Open maxupp opened 4 years ago
Current implementation does not provide this out of the box. One needs to implement that functionality.
I have analyzed your message once more and I am not sure what you are trying to achieve. There are two layers of fitting: Level 1: choosing whether to use trend/dumping/arma/boxcox ... (for example trend=True, arma=False) Level 2: finding optimal values for parameters for case(s) from level 1 (for example alpha=0.9, p=1)
If you only need the level 1 then have a look at examples/detailed_tbats.py, lines displaying components.* values. You can use those directly to construct particular case of TBATS (after level 1 optimization).
Is this what you need?
Yes that's part of what I want to do. Optimally I would also like to have level 2.
Thanks for bringing up this discussion :) I have a similar use case, where I let tbats do level 1 & level 2 jobs and saved the fitted_model, then I re-used the fitted_model when new data came in - according to https://github.com/intive-DataScience/tbats/blob/master/examples/re_fit_model.py , and the comments " # Re-calculate model for new observation, it will not change model parameters", my understand is, re-fitting the model with new observations will calculate y_hat for each of the new data point and make steps of forecast, but it will not redo level 1 and level 2. So the model config and parameters will stay the same as fitted_model, is my understanding correct?
Yes. It will keep parameters the same. It only recalculates model state to be able to predict for next days after new observations.
I'm looking for the ability to extract the parameters from the fit tbats estimator.
I would then like to recreate this model and fit it directly without evaluating all the other candidates. The background is that I am building a forecasting server where model selection should only be done once, and every time after that, the user provides the parameters that he was originally served.