It would be easy to match an obfuscated stock market dataset with some third party dataset, and this has happened on many Kaggle competitions (data leaks). That's why the encryption here is important.

But how do you encrypt a stock's historical performance without removing the information (performs better in summer, went up after 9/11…) hidden inside it?

You can add noise, but I doubt that will be enough.

See https://en.wikipedia.org/wiki/Homomorphic_encryption, which they claim to use.

There's no way in hell they're doing ML on cipher text. It's like orders upon orders of magnitudes too slow.

The model has no idea what stock it is predicting. It is just a random ID that represents a security. Further, there are no encrypted charts and no ability to backtest your model outside of the small amount of data provided by Numerai. Download the csv's. They are much smaller than I expected.