0
$\begingroup$

I have a minutely dataset for a year duration. It has a daily seasonality. This would imply a seasonal period of 1440 according to https://robjhyndman.com/hyndsight/seasonal-periods/ .

I thought of using the SARIMA (𝑝,𝑑,𝑞)×(𝑃,𝐷,𝑄)𝑠 model with the following parameters : order : (𝑝,𝑑,𝑞) - (6,0,0) - ACF and PACF hinted at a pure AR process

I have trouble determining the (𝑃,𝐷,𝑄).

𝑠 - 1440

However even for (𝑃,𝐷,𝑄) = (1,0,0), the model.fit takes a really long time.

train size - 216,000 data points.

I suspect it is due to the high 's' value.

Is SARIMA not meant to handle minute level data with daily seasonality?

$\endgroup$
1
$\begingroup$

Well of course you have a computation problem since you have lag of 1440 and you probably use maximum likelihood estimator (conditional) so to initialize it has to solve a 1440 linear system + multiplying more than 1440 matrices by each other ...

So it's quiet impossible to use SARIMA here, the best way would be to go with ARIMAX with the creation of a dummy variable that is 1 every 1440, 0 otherwise.

$\endgroup$
0
$\begingroup$

https://stats.stackexchange.com/search?tab=newest&q=user%3a3382%2016%20minute details how to model data taken 15 minutes apart. This is easily extended to readings taken 1 minute apart . The whole idea is that one can and should use deterministic dummies and develop 60 models by using daily predictions.

$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.