In order to achieve scalable and robust time series forecast models, I am currently experimenting with metalearner ensembles. Note, that I am also using a global modeling approach, so all time series are "learning" from another. In my example I want to predict the monthly demand for 12 retail products one year ahead (4 years of training data available) I also choose different datasets to test the following.
As base models, I am fitting 6 XGBoost Models with 6 different learning rates from 0.001 to 0.65. The other parameters I do not tune (parsnip library defaults). For the metalearner I am also using an XGBoost Model with the following tuning grid:
Note, that I choose a very small range for the eta parameter!
The grid search resulted in mtry = 16, min_n = 13, tree_depth = 7 and a learning rate of 0.262. Trees were set to 1000 and early stopping parameter was set to 50 iterations.
As you can see, the predicted line of the ex post (training forecast on out-of-sample test set) almost perfectly matches with the actual values. Also there is an almost perfect bias, variance tradeoff. This seems odd, because it should be really hard to achieve in real world problems.
Now I am asking myself, if this approach is just exactly what I need, or if this