created a new udf for multivariate

[IshSiva]

time series forecasting

[xzdandy]

Hi Ish, what is the difference between the forecasting in this PR and existing forecasting in EvaDB?

[IshSiva]

Hi Ish, what is the difference between the forecasting in this PR and existing forecasting in EvaDB?

Hi Andy, this is for multivariate forecasting. The previous PR works for univariate forecasting only.

[xzdandy]

Hi Ish, what is the difference between the forecasting in this PR and existing forecasting in EvaDB?

Hi Andy, this is for multivariate forecasting. The previous PR works for univariate forecasting only.

Thanks Ish! Does multivariate subsume univariate? If so, we can direclty change the existing function instead of creating a new one.

[IshSiva]

Hi Ish, what is the difference between the forecasting in this PR and existing forecasting in EvaDB?

Hi Andy, this is for multivariate forecasting. The previous PR works for univariate forecasting only.

Thanks Ish! Does multivariate subsume univariate? If so, we can direclty change the existing function instead of creating a new one.

Hi Andy, Prophet works for univariate also. I think we can update to having a single function.
CREATE FUNCTION MultivariateForecast IMPL 'evadb/functions/multivariate_forecast.py' train_end_date '2005-02-07' type 'MULTIVARIATE' model 'PROPHET' horizon 365 frequency 'd' ;

Will this syntax work?
we can have default parameters for univariate (ARIMA) and multivariate (PROPHET) types.

@americast can you also share your thoughts on this

[americast]

Hi @IshSiva, thanks for the PR. I think for multivariate forecasting, the user would have to also provide multiple columns that need to be considered for forecasting. Thus, the user would have to provide the target column, as well as columns that would be considered as additional regressors.

You can check out statsforecast's implementation at this link. If you would like to directly use prophet, it's even simpler as they have an add_regressor option..

[IshSiva]

this link

Hi @americast , I have used this statsforecast implementation of Prophet for the multivariate forecasting. The target column can be given as an additional argument in the create udf function. So can we just update this new function to support univariate also? (because it is just an extension).

[americast]

@IshSiva I think we can give statsforecast's Prophet Holiday a try. The only change that might be needed is support for more columns as user input while creating the UDF that would act as a reference for forecasting. The model option is also a parameter already implemented (I think I missed adding it to the docs), so the user can simply choose AutoARIMAProphet if they need multivariate forecasting. Or, we can create an alias for the same.

[jarulraj]

Let's go with just statsforecast and neuralforecast.