How every number is made

Where the data comes from

Every trade, position and profit-and-loss figure comes from Polymarket’s public APIs (leaderboards, trades, positions, public profiles, order-book price history) and the Polygon blockchain. We don’t estimate or invent trades, and we don’t buy private data.

• Trades stream in live. Leaderboards, stats and positions refresh every few minutes.
• Every screen shows when its data last updated. If an upstream source is down, you see the last good data, marked as stale, never a guess.
• Each trade links to Polymarket so you can verify it yourself.

Star rating (1–5★)

Stars are our primary rating and the only one we rank by, because in testing it is the signal that held up best (see How we test).

A trader’s last 300 settled or closed positions are scored on four components:

Cutoffs are calibrated on real top traders so that roughly the bottom 30% get 1★, then 2★ up to the 55th percentile, 3★ to the 75th, 4★ to the 90th, and the top 10% get 5★. Fewer than 15 settled positions means no rating.

Two details matter a lot and were caught by our own backtest:

• Early exits count at what was actually realized. A position sold before resolution isn’t scored as a full win or loss.
• Unredeemed losses are included in a consistent time window. Losing positions are often never “redeemed” on Polymarket; ignoring them flatters a trader, and including all of them forever buries long-time traders in old losses.

Skill evidence

Big profits can be luck. For each trader we ask: if every bet had gone the other way at the same stake, how often would a record look this good? This is a sign-randomization test, the same idea used in the academic work on Polymarket skill (Gómez-Cram et al., 2026).

• Proven skill: p < 0.01 with a positive edge.
• Promising: p < 0.10 with a positive edge.
• Could be luck: no statistical evidence either way.
• Not enough data: fewer than 15 resolved positions.

The p-value uses an exact-moment normal approximation of the randomization distribution rather than thousands of random draws; results are nearly identical (rank correlation 0.177 versus 0.169 with 2,000 Monte Carlo draws) and it is fast enough to run live.

How we test: the monthly Edge Report

A rating is only useful if it predicts what happens next. Every month we:

1. Build a pool from Polymarket’s top-100 leaderboards (day, week, month, all time).
2. Score every trader using only positions that resolved before a cutoff date.
3. Measure their actual return on the positions that resolved after it (the test window).
4. Compare signals: rank correlation (Spearman) between score and future return, and the return gap between the top and bottom fifth with a 95% bootstrap range.
5. Publish the result, including signals that failed and months when ours get worse.

Latest result (October 2026, 132 traders, test window of 45 days):

Known limits: the pool comes from today’s leaderboards, which favours recent winners (survivorship bias); every signal shares that bias, so comparing them is still fair, but absolute returns are flattered. 132 traders and 45 days give wide ranges. Up to 1,000 recent positions per trader means very high-frequency traders are under-covered. Full report and raw numbers →

Smart Money

For each open market we add up the positions of this month’s top traders rated 3★ or more, ignoring positions under $50 and markets that have settled. The leading outcome is the one most of those traders back.

Divergence is their share on that outcome minus its market price, after adding two neutral “traders” at the market price, so that two people agreeing doesn’t read as certainty. +20 pts means rated traders are much more confident than the market. It is a description of positioning, not a forecast.

Alert context

Alerts compare a trade with the trader’s own history:

• Conviction: trade size versus their usual order.
• New territory: first trade in a category.
• Consensus: other traders you follow on the same side.
• Lateness: how far the price has moved since they traded, so you can see if the opportunity is already gone.

Fills of the same order are grouped into one alert, and there is an hourly cap per trader so a busy trader can’t flood you. These describe the trade. They aren’t recommendations.

Copy simulator (Pro)

Replays a trader’s orders over a window (default: 30 days, $1,000, 2-minute delay) as if you had followed them:

• Fills of the same token and side within 5 minutes are merged into one order.
• Their median order maps to a fixed share of your bankroll; bigger or smaller orders scale with it, capped per order and by the cash you have left.
• You enter at the market price after the delay (from order-book price history), then lose 1¢ to slippage. When they sell part of a position, you sell the same share.
• Resolved markets count at 1 or 0; open ones at today’s price.
• Fees, order-book depth and partial fills are not modelled, so real results would usually be somewhat worse.

The result is labelled copyable, partly (speed matters), lost (the edge didn’t survive the delay), no edge or insufficient. It is a historical simulation, not a promise.

AI summaries

The one-line trading style on a profile and the weekly brief are written by Anthropic’s Claude, server-side, from structured public trading facts only (categories, typical prices, holding times). No personal data about you is sent. The model is instructed to describe, never to recommend, and output is cached for 24 hours. If it is unavailable we show a template sentence built from the same facts.

What we don’t do

• No affiliate links, no paid placement, no sponsored traders.
• No trading, deposits or wagering on your behalf, ever.
• “Unusual activity” means a statistical outlier, never an accusation of insider trading.
• Nothing in Edgewatch is financial, investment or betting advice.