What this covers

Patrick interviews Richard Craib, founder of Numerai, about how the hedge fund is built. Numerai crowdsources machine learning work on a carefully specified prediction problem, not stock picks. It publishes its full dataset free, but the roughly 50 features are anonymized. Contributors can fit models without knowing what the axes mean, and the firm inspects only their predictions. Craib says that makes the final signal uninterpretable and that Numerai doesn't mind.

The middle of the conversation is about mechanism design. Craib ties the wisdom of crowds to Crowd Wisdom Conditions: averaging correlated opinions adds nothing, while orthogonal signals add up. An open payout system also has to resist a Sybil attack, so Numerai allocates rewards through a variant of a multi-unit Dutch auction where extra accounts gain nothing. He also argues that making exit easy increases long-run engagement, citing PayPal withdrawals, and that blockchain's real value may be in enabling unusual companies like Numerai more than consumer products like CryptoKitties.

The conversation also covers finance and modeling. Craib stresses that data comes first, and that preprocessing should make the past look like the present. He explains why a move from 52 to 53 percent accuracy can shift the Sharpe ratio by a point or two. The more contentious claims are that most big funds' machine learning teams rarely reach production, that Renaissance Technologies may owe its edge to data more than modeling, and that conventional funds, staffed alike and buying the same data, sit in a poor local optimum. One surprise is that a simple average of staked models is very hard to beat.

Sharpest takeaway

Richard Craib argues that Numerai's edge comes from crowdsourcing machine learning models on obfuscated data and using cryptocurrency staking to filter out low-quality predictions, and that interpretability and conventional hedge fund modeling matter less than data quality and skin in the game.

  • Crowd-sourced models on hidden, anonymized features outperform what a single firm could build in-house, and staking raises Sharpe from about 1.5 to 2.1
  • Staking acts as a cost filter against spam and low-effort predictions, which blockchain escrow makes credible
  • Quant hedge funds may win on proprietary data rather than modeling sophistication

The argument · threads2 threads · 65 claims
0.71

Numerai's crowd-driven process mirrors the standard quant pipeline, where edge comes from data.

2 pointscentrality 4/5
  • The standard quant pipeline of feature engineering, feature selection, hyperparameter tuning, model selection, ensembling, and production maps onto Numerai's process, with the crowd handling the model-building and selection stages.

    “engineering the features themselves, doing whatever transforms you need to, in your case, regularizing them, selecting the features”

  • Patrick observes that elite quant firms' edge may come more from informational, data-based advantages than from analytical modeling, which is the opposite of common assumptions.

    “It's interesting that maybe places with all these PhDs, maybe it's a data advantage, not a modeling one.”

0.29

Balancing both marketplace sides leaves Numerai exposed to competitors offering better payouts.

1 pointcentrality 2/5
  • Building a marketplace requires balancing two sides at once, and the risk for Numerai is that a competitor offering better payouts or service to data scientists could draw them away.

    “I'm obsessed with marketplace businesses in general. I just think it's so interesting when a business has to build two sides of a market effectively”