How Numerai Works: Tournament, Staking, and the Meta-Model

Numerai is a hedge fund powered by a global tournament of data scientists. Participants download free, obfuscated financial data, build machine learning models, submit predictions, and stake NMR on their confidence. Numerai combines those predictions into a stake-weighted meta-model and trades real capital on it. The structure is unusual enough that one resolved round in 2023 burned 38,174 NMR from staker bonds in a single week while another in 2021 paid out 33,577 NMR — both at the same protocol, governed by the same formula.

This post covers the three tournaments, the data, the round lifecycle, scoring, staking, and how the meta-model is assembled.

The Three Tournaments

Numerai runs three concurrent tournaments, each targeting different markets and data paradigms:

• Classic (Tournament 8) — Numerai provides a fully obfuscated dataset. You build ML models on their features without needing any financial domain knowledge.
• Signals (Tournament 11) — You bring your own data and generate original signals. Numerai neutralizes your signal against its existing ones to extract the unique component.
• Crypto (Tournament 12) — Targets cryptocurrency markets using Numerai's obfuscated crypto data.

The Data

In Classic, each row represents a stock at a point in time. Real financial features (P/E ratios, volume, momentum, and so on) are mathematically transformed and binned into 5 discrete values (0 through 4). Stock IDs are randomized per era, so you cannot track individual stocks across time.

An era is one weekly time slice. Training and validation data spans thousands of eras and is refreshed every few months. Live tournament data is released fresh each round.

You don't need a finance background to compete. Obfuscation lowers the finance-domain barrier, but model quality still depends on validation discipline, feature handling, and tournament-specific scoring.

Round Lifecycle

Each round follows a defined timeline from data release to payout settlement:

1. Open — New round data is released. Participants download it and run their models.
2. Close — Submission deadline. All predictions must be uploaded.
3. Stake Deadline — Last chance to adjust your NMR stake amount.
4. Scoring — Predictions are evaluated against realized stock returns over ~20 business days.
5. Resolved — Final scores are calculated and payouts settle.
6. Finalized — Round is complete and immutable.

Because the target measures roughly 20 business days of returns, each submission takes about a month to fully resolve, even though new rounds open every week and daily rounds run in parallel.

How Scoring Works

Numerai uses four primary metrics to evaluate predictions. Each measures a different dimension of model quality:

MMC (Meta Model Contribution)

The covariance of your predictions with the target after neutralizing to the meta-model. MMC measures how much unique value your model adds beyond the crowd, which is why it's the most important metric — it's your direct contribution to the hedge fund. See MMC vs Correlation for how it compares to raw correlation.

BMC (Bonus Meta Model Contribution)

Same concept as MMC, but neutralized against Numerai's stake-weighted benchmark models instead of the full meta-model. BMC rewards originality relative to Numerai's own baselines.

CORJ60 (Correlation Jerome 60)

Correlation of your predictions against a specific target with a 60-day lookback window and a 2-day data lag. A more direct measure of predictive accuracy.

FNCv3 (Feature Neutral Correlation v3)

Correlation with the target after neutralizing your predictions to Numerai's features. FNCv3 isolates predictive power that isn't just linear feature exposure, rewarding models that capture something the raw features miss.

The Payout Formula

Since January 2024, payouts are calculated as:

Payout Score = 0.5 x CORJ60 + 2.0 x MMC

The 2x multiplier on MMC heavily incentivizes unique, additive predictions over simple correlation with the target.

Staking and Payouts

Staking puts real money behind your predictions, aligning participant and fund incentives.

• Minimum stake: 0.01 NMR
• Maximum payout: +25% of your stake per round
• Maximum burn: -25% of your stake per round
• Payout factor: A per-round multiplier that modulates actual payouts based on tournament conditions, visible on the trends page

Payouts and burns settle after round resolution. Burned NMR is permanently destroyed, not redistributed to other participants.

The Meta-Model

All staked predictions are combined into a single stake-weighted meta-model:

Models with more NMR staked get proportionally more influence. Models that perform well accumulate stake; models that perform poorly burn stake and lose influence. Over time, bad predictors get priced out.

Roughly 4,800 staked models feed the meta-model each round in recent samples, but the top single staker now controls about 13% of total weight on its own — so effective influence is much narrower than the headline count. The leaderboard ranks each one by reputation and returns.

Key Concepts

Obfuscated Features All data is transformed so participants cannot reverse-engineer Numerai's trading universe. This protects Numerai's IP while enabling anyone — regardless of financial background — to compete.

Prediction Neutralization Used in MMC and BMC scoring to remove prediction components already captured by the meta-model or benchmark models. The residual signal is what counts.

Feature Exposure Models with too much linear exposure to known features tend to perform poorly out of sample. FNCv3 specifically penalizes this. The best models capture nonlinear relationships that the raw features miss.

NMR Token

NMR (Numeraire) is an ERC-20 token on Ethereum with a hard cap of 11 million tokens. No additional NMR can be minted; rewards come from existing supply, while burns permanently remove tokens. The token's primary utility is staking in the tournament — see NMR Token Economics for the full supply story.

Why It Matters

Numerai solves a hard problem in quantitative finance: how do you combine work from thousands of independent researchers without them copying each other or gaming the system? Four mechanisms do the work together: obfuscation prevents reverse-engineering of the trading universe, staking forces participants to risk real capital, MMC scoring rewards originality over conformity, and the stake-weighted meta-model aggregates diverse signals into a single trading input.

The result is a hedge fund whose alpha comes from a global tournament. You can track the meta-model's inputs in real time on nmrdash; the trends page has the historical stake and payout data.