Is it somehow possible to use

max_horizon

arg in DistributedMlForecast? Or only through some workaround?

Hey, it's not currently possible

Rough idea - something like Mlforecast.preprocess to generate df with target columns and then just Distributed.fit in a loop should work right?

Yes, but we'd have to add the max_horizon argument to the distributed preprocess and update the output schema to consider the case when there are several targets for that to work. Then we could add the for loop to fit. It's a lot of work and if you have enough data to require distributed training it'll probably take a lot of time to train. Are you certain that the direct approach works better for your use case?