How reliable is CrowdIntel's insider detection?

Reliability rises with sample size. Individual trade flags are probabilistic. Published investigations clear a quantitative bar combining sample size, excess win rate, realized PnL, and statistical significance versus the relevant base rate.

Can the detection engine be gamed?

Any deterministic detection system has edge cases a determined adversary can exploit. CrowdIntel treats scoring as a living system — thresholds, signal weights, and the specific feature set are revised as adversaries adapt. The public methodology deliberately omits parameters that would only help someone trying to evade detection.

Why shrinkage-adjusted win rate instead of raw?

Raw win rate punishes long-track-record sharps and rewards lucky newcomers. Shrinkage estimation pulls small-sample results toward a prior and only lets data override the prior when the sample is substantial. A wallet with 80% over 200 bets is far more credible than 90% over 10 bets, and shrinkage formalizes that intuition.

Does CrowdIntel ever retract an investigation?

Yes — if new resolved bets drop the cluster below our publication bar, or if evidence suggests an investigation was opened on misclassified data. Investigations are living documents; they can be merged, split, or retired as evidence evolves.

Is the detection engine open source?

No. The scoring stack is proprietary. The principles are transparent; the parameters are not, because that would undermine the product's core value for everyone who uses it.

Methodology

How CrowdIntel detects Polymarket insiders

Every trade on Polymarket leaves a signature on Polygon. CrowdIntel reads those signatures at scale through a multi-signal statistical engine. Below is the public overview. The specific parameters stay proprietary — publishing them would help the people we're trying to catch.

What we look at

Every on-chain Polymarket trade is scored against a set of signals that collectively separate skill, luck, and information advantage. Signals group into four families:

Trade shape

Characteristics of the trade itself — size, timing, odds at entry — relative to the market's baseline behavior. Insiders usually leave footprints in how, not just when, they trade.

Wallet history

Lifetime win rate, realized PnL, bet concentration, and prior accuracy in the market's category. A track record either supports or undermines the pattern of an individual trade.

Network topology

On-chain links between wallets — shared funding sources, overlapping counterparties, correlated timing. One entity running many wallets to mask size or inflate consensus leaves the hardest-to-fake footprint of all: the funding trail on Polygon.

Cluster reputation

When a group of linked wallets has already proven itself across many resolved bets, future trades from the same operators inherit that weight. Reputation is earned through statistical track record, not claimed through branding.

When CrowdIntel opens an investigation

A flagged trade is a starting point. A public investigation — where CrowdIntel names a cluster or wallet and publishes evidence — is a much higher bar. Investigations must clear a combination of:

Sufficient sample size. Small-sample anomalies are noise. We require enough resolved bets that luck alone can be ruled out at standard significance levels.
Excess win rate versus the category base rate. Beating 50% by a little is not interesting; beating it by enough to reject a no-skill null hypothesis is.
Positive realized PnL. Win rate can be inflated by betting heavy favorites while losing money. We require the wallet or cluster to actually make money on their supposed edge.
Statistical significance. We report a p-value against the null hypothesis of no edge. Investigations require it to clear an internal threshold — stricter thresholds unlock higher visibility.
Coordination evidence (for cluster investigations). Multi-wallet investigations require the wallets to be linked on chain by more than coincidence.

The exact numeric thresholds are not published. They are tuned against observed adversarial behavior and revised without notice.

Shrinkage-adjusted win rate

Raw win rate is noisy. A wallet with 2 wins in 2 bets shows 100% — meaningless. CrowdIntel computes a Bayesian shrinkage estimate on every wallet — a posterior that pulls small-sample results toward a category-appropriate prior and converges to raw win rate as sample size grows.

In practice, a wallet with 80% over 200 bets is ranked above a wallet with 90% over 10 bets — because the shrinkage estimator refuses to be fooled by the small sample. Both raw and shrunk win rates are displayed on every wallet profile; leaderboards rank by the shrunk value.

Data pipeline

1. Polygon blockchain. Source of truth. Every Polymarket trade, USDC transfer, and market resolution is read directly from Polygon state.
2. Custom indexer. A subgraph ingests every Polymarket contract with short cursor lag from confirmation.
3. Enrichment. Each new trade is scored, linked to its funding trail, and assigned to any clusters it participates in.
4. Outcome tracking. Resolved markets update wallet statistics — win rate, PnL, category-specific edges, and the shrinkage estimator.
5. Investigation engine. Periodic evaluation opens new investigations when thresholds clear and re-scores existing ones as new bets resolve.

Known limits

Transparency on what CrowdIntel cannot see:

Sophisticated laundering. Operators who route funds through many hops or fresh exchange withdrawals can partially defeat on-chain clustering. Detection is updated against patterns as they emerge.
Off-chain coordination. Chat-room tipping, dark pools, and private signal groups leave no on-chain footprint unless members share funding.
Identity. CrowdIntel proves statistical anomalies. It does not and cannot prove who operates a wallet.
Early-stage markets. Thin, brand-new markets generate unstable scores regardless of method. We weight them accordingly and require more resolved bets before taking them seriously.

Reproduce the evidence

Every wallet page links to Polygonscan. Every trade has an on-chain tx hash. Every investigation lists the wallets, funder, and supporting evidence. Our scoring parameters stay private; the raw data is public and any individual claim can be verified against Polygon directly.

Want to go deeper?

Every term used here is defined in the glossary. See the methodology applied on any whale profile or investigation.