How to harmonize comparison of local and global models?

[jmbejara]

Right now the local models allow retraining at every step of the forecasting whereas global models do not

[jmbejara]

• One easy solution is simply allow retraining at every step. This is likely slow and expensive.
• Another solution is to allow only one epoch of retraining at each new step, while starting it with a hot start of 100 epochs for the first step.
• Another solution is to figure out how to restrict the local models to only use the fit at the training data.

[thearshkumar]

I checked online to find a way to get Local models to start predicting from the series we provide.

Turns out the models below are not able to fit to the specs, but the other local models can still be provided a series to predict off of.

ExponentialSmoothing
Simple ExponentialSmoothing
Theta
Prophet
Naive models that we're using*

I tried using sktime for exponential smoothing and even that didn't allow to use a pre-trained model to predict using some other time stamp which is not part of the training series. This leads me to the conclusion that exponential smoothing would not be supported anywhere due to the intrinsics of how we train and infer from the models.

We discussed Auto-regressive models and those(like ARIMA) have an option to predict from any series provided.

This is making me think that either we should switch up our strategy on how we get predictions, or we can remove these 4 models and keep the current methods.

*Although the Naive models don't support this, I think it is still reasonable to just retrain them(or predict using synthetic data) and that would still be relevant because they're really simple models that are taking values from the past to predict.