Managing Time Series Compute Costs

Question

This is a question about how to build time series models with a reasonable compute cost.

I have about 100 items to forecast.

For each of these forecast items, I am building 8 separate models to accommodate limitations in the exogenous data I am using to build the models. So, one model is built to forecast steps 1 to 5 , a second model to forecast steps 6 to 11, etc.

Due to the limited amount of data I have, I'm building SARIMAX models, using brute force grid search to find optimal sets of hyperparameters.

To test individual sets of hyperparameters, I am using a simple holdout data set. I would prefer to use some sort of time series cross validation to get more reliable measures, however that quickly increases my already high computing cost up by maybe 5 to 8 times.

Is there a way to get the benefits of time series cross validation without the enormous computing cost?

Answer 1

Your best bet is to try to reduce the computation for your current framework and then just expand that into the CV time slices.

So you only have your exogenous variable for the first 5 steps and then you don't so you leverage a model without the variables after that point?

If that's the case I would say you can try to just use some average of each exogenous variables or something like that so you can just use one model or potentially forecast for those variables if applicable.

Then you can do certain tests to reduce the parameter space of the SARIMAX orders. If you use python you could look at pmdarima which is based on the auto_arima funtionality in r. For example, a simple test could be to test for seasonality and if there is none then you don't have to worry about the seasonal differences.

Those steps should get you some nice gains in computation cost.