Multivariate Time Series

[asifmustafa87]

Hi,

I want to know like if Nixtla supports Multivariate Time Series Forecasting. If so, where can I find the workout examples?

[kdgutier]

Hi @asifmustafa87,

1. All NeuralForecast's models use cross-learning optimization, you can think of them as multivariate/global models.

2. Additionally you can use the exogenous variables inputs to directly leverage predictive correlations across series, here is an exogenous variables example notebook.

3. We are still working to allow the network's outputs to accept multivariate distributions, or multivariate losses.

4. Finally, we are already modeling the forecast horizon as a multivariate regression problem. As we tend to prefer the multistep forecasting strategy over the recurrent strategy due to its computational efficiency and accuracy.

Let me know if this helps.

[candalfigomoro]

Hi @kdgutier ,

in Darts you can use multivariate timeseries so that "each dimension consumes the (lagged) values of all other dimensions".

Is there an easy way to do this in neuralforecast (as in Darts), or do we have to do it manually by adding the other time series as exogenous variables to each time series?

Thanks

[kdgutier]

Hi @candalfigomoro,

Thanks for using NeuralForecast. We are currently working towards a full-multivariate model if you want to try it check this notebook:

• https://github.com/Nixtla/neuralforecast/blob/main/nbs/models.stemgnn.ipynb

In my experience unless the dataset has plenty of Granger-causal relationships a Vector Autoregressive approach tends to make the models unnecessary complicated, as the models tries to learn noisy cross-series relationships. I would definitely compare the predictions of the full-multivariate model to one where you manually feed the model the exogenous variables through the stat_exog_list, hist_exog_list, futr_exog_list parameters.