Hello Nixtla Team, I hope you are doing well. I would like to find out if any of the following models support cross-learning when multiple time series are passed in long format: NBEATSx, NHITs, TSMixerx, TIDE, BiTCN, LSTM, and RNN. If they do, could you please let me know how to disable that feature? Is there a parameter or any mechanism that can help to turn down that feature? Thank you! Best regards, Sapna

What is cross-learning?

Let's say you input 100 unique time series in a long format to train a model and generate forecasts. Will the model be able to learn the global patterns across the different time series and then generate forecasts based on those patterns, or will it only learn the seasonality and trend specific to each individual time series?

It will learn a global model. Only statsforecast will learn one model for each unique ide

Is there a way to set it off using any parameters? I don’t want the model to cross learn as each unique time series is different from one another.

then you must train one model per unique id. but this is very uncommon in machine learning approaches. if you want to focus on individual patterns, you might want to try statsforecast, which learns one model for each unique_id. there are all sort of models available, like arima, ces,

I want to use DL models and want to pass exogenous variables. With statforecasts there are very limited options where I can pass exog variables

then you have to train one model for each unique_id 🙂

Got it!

Note that it kind of doesn't make sense to train a NN per time series, these models are in general too expressive to be fit on just a single series. Just train a univariate model such as NHITS or BiTCN, that will usually give much better performance than a model per time series.

It makes sense. I would like to use a model that can capture the complex and non-obvious patterns in the data, and neural network models are the best candidates for this purpose. Additionally, the time series I am dealing with is heavily dependent on external features, so univariate forecasting will not be effective in this situation.

I think we can still learn a lot here. Keep us updated, please.

Ok - note that univariate vs multivariate forecasting and ability to use exogenous features are two different concepts. E.g.: • NBEATS is a univariate forecasting model that can't handle exogenous features; • NHITS is a univariate forecasting model that can handle exogenous features; • TSMixer is a multivariate forecasting model that can't handle exogenous features; • TSMixerx is a multivariate model that can handle exogenous features.

Yeah. I am using only those models that support exog variables.

Where are these series coming from? What do they show? You can add encode the ids so the model can generate different patterns for the different series even if you train them in one big global model.

Sorry I don’t have the liberty to share much details here. How can I encode the ids? Can you explain with an example?

So lets say you have regions of a country and each region resembles some specific patterns. But still there are some similarities because they are all from the same country (e.g.: same public holidays, same or similar education and purchasing power). Then you can encode them into numeric values. The easiest would be one-hot-encoding, where you create a new column for each region and you encode them as 1/0 values (so 1 if it is that specific region and 0 if not - just search for one hot encoding). Then the model can learn patterns that all regions might have in common, but also has the possibility to learn very region specific patterns. Like this you often can combine the power of a global model (learning from all regions), while it still has the flexibility to adapt based on the individual regions. but one hot encoding can be very small if you have a lot of regions. then you might want to try other ways of encoding categorical information like target encoding.