Hi, actually, when I mentioned “returns” I meant exactly this kind of data. It works slightly better with loss function that cares more about the sign of prediction (comparing to MSE), but still is not good enough. I also want to compare results with GARCH model later.