Building models for the Numerai tournament

Many people who learn machine learning want a hard problem with real data behind it. Numerai runs a tournament that gives anyone a free dataset of stock market data to try.

We are a small research project. We build our own models from scratch, on our own servers, and we are entering them in the tournament. This page sets out how the tournament works and how researchers test a model before they trust it.

What the tournament is

Numerai calls it a data science tournament. Numerai gives out the dataset free. Participants build models on it and send in predictions, and Numerai scores those predictions over time.

Numerai also runs a hedge fund. Numerai combines the staked predictions into one ensemble, which they call the Meta Model. The fund trades on that ensemble.

What the data holds

Each row in the dataset is one stock on one date. Numerai describes the stock with features known on that date. Its target records how the stock went on to perform against the rest of the market. Numerai builds the target this way because its fund looks for returns that broad market moves leave unexplained.

In version 6.0 the main target looks 60 business days ahead. The training file runs to 28 December 2018, and the validation file starts on 4 January 2019. Most of the new features come in five bins, numbered 0 to 4. Numerai publishes the files in Parquet, a format built for large tables of columns.

Numerai released version 6.0 on 16 October 2026. Participants can send in predictions made with it from the round that starts on 6 November 2026.

Testing a model fairly

Numerai groups the rows by date, and calls each group an era. Its guide advises researchers to treat each era as one data point. Targets on nearby dates can cover overlapping future periods, so a test needs care.

Numerai trained its earlier benchmark models only on data from before the dates they predicted. For a 60-day target, its staff left a gap of 16 eras between the training data and each test period.

The scikit-learn guide calls it overfitting when a model scores well on data it has already seen and fails on new data. For data in time order, the guide recommends tests that always come after the training dates. Numerai gives a similar warning. Researchers who tune a model too often against the validation years can end up fitting it to those years.

How the scores work

Numerai releases new live data every day from Tuesday to Saturday. A participant sends in one prediction for each stock in it. Numerai scores the order of the predictions, so only their ranks count.

The two main scores are correlation with the target (CORR) and contribution to the Meta Model (MMC). MMC measures how much the unique part of a prediction adds to the Meta Model. A round scored on a 60-day target takes about 12 weeks to resolve. Numerai keeps every score on a model's record for good.

Staking and the NMR token

Participants can stake NMR, Numerai's own token, on their predictions. Numerai pays out more NMR for a positive score. For a negative score it burns part of the stake, which destroys it. Numerai scores unstaked predictions in the same way, and only staked ones carry weight in the Meta Model.