nechepurenko, 2026a).
On-Chain Wallet Features:
- Wallet Novelty Score: A composite of wallet age, participation history, funding source concentration, and entry timing relative to resolution.
Category-Conditional Priors: The system focuses on three high-risk categories: Military/Geopolitical, Corporate Disclosures, and Regulatory Decisions.

D. The Deadline-ILS Extension

A critical finding of the pilot study was that documented insider cases are predominantly deadline-resolved. The paper extends ILS to this domain:

Mechanism: Instead of anchoring to a news event, the score anchors to a Bayesian baseline belief ( $\theta_{T_{open}}$ ) derived from a constant-hazard survival model.
Formula: $ILS_{dl} = \frac{p(T_{event}^-) - \theta_{T_{open}}}{p(T_{resolve}) - \theta_{T_{open}}}$ .
This allows the detection of informed traders who know an event is imminent in a "Will X happen by Date Y?" contract.

3. Key Contributions

Reformulation: Shifts the problem from post-hoc forensics to real-time inference, requiring causal features.
ILS Metric: Introduces a theoretically grounded label generator linked to proper scoring rules (Murphy decomposition).
Typology & Scope: Defines strict operational boundaries (Event vs. Deadline) to prevent false positives, validated by empirical data.
Hybrid Architecture: Combines classical microstructure measures (PIN, VPIN) with unique on-chain wallet features (novelty, funding concentration).
FFIC Inventory: Releases a curated validation set of 8 documented insider trading episodes (e.g., Iran strikes, Google Year-in-Search) mapped to specific market IDs.
Deadline Extension: Formalizes the extension of the framework to deadline contracts, addressing the structural gap in existing literature.

4. Results & Empirical Findings

The paper reports a pilot empirical study on 725 event-resolved markets and a validation audit of the FFIC inventory.

Negative Pilot Findings:
- Using a resolution-anchored proxy ( $T_{resolve} - 24h$ ) for $T_{news}$ failed to separate informed markets from controls.
- Reason: The proxy was structurally misaligned. For event-resolved markets, the actual news often precedes the resolution by days, making the "pre-news" window actually "post-news."
- Anchor Sensitivity: ILS values were highly sensitive to the choice of time anchor, failing the robustness check.
Case Study (Barak Market):
- A deep dive into the "Ehud Barak" market showed that while a proxy-based ILS was high (0.55), using the article-derived timestamp (actual news release) shifted the ILS slightly but revealed that the "early" trades were actually reactive to public news or resolution-criteria arbitrage, not insider information.
- Conclusion: ILS alone is insufficient; it must be combined with wallet-level features (e.g., identifying professional vs. new wallets) and strict scope checks.
Structural Audit:
- The audit of the FFIC inventory revealed that 100% of documented insider cases were deadline-resolved, falling outside the original ILS scope.
- This necessitated the Deadline-ILS extension (Section 7) to make the framework applicable to real-world insider trading patterns.

5. Significance & Future Work

Public Good: Provides the first transparent, replicable, real-time surveillance tool for decentralized prediction markets, addressing the "authority paradox" where markets gain social credibility while having immature microstructures.
Methodological Bridge: Connects the literature on proper scoring rules (forecasting accuracy) with blockchain forensics (informed trading detection).
Open Source: The authors release the ForesightFlow Insider Cases (FFIC) inventory, the full resolution typology for 911,237 markets, and all system code.
Next Steps:
- Full empirical evaluation of the Deadline-ILS on the 2026 U.S.–Iran conflict cluster (reported in a companion paper).
- Training the detector model on the refined, deadline-extended dataset.
- Addressing infrastructure limits (e.g., The Graph indexer failures for billion-dollar markets).

Conclusion: ForesightFlow establishes that informed trading in prediction markets is detectable in real-time, but only if the methodology accounts for the specific contract typology (deadline vs. event), uses robust news timestamps, and integrates on-chain wallet behavior with microstructure signals. The framework moves the field from retrospective "whodunit" analysis to proactive market integrity monitoring.