search

winedarksea / AutoTS

Automated Time Series Forecasting
MIT License
1.1k stars 100 forks source link

Theta Template Eval Error #216

Closed emobs closed 9 months ago

emobs commented 9 months ago

[AutoTS version 0.6.5]

Not sure if this is a true bug, but on all Theta model templates I get this error: Template Eval Error: ValueError('freq T not understood. Please report if you think this is in error.') The frequency of my input data is 5T and set in the model definition, so not automatically determined by 'infer'.

winedarksea commented 9 months ago

Could you print model.used_frequency and post that? I made an update to the inferred frequency process and that might be inferring incorrectly? otherwise could be an issue with the Theta model directly

emobs commented 9 months ago

I think this is what you need: Data frequency is: 5T, used frequency is: 5T If not, please let me know.

winedarksea commented 9 months ago

That's helpful. Means the bug is specific to the Theta model. https://www.statsmodels.org/dev/generated/statsmodels.tsa.forecasting.theta.ThetaModel.html

I have confirmed that if I specify period specifically (as an int) and not use the auto 'None' for periods it will prevent the issue.

But regardless Theta is too slow at scale. I will for now add period as a not None value if models_mode="deep". You could also try the model manually with a not None value for period, seeding that in as a template if you want to try running it now