`stream` functionality

[totvi]

I have read that the stream function should be able to update (at least?) the model parameters. However, I'm not able to find models that has this method implemented. Am I wrong, or hasn't this been implemented yet? I can see that there is a stream.ARIMA, but that it is only a comment in the current devel version?

What is the anticipated behaviour of this compared to refit? I imagine that I would have an use-case where the data is streamed over time, and that updating a model with new arriving data should be easy. Say, that I have data from 2025 and fit a model to these data, e.g., ETS. Then, later I get data from 2026. Would stream(model_2025, newdata = df_2026) and refit(model_2025, newdata = df_2026) be different in compute time and/or outcome? And what would be the main difference to model(MODEL, .data = bind_rows(df_2025, df_2026))?

Thanks for a very impressive ecosystem of packages!

[mitchelloharawild]

Thanks for your kind words, yes - that is the general idea behind stream(). As you correctly identified, the only implemented method is for ARIMA currently (the internal data structure for other models need refactoring to support streaming, which is a substantial endeavor). The ARIMA method is commented out because it needs more thorough testing, I'm 90% sure it works correctly from the testing I did when I wrote it.

• refit() takes a model and applies it to a new dataset
• stream() takes a model, and extends its fit with new data

Assume we have three datasets - full (all time), train (first 80%), and test (last 20%).
The following equivalences and differences should apply:

• refit(<mdl train>, full) === stream(<mdl train>, test)
• refit(<mdl full>, full) != stream(<mdl train>, test) (the estimated coefficients differ, but the underlying data matches)
• refit(<mdl train>, test) != stream(<mdl train>, test) (the initial parameters apply differently, and the underlying data lengths differ)

The advantage of stream() is mostly compute time, it removes the need to apply the model across all historical observations to extend the fit by a few new observations.