Nixtla's Open Source Time Series Ecosystem.

Nixtla Community

Hey. The AutoARIMA is AutoSARIMAX if you provide season_length and exog

<@U03UX9KR75W> guess it can only handle one seasonality as standard SARIMA does? Can MSTL handle exogenous?

Yes, only one seasonality. The MSTL uses a separate model to forecast the trend, so if you use ARIMA as that model it can handle exogenous

We just released 1.7.4 which has the fixes for the TBATS, so you should get a better performance with it now

I checked out your nb <@U03L39VLJBA> and I have two recommendations:

• The data has negative values, so the Box-Cox transformation of the AutoTBATS will be disabled. If you want to enable it, you can use a transformation, for example, add the absolute value of the minimum plus 1 and then use the AutoTBATS. 
• I saw that in your nb you only used 96 as seasonal period, but using 96 and 96*7 seems to yield better results. 

Yes <@U03EP6RE8JU> I am literally adding second seasonality now, thanks.

Screenshot 2024-04-09 at 12.13.50.png

<@U03EP6RE8JU> <@U03UX9KR75W> just to let know some additional warnings when running new version

Screenshot 2024-04-09 at 12.24.46.png

<@U03UX9KR75W> is there an example how I can set up trend forecaster with exogenous variables please?

<@U03EP6RE8JU> I have now added 96*7 which is weekly seasonality, the question I have is about annual seasonality as arguably temperature variation over the year is larger than during the week. 

Can triple seasonality be handled in the TBATS implementation, will passing 2922 for monthly one be correct? 

From what I recollect the mathematical version of TBATS can handle multiple seasonalities. Does implementation reflect this?

Hi <@U03L39VLJBA>, yes StatsForecast's version of TBATS can handle N seasonal periods. I'm not sure where the 2922 came from (maybe it was 96*31=2976?). What we can't handle yet is fractional seasonal periods (for example, 365.25 for yearly)

Screenshot 2024-04-09 at 19.00.08.png

here's how it looks with the 3 seasonal periods and the data transformation so that everything is positive

<@U03EP6RE8JU> from fractional year, I didn’t realise TBATs can’t handle fractional year yet (maybe new feature request?) I will try based on full year.

The picture looks like a great forecast, thank you. I will add 3rd seasonality.

yes, that will be a feature soon, its just that we need to think more carefully how to select the Fourier terms when the seasonal period is not an integer.