1) Bullshit marketing where they advertise the usage of "homomorphic encryption". If you've looked at their datasets they are most likely taking some fundamental and pricing factors and then performing min/max scaling. Nothing too crazy... and I think it would be possible to reconstruct which stock each observation corresponded to given some effort.
2) Professional quants care a lot about the way their features are constructed. The type of data you are looking at affects the type of model you would use... e.g. some kind of GARCH process for lagged time series data vs something different for sector betas, etc.
3) Given (1) there is an insider trading opportunity to reconstruct the underlying data and then submit predictions based on unknown or illegally sourced features.