Numerai is a quant hedge fund built on thousands of crowdsourced machine learning models.

The Numerai dataset consists of standard financial market data for a universe of roughly 5,000 global equities. We assemble it from several market data providers and consistently research both traditional and unique features in data that's available over a 10-20 year timeframe.

Numerai gives away data so that users around the world have free, hedge-fund quality data to build their machine learning models -- data they would otherwise not have access to. From inception, Numerai has believed that open participation, in the form of readily available data, is a core component in crowdsourcing solutions to significant data science problems.

Finding a way to share data without revealing it, a solution that allows us to freely distribute a high-quality set of financial data, was also an early, core innovation that Numerai built. We needed a way to secure the value of the data set while simultaneously retaining its unique features for the purpose of building predictive models with machine learning.

Numerai data has significantly expanded over several iterations: the original data set contained just 40 features and a single target variable. In subsequent releases, the feature count expanded to 310, then to 1100 and now sits at more than 2,000 features. We also now offer multiple targets on which to train your models. More data means better model performance.

The ability for Numerai tournament participants to stake on the predictive accuracy of their models is another core innovation we designed. "Skin in the game" both rewards accurate predictions and penalizes inaccurate submissions. Effectively, it is a way to align incentives and engender trust between Numerai and those submitting their models.

Staking is done with NMR, a crypto-token Numerai launched in 2017. NMR trades freely on several of the major crypto exchanges (Coinbase, Binance, etc.) and its utility is solely to stake on models in the Numerai tournament. Based on a scoring metric, users receive a fractional payout on their stake for performant models, with a portion of the stake being burned (destroyed) for poorly performing models. Stake payouts and burns occur on a blockchain with no direct financial benefit accruing to Numerai.

The Numerai data tournament occurs daily; Participants must submit ranked signals for all ~5,000 global equities in the Numerai data set, forecasting performance over a 20 day (four week) period. Only individual stock predictions, and not any actual models nor code are submitted.

While staking is not required, only staked submissions will be incorporated into Numerai's Meta Model which ultimately manages the trading for and the portfolios in the Numerai hedge fund.

As of July 2023, approximately 5,400 staked models have been submitted for the most recent data tournament, with a staked amount totaling $13 million.

The Numerai Meta Model is an ensemble of the individual models that are submitted as part of the Numerai data tournament. Crowdsourcing predictive models and then combining those signals at scale is another unique innovation that Numerai has developed.

The concept of a performant meta model is predicated on the idea that a broad and diverse selection of uncorrelated models will more accurately predict the outcome of a complex data science problem. Particularly, in the case of predicting financial markets where high accuracy and low variance (a high Sharpe ratio) matter more, this kind of ensembling exercise is doubly helpful.

Seven and half years of iterating on this concept of meta model supremacy (better than simple regression, better than an individual handful of correlated machine learning models) has proven out. Numerai's ability to design the right kind of targets in our data set and our ability to neutralize the impact of traditional risk factors have also contributed to the growth in performance of the Numerai hedge fund over the past several years.

The Numerai hedge fund consists of equally weighted long and short portfolios with approximately 750 individual equities in each portfolio. The beta of the combined long/short portfolio is close to zero, uncorrelated to broad market movements and also not exposed to traditional market risk factors such as country, sector or market cap groupings.

Through several years of iterating on things like data, staking and meta model composition, Numerai has been able to consistently construct a market portfolio where 80%+ of the return is derived from true idiosyncratic (stock selection) risk, rather than the usual factors that drive returns. During times of severe market dislocation like we saw in 2020, a beta-neutral portfolio that's also not exposed to traditional factors will significantly outperform both the market and other purported "market neutral" funds.