What is Numeraire (NMR)?
👋 New to this? Just start reading at the top — it begins in plain English and gets more detailed as you scroll. Jump to any level:
🟢 The simple version
Plain English — no jargon. Start here.
Numeraire is the token used by Numerai — a hedge fund that crowdsources its stock market predictions from thousands of data scientists worldwide, who stake NMR on their own models and earn or lose it based on how accurately their predictions perform in live markets.
Numerai — the hedge fund with no employees
Numerai is a hedge fund founded by Richard Craib in 2015 with an unusual model: instead of hiring in-house quant analysts, it runs a continuous global data science tournament. Anyone can participate — download Numerai's dataset (stock market data obfuscated to prevent reverse-engineering), train a machine learning model, submit weekly predictions, and stake NMR on your confidence in those predictions.
Numerai aggregates thousands of models into a "meta-model" and trades based on this consensus. Models that perform well earn NMR; models that perform poorly lose the NMR staked. This creates a mechanism where the best analysts in the world are economically incentivised to contribute their best work — and are automatically penalised if their models only perform well on training data (overfitting) but fail on live market data.
Why NMR is a novel use of blockchain
NMR's use case is genuinely distinctive. It is not primarily traded — it is used as a stake. Data scientists lock NMR on their predictions as a credibility signal. The staking creates "skin in the game": you cannot earn NMR rewards without risking NMR. This eliminates the incentive to submit low-effort or random predictions just to participate. The blockchain enforces the stake automatically — no human at Numerai needs to police participation quality.
What NMR is used for
NMR is staked by tournament participants (earned if your model performs, burned if it doesn't). It is also used in Numerai Signals — where participants can stake on predictions for any global stock, not just Numerai's curated dataset. Total supply: 11 million NMR (very small — one of the smallest supplies in crypto).
Is it legal in India?
Yes. NMR is a VDA under Indian law. 30% tax and 1% TDS apply. See India regulation.
🟡 A bit more detail
For when you want to go a little deeper.
How the Numerai tournament works
Each week, Numerai releases an encrypted dataset — real stock market features but with identifiers removed, column names hidden, and values normalised. Participants build machine learning models to predict future stock returns from these features, without knowing which stocks they are predicting. This obfuscation prevents participants from front-running Numerai's trades or using their model insights for personal trading.
After submission, Numerai scores each model over a 20-day live market period. The scoring metric is a combination of correlation with actual returns and MMC (Meta-Model Contribution) — how much your model adds to the ensemble beyond what other models already provide. Models with high MMC that provide unique signals earn disproportionately more than models that merely duplicate existing predictions.
Numerai manages real money using the aggregated predictions from its tournament. The fund's trading strategy is driven entirely by the meta-model consensus. Thousands of data scientists globally contribute models, with participants from 140+ countries. The top earners include professional quants, academic researchers, and self-taught data scientists. Numerai Signals extends this to any global stock market data that participants source themselves — predicting returns on equities not included in Numerai's dataset, with NMR staking on those predictions as well.
The burn mechanism
When a participant's model underperforms, NMR is burned — permanently destroyed, not redistributed to other participants. This deflationary mechanism reduces NMR supply over time. The burn is proportional to the staked amount and the degree of underperformance. A model with catastrophic performance over an extended period can lose its entire stake. This burning mechanism gives NMR a deflationary supply dynamic tied directly to the quality of predictions — better participants win more, worse participants burn their stake, reducing overall supply.
Key Numerai metrics: total NMR staked across the tournament (a proxy for participation quality and confidence), model count, NMR burned per epoch (deflationary health), and fund performance (not publicly disclosed in detail but Numerai has reported positive returns). NMR is one of the few crypto tokens with a genuinely novel and functional use case that has operated continuously since 2017. Live data: CoinGecko · Numer.ai.
🟣 The full technical picture
For the technically curious.
Erasure Protocol and the staking mechanism
Numerai built the Erasure Protocol — an open-source Ethereum protocol for staking on information and burning stakes when predictions are incorrect. NMR staking and burning is implemented via Erasure's smart contracts. The protocol is general-purpose: anyone can build an application where participants stake NMR on claims they make, with an autonomous burning mechanism if those claims are falsified. Numerai's tournament is the primary application of Erasure, but the protocol's architecture allows arbitrary staking use cases.
Source: Craib, R. (2017). Numeraire: A Cryptographic Token for Coordinating Machine Intelligence and Preventing Overfitting. numer.ai/nmr.pdf · Numerai documentation: docs.numer.ai
The meta-model and ensemble aggregation
Numerai's meta-model aggregates thousands of individual model submissions into a single prediction vector used for trading. The aggregation is stake-weighted — models with more NMR staked have higher weight in the ensemble. This creates a mechanism where confident, well-capitalised participants have more influence on Numerai's trades. The meta-model is recomputed weekly as submissions arrive. Individual model performance is evaluated against the actual meta-model's decisions (MMC) as well as raw market correlation — incentivising diversity: a model that perfectly duplicates the meta-model earns zero MMC.
Numerai's data obfuscation uses a combination of feature engineering, normalisation, and identifier removal to prevent reverse-engineering of the underlying stock universe. The target variable (what models are predicting) is a 20-day stock return normalised within each market sector to remove systematic market-wide moves — participants are predicting relative performance within sectors, not absolute returns. This normalisation means models that simply predict "tech stocks go up" earn zero because all tech stocks are normalised against each other. Only genuine relative alpha — predicting which specific stocks outperform their sector — earns returns. This design specifically rewards sophisticated fundamental and quantitative analysis over simple momentum or sector-trend strategies.
Key facts
- Application: Crowdsourced hedge fund predictions
- Staking model: Stake NMR → earn (model performs) or burn (model fails)
- NMR supply: 11 million (very small, deflationary via burns)
- Protocol: Erasure Protocol (open-source staking/burn)
- Tournament: Weekly, 20-day live market evaluation
- Founded: 2015 (Richard Craib)