covariate and forecasting target

[rakshitha123]

covariate and forecasting target

[kashif]

thanks @rakshitha123 I'll have a look, I don't have my own data but wish to use your repository for my models... the dataset convention I have is a bit more verbose in defining the covariates as well as categorical features explicitly and I am trying to convert from the tsf format to my convention...

[rakshitha123]

For our current benchmarking model implementations in this repo, we have considered all series as target series, not covariates.

However, you can use the name of the series (series_name attribute) to distinguish between target series and covariate series. Within benchmarking functions, then you can input the series into models as targets and covariates separetly. Although our .tsf format supports the inclusion of covariate series, the models in our repo do not have this implementation as our focus was to run the benchmarks in their most simple format without covariates.

[kashif]

right, so my question was how does the tsf format supports covariates? will the series_name have an fcst_ at the start to distinguish it as a target? For example in the temperature_rain dataset the series_name is just a unique name for each time series... while the obs_or_fcst column has fields with names that seem to encode if they are forecast target or covariate... although I am not sure about that either...

[rakshitha123]

Sure, you can use a prefix with a series_name to identify whether it is a covariate series. We have done something similar in one of our latest project available at https://github.com/rakshitha123/SETAR_Trees. If you see a dataset with covariates in the "datasets" folder (e.g. Rossmann with covariates), you'll find covariate series in their starting with prefix such as "Customers", "Open" and "Promo" where the target series start with the prefix "T".