Skip to content
Convexly
Lesson 7 of 12

Concentration

When one event carries 60% or more of a wallet's net result, the read carries the concentrated style label: the per-position edge can be clean while the dollar outcome rode on a few events.

The answer first

Concentration measures how much of a wallet's net realized result comes from its single biggest event. At 0.6 or above, the read carries the concentrated style label: the dollar result is mostly an echo of one outcome, so however clean the equal-weighted per-position edge is, the path anyone following the wallet would live through is the lumpy one, not the average. The label travels with the verdict and never changes it; the sufficiency floors (30 resolved positions, 20 independent events) and the interval decide the state.

The intuition, with one worked example

Take a made-up record. A wallet nets $100 across 50 resolved positions, and $80 of it came from one event. Concentration is 80 / 100 = 0.8, above the 0.6 cap, so the concentrated label attaches. Whatever its edge and win rate say, the dollar outcome is mostly one good call (or one lucky one; the record cannot tell you which), and the caveat travels with the verdict. Now make the same wallet's biggest event worth $150 while everything else lost $50 combined. The net is still $100, and concentration is 150 / 100 = 1.5. The ratio exceeds 1 whenever the biggest win is larger than the whole net result because the rest of the book lost money. Values above 1 are real and observed on live records, and they mean extremely top-heavy, not broken math.

A real row: Edge Score 95.8, unreadable record

A real row from the frozen 2026-06-09 scan of our own published top-50 cohort (full table at /research/top50-skill-scan). Wallet 0x0d83ca sat near the top of the board with an Edge Score of 95.8 and positive net PnL. Its skill read:

  • Realized entry edge +1.9pp across 35 resolved positions, 95 percent interval [-7.8, +11.9].
  • Concentration ratio 2.70: the single biggest winning event was 2.7 times the net result, meaning everything else netted negative.

A reader who saw only the board position would call this wallet elite. The resolved record says two things instead: the interval includes zero, so the edge is not separable from chance, and the concentrated style attaches because the result is one event. (The frozen 2026-06-09 scan predates the 2026-07-15 reframe, so its artifact reported this row as too concentrated to read; the numbers are unchanged.) In the same scan, 15 of the 35 testable wallets crossed the same threshold, which is a large part of why raw profit leaderboards mislead.

The actual method

  • The ratio is the biggest single event's realized result divided by the net realized result. It is not bounded to [0, 1] by design; it is null when the net result is not positive, since the share of a non-positive total is not meaningful.
  • The 0.6 cap is an operator-set conservative floor, reused from the cohort screen in the skilled-wallet registry. It is deliberately not tuned per wallet.
  • Since the 2026-07-15 realized-edge reframe the threshold attaches a style label instead of demoting: the verdict comes from the sufficiency floors (30 resolved positions AND 20 independent resolved events, fail-closed) plus the interval, and the label travels with it. A skilled verdict with the concentrated style renders with a caution accent and a mandatory caveat, never the bare positive accent.
  • More positions do not clear the concentrated label. Only a record whose net result stops depending on one event clears it.
  • A cousin check runs on categories: when more than 60% of a record sits in one market category, the evidence grade in the cohort deliverable is capped and the dependence travels as a named caveat.

Where you see this on the site

The analyzer reports the concentration value with every read, and the concentrated style note appears on the verdict when it crosses the cap. The same label drives the board annotations, the badge copy, the API styles field, and the binding caveat column in the enterprise deliverable. The frozen wording lives in the lexicon, and the threshold is documented on the methodology page.

Why it was once called a flag

Before the 2026-07-15 realized-edge reframe the same 0.6 threshold demoted the read to a flagged state, which is why older material calls it the concentration flag. That behavior is retired. The label describes how the money was made, the sufficiency floors and the interval decide the state, and the flagged state is reserved for a wallet-integrity screen that is not live.

What this does NOT mean

The concentrated label is not an accusation and not a verdict on the wallet. Concentrating on one high-conviction event can be a legitimate style, and since 2026-07-15 the machine treats it exactly as that: a style, stated plainly next to the verdict. A record with an interval that clears zero and a concentrated label reads as skilled on the per-position record, with the caveat that the dollar outcome rode on a few events and that concentrated records can mask gamed activity. The label is not permanent: it describes the record as it stands, and a record that broadens its base of resolved results reads differently.

Next lesson: Edge Score

Convexly publishes new methodology research roughly every 6-8 weeks plus the /learn series on a rolling cadence. Get the next paper in your inbox when it ships:

Frequently asked

What does the concentration number measure?

The biggest single event's contribution to the wallet's net realized result: biggest event divided by net total. At 0.6 or above, the read carries the concentrated style label: one outcome carries so much of the dollar total that the headline figure is mostly an echo of that event, even when the equal-weighted per-position edge is clean. The label is about how the money was made, not about the wallet's character, and since 2026-07-15 it never changes the verdict.

How can concentration be higher than 1?

When the biggest winning event is larger than the whole net result because other events netted negative. A wallet whose biggest event contributed 150 units while everything else lost 50 units nets 100, and its concentration is 150 / 100 = 1.5. Values above 1 are real and observed; they mean the record is even more top-heavy, not that the math broke.

Why do more positions not clear the label?

Because the label is about proportion, not count. Adding 200 tiny positions to a record whose net result still comes from one event leaves the same composition: the dollar figure remains an echo of that single outcome. The only thing that clears the concentrated label is a record whose result no longer depends on one event. Sample-size problems are handled separately, by the 30-position and 20-independent-event floors, which more resolved, independent positions do fix.

What was the concentration flag?

The older name for the same threshold. Before 2026-07-15 a ratio at or above 0.6 demoted the read to a flagged state; since the realized-edge reframe it attaches the concentrated style label instead and never changes the verdict. The guardrail moved to the independent-event floor: the realized-edge interval is an event-clustered bootstrap, so its effective sample size is the count of independent resolved events, and a would-be-skilled read with fewer than 20 of them is capped at thin sample. That floor fails closed when the count is missing.

How common is high concentration among top wallets?

Common. In Convexly's frozen 2026-06-09 scan of its own published top-50 Polymarket cohort, 15 of the 35 testable wallets crossed the 60 percent threshold (that scan predates the 2026-07-15 reframe, so it reported them as too concentrated to read; today the same records would carry the concentrated style label on their interval read). Barbell-shaped records are the norm at the top of profit leaderboards, which is a core reason raw PnL rank is a poor skill signal.

Related explainers

Related reading