Reading Numerai Rankings and Committed-Capital Returns

A leaderboard rank is not a payout rank, and sorting a table by stake should not renumber a model’s published rank. Check published rankings on Numerai; the nmrdash dashboard shows aggregate model scores, rather than a ranked list of individual models.

Revised September 8, 2026. Earlier current-ranking examples and legacy-only column guidance are retained in repository history, not presented as today’s standings.

Use the actual reputation field

Classic displays explicit CORR60 and MMC60 reputation fields. Crypto retains its CORR/MMC fields. Signals shows available Alpha/MPC and neutral-score fields. A missing new field remains missing; an older field with a similar name is not silently substituted.

Reputation also aggregates a history of scores. It is not the current round’s score and does not by itself determine a position’s payout. The metric reference lists the required horizon and version context.

Read returns against committed capital

Numerai’s published model/account return convention counts outcomes settled in a window against NMR actually committed, including initial locked capital and additional funding beyond reusable releases. Idle funds are excluded, and recycling a settled position does not repeatedly increase the funding basis. The upstream return definition is the authority for those columns.

That is distinct from a settled position return, where net NMR is divided by that position’s principal without annualization. The staking return distribution uses this position basis. A 13-week return and an individual daily-round percentage need not agree even for the same model.

Rank changes require a dated cohort

A model may move because its score changed, its reputation definition changed, or other models entered or left the comparison. Historical rankings need the snapshot that existed at that time. I would not use today’s survivors to reconstruct an earlier leaderboard, or interpret a missing rank as an ordinal row number.

Method: this is a definition guide, not a current rank distribution or backtest. Read each published ranking against its explicit upstream score field, observation date and return convention.