Numerai uses NMR staking to attach financial consequences to submitted predictions. The useful signal is a participant’s willingness to risk NMR on a model. A payout rewards the configured prediction scores; it does not distribute the hedge fund’s investment return.
Revised September 8, 2026. This explanation replaces the earlier universal payout formula and 25% cap.
A prediction, a position and a portfolio are different things
Classic supplies obfuscated financial features and targets. Signals participants bring stock predictions built with their own data; Crypto has its own market universe and scoring setup. A model slot is an identifier for submissions, and its underlying strategy can change.
Numerai combines submissions into a meta-model, then uses that information in a separate portfolio process. Stake weights influence the prediction aggregate. They do not give a staker ownership of securities, fund assets, or Numerai. The SEC holdings page reports a different layer of this system.
Each atomic round needs its own principal
Classic’s observed atomic boundary is round 1343, opened August 28, 2026. The round principal fell from about 796,000 to 12,500 NMR across the boundary. That change alone is not evidence that investors withdrew almost all their capital.
The new unit is a separately funded (round, staker, model) position. Many positions remain open together. For participation counts, any positive principal qualifies: a fixed 1 NMR threshold would exclude most fractional positions at this boundary.
The round configuration determines what pays
Classic round 1343 specifies CORR60 and MMC60, with multipliers 3 and 15. CORR60 measures correlation to the 60-day target. MMC60 measures contribution after accounting for the meta-model. BMC means Benchmark Model Contribution; its presence in an API response does not make it a payout score.
For this atomic configuration, position return is clip(3 × CORR60 + 15 × MMC60, −1, 1). Multiply by that position’s principal to get its net NMR outcome. The coefficients, target and clip must come from the relevant round, not a formula copied across tournament history. See the metric guide.
Capital moves through several states
Idle strategy funds can finance a new position. Active principal remains committed until resolution. A confirmed claim can release principal and its net outcome. Constant and compound allocation modes treat released gains differently; neither allows the same locked token to fund every overlapping position.
Automated restaking can carry a loss as deferred burn debt for the same model. That debt cannot finance another stake or be withdrawn. Direct claims can finalize the burn. A negative economic outcome and the time tokens leave supply therefore belong in separate measures. Numerai’s atomic staking documentation describes these states and claim rules.
What the dashboard should communicate
Read current balances as capital, model counts as participation, scores as prediction diagnostics, and confirmed rewards less losses as economic outcomes. Treat a scheduled round end as a schedule, not a receipt timestamp. I would judge the migration by reconciled balances and model participation before interpreting a fall in per-round principal as lost confidence.
Figure methods: public Numerai responses captured September 8, 2026; cutover counts include every positive stake. The recycling figure is hypothetical, assumes two 10 NMR positions earning 10% per complete cycle, and excludes fees and losses.