robjhyndman / forecast

Forecasting Functions for Time Series and Linear Models
http://pkg.robjhyndman.com/forecast
1.11k stars 341 forks source link

non-automatic TBATS? #965

Closed smxzehvb closed 2 weeks ago

smxzehvb commented 2 weeks ago

Is there a way to force TBATS to run with a certain configuration? Say, I don't want the function to try and find the best model out of many parameter configurations. I only want a single specific configuration.

e.g. if I want exactly this model, and nothing else tested under the hood:

Is this possible? Thanks.

robjhyndman commented 2 weeks ago

Yes. Please read the documentation. help(tbats) or https://pkg.robjhyndman.com/forecast/reference/tbats.html

smxzehvb commented 2 weeks ago

Ok, yeah, everything I asked for, except fixing an order for the ARMA errors.

So even if you set use.arma.errors=TRUE, could the model still select order = (0,0) under the hood? Would that end up equivalent to use.arma.errors=FALSE? Or is it forced to pick an order > (0,0)?

I'm asking about fine-grain control because I want to pick the combination of TBATS params based on out-of-sample forecast error, rather than in-sample AIC. And I don't want to iterate over inadvertently redundant combinations of parameters. Cheers

robjhyndman commented 2 weeks ago

The automatic selection of the ARMA errors can choose $p=q=0$ if that gives the lowest AICc value. Additional arguments to tbats() are passed on to auto.arima() which gives you some control over the model selected, but not enough to fully specify the model.

smxzehvb commented 2 weeks ago

thanks