ForesightFlow: Real-Time Detection of Informed Trading in Decentralized Prediction Markets

Abstract

Decentralized prediction markets such as Polymarket aggregate

dispersed beliefs into continuously updated price signals, but their

on-chain transparency and pseudonymous participation also make them

an unusually fertile environment for informed trading on material

non-public information. Recent empirical work has documented hundreds

of millions of dollars in anomalous profits on Polymarket between

2024 and 2026; existing detection approaches are almost exclusively

post-hoc and offer no actionable signal during the window when

informed flow is actually moving prices.

We propose ForesightFlow, a real-time detection framework that

combines classical microstructure measures of informed trading —

adapted to the discrete, bounded structure of binary outcome

markets — with on-chain wallet features unique to decentralized

platforms. We introduce the Information Leakage Score (ILS), a label

generator that quantifies how much of a market's terminal information

move was priced in before the corresponding public news event, and

show that ILS admits a clean interpretation in terms of the Murphy

decomposition of the Brier score. As an empirical anchor we release

the ForesightFlow Insider Cases (FFIC) inventory, a curated

validation set mapping eight publicly documented insider episodes to

concrete Polymarket market identifiers. The historical-backfill stage

of the data pipeline is operational, indexing 911,237 markets

(865,725 resolved) and 17.9 million trades across 10,410 markets

above the $50K volume cutoff at which subgraph trade history becomes

reliably available.

A pilot empirical study tightens the methodology in three ways: a

resolution-anchored proxy for T_news does not separate event-resolved

markets from a matched control population; high ILS on high-consensus

markets reflects a formula edge effect; and the proxy-based ILS

distribution is not robust to anchor choice. We specify three

operational scope conditions for ILS — restriction to event-resolved

markets with substantive uncertainty at T_open, recovery of

article-derived T_news, and a robustness check across multiple anchor

offsets. A proof-of-concept article-derived T_news recovery on the

Epstein-files Barak market shifts ILS only from 0.553 to 0.570

relative to the proxy, suggesting that proxy quality was not the

binding constraint and that wallet-level features rather than ILS

alone are required to identify informed flow.

An attempt to scale to the full FFIC inventory revealed that

documented insider cases are systematically deadline-resolved ("Will

event X occur by date Y?"), falling outside the original ILS scope.

We extend the score to a deadline-ILS variant anchored to the

underlying event timestamp, implement a categorical exponential-hazard

baseline for the time-to-event distribution, and apply the extension

to the 2026 U.S.–Iran conflict cluster. On the cleanest applicable

case ("US forces enter Iran by April 30," $269M cumulative volume),

the deadline-ILS shifts from −0.331 at the resolution-anchored proxy

to +0.113 at the article-derived event timestamp, demonstrating that

the extension distinguishes signal from proxy artefact. The empirical

evaluation surfaces five concrete methodological refinements that

constrain any future application of the framework: a positive lead-

time requirement, regulatory sub-categorization, corporate-disclosure

sample expansion, continuous CLOB price collection, and continuous

per-trade collection from market opening.

System code, the FFIC inventory, and a full resolution-typology

classification of the 911,237-market corpus are released openly at

github.com/ForesightFlow.

Keywords: prediction markets, informed trading detection, real-time market surveillance, information leakage score, market microstructure, Murphy decomposition, hazard rate estimation, PIN VPIN Kyle's lambda

JEL Classification: D82, G14, G18, G28, C53, C58, C41