It is a virtual certainty that Numerai is lying about using homomorphic encryption. The only known algorithms for FHM encryption require specialized algorithms to perform computations on the encrypted data. You can't just run a standard neural net over some homomorphically encrypted data and expect an interpretable result. Yet, Numerai claims that this is exactly what's possible with their data. This is clearly false. They are probably obfuscating their private signals in some extremely trivial way.
Unclear tbh, the only thing I could tell from numerai is that the data is time series.
The evaluation of the predictions isn't public afaik so you can't tell.