Hi all,
I would like to use my prediction from one model as a target in the next model, but I got this error:
Any idea what is going wrong?
What are the use cases for knowledge distillation under the current system?
I think we can't get the coefficients from linear models
This is coming.
This is not supported, sorry.
Ok that explains why it does not work. I will try it out on my own system and see the merit. I wanted to predict my error and use feature importance to see if it was a way of identifying where the model is better. Could be a nice way of saying my model is great in mid caps in the X sector, but really bad at micro caps in the Y sector for example. Or great just after earnings, but horrible just before.
So you want to stack your models?
Yes, probably not more than one level as I imagine there are diminishing benefits.
I want to look into other ways of determining confidence level in the predictions, but this seemed like a simple way to start. Maybe I will try it and it will be worthless, or maybe it will help me find a great sub-universe where my predictions have much lower error.
What about custom stock factors? I am getting an error on these as well:
Invalid formula, ensure you're using future data.
You will need to use FHist("$$factor", -weeks) with your stock factor to use it as a target, as the system needs to see NA values from the target formula for at least one week to consider it valid and to determine lookahead. That means the stock factor itself should not have future data, and FHist should express exactly how far to shift it.
1 Like