Question: Probabilistic model about autots

[nsankar]

@winedarksea Thank you for the detailed explanation and guidance. I will assimilate this and explore . Perhaps, It might be a nice feature addition for probabilistic forecasting if NGBoost can be integrated as a part of AutoTS .https://stanfordmlgroup.github.io/projects/ngboost/

[winedarksea]

All models should give 'reasonable' percentile forecasts - you would set prediction_interval=0.5 to get the 25th/75th you asked for, then retrieve the .upper_forecast and lower_forecast from results.
Models are, however, grouped into two groups, those which are "probabilistic" and those which are not. Probabilistic models (you can see which models are in the table at the end of extended_tutorial.md or as you mentioned, you can use model_list="probabilistic") are those models where the probabilities are generated by the model uncertainty - exactly the sort of probabilistic behavior most people are expecting.
The second group of models which are not probabilistic, are those where the uncertainty interval is generated based on the variability of the data and not the variability of the model parameters. These tend to generate slightly wider intervals, but in my experience are still usually reasonable. In fact, these data-driven intervals are sometimes better than the model driven intervals, but being more unusual, I figured I should warn users about them.
If you care about having highly accurate forecast intervals, you can use the metrics SPL and Containment in your metric-weighting, both of which evaluate how well those upper and lower forecasts do over the cross validation. SPL is better for optimization, while containment is a bit more intuitive to human review.
Finally, the tensorflow-probabilistic models you mentioned are experiment (again, see the extended_tutorial.md table for which are experimental). I had a problem with them crashing on some GPUs (something to do with the CUDA configuration? I have never gotten around to exploring it in detail).