Hi, does AutoMLForecast concern time series features optimization (such as lags) or model parameters as well (learning_rate)?

By default it optimizes both but you can set it to only tune one

In that case, I do not have to give a search grid? And is differencing also done by AutoMLForecast?

There are default search spaces, differences are included in them when possible (series are long enough)

Thanks! Can I find the default search space somewhere?

They're in the auto module, the model ones are fixed but the transforms are defined programatically

And does it also help finding the best combination of exogenous variables?

No, it'll always use the ones you provide. If you want to tune that you'll have to define a custom objective

But don’t tree-based models inherently perform feature selection, such as a RF or a XGB?