How can we perform autoML per unique id, so not a global one. Is that possible?

Sure, just write a loop that iterates over all your unique_ids. Note sure why you'd want that, though, it's then maybe easier to just use a StatsForecast model that creates a unique model for every timeseries. Optimizing per series is a bit weird when using a model that ingests all unique series and of which the accuracy benefit comes from pooling unique_ids in the same training set (such as MLForecast models, or NeuralForecast models)

Thanks. Does AutoARIMA fit a unique model per

unique_id

? Would a general model perform better? I was thinking about optimizing per unique series to improve accuracy.

Yes. Yes. If the time series exhibit any similarity, I'd always pick a general model, these perform better.

Since each

unique_id

has somewhat different dynamics, is there another solution you’d recommend to capture these differences while still benefiting from a general model? I am trying to do a comparative study against the statsforecast, that is why I am asking.

A properly trained general model should already capture these differences and should outperform statsforecast. E.g. a well engineered LightGBM

Even when a lot of unique ID have very different patterns?

Yes

Hi, what do you mean by 'well-engineered' LightGBM?

LightGBM with good features, good hyperparameters, good cross-validation setup

With good features, would you also recommend including unique id specific variables as dummy variables? So for example, age, gender etc.

Sure, just try out and see what it does to performance.