Numeraire is a platform designed by Numerai to improve the quality and reliability of predictive models in financial data science competitions. It addresses the problem of overfitting—where models perform well on historical (training) data but fail on unseen (real-world) data—by introducing a staking mechanism tied to model confidence.

Participants in Numerai’s machine learning tournament build predictive models using provided data and submit their predictions without accessing raw datasets. To signal confidence in their models, data scientists stake tokens, and their compensation is based on how well their models perform on unseen data. This creates an incentive for participants to focus on building models that generalise, rather than optimise for the historical dataset.

The platform also introduces a novel mechanism that aligns economic incentives with performance, helping to surface models that are more likely to succeed in real-world market conditions. The end goal is to enhance the investment strategies of Numerai’s hedge fund by integrating crowd-sourced intelligence while reducing susceptibility to data overfitting.

NMR is an Ethereum-based ERC-20 token used as a staking mechanism within the Numerai ecosystem. Its primary functions are:

• Staking on predictions: Data scientists submit predictions and stake NMR tokens to express their confidence in their model’s performance.

• Incentivising performance: Participants with accurate models on unseen data are rewarded in fiat currency and their NMR stakes are returned.

• Penalty for poor performance: If the predictions underperform, the staked NMR may be destroyed (burned), introducing an economic cost to overfitting or poor modelling.

The token is not used to pay fees or access the platform. Instead, it serves as a mechanism to reveal users' confidence in their models and to align incentives across participants. The staking process is governed by a Dutch auction system that adjusts rewards based on performance and relative confidence.