It's actually my understanding that the core of Medallion is the data processing engine. The models themselves aren't that much more advanced than everybody else (iirc). Rather, financial data is extraordinarily noisy and often not i.i.d. and thus things like deep networks, applied alone, will drown in noise. The use of techniques from advanced statistical signal processing, information theory, conpressed sensing, information geometry, etc. allows the data processing engine to automatically and autonomously extract this signal from a sea of noise.
Ya, that's likely correct. It's all about how you transform the data you feed into the model, and which data you choose to look at, not to mention how you group companies together. That part is huge too. Once you've done all that correctly, off the shelf ML models will solve your problem just fine.
