Hi everyone. I have a general question (I believe more in the computational side, less of the modelling aspect). Imagine a scenario that I have a timeseries (with exogenous covariates), and I am able to successfully train a model (given a set of metrics) for forecasting a given horizon. How much of a problem would it be using that same model for “backcasting”, in complement of forecasting? For simplicity, I consider that conditions of the model training to be the same in previous time (when compared to the timeseries sample used for training).

Hey. Are you using a specific library (statsforecast, neuralforecast)?

Hi @José Morales! I Intend to use #C05CAFFR22H, #C032L838318 and #C031M8RLC66, if all three were able to do that. Basically is an Earth Sciences modelling work, and I would like to benchmark on ARIMA, XGBoost and N-BEATS/N-HiTS if possible.

And do you want predictions on the training set or a different one in the same training period?

In a sense is the same “out-of-sample” that we have when doing regular forecasting, but backwards, in time. So have a time interval used for training/testing, and a separate validation set to check. In my thoughts I assume that the model captures what is possible given the time span of the training data.

So training on the training dataset and then backtesting on the validation set?

Yes. Probably there will be a validation test for future, and another for the past (considering model training on training + testing partitions, in which testing would be a future one)

Sorry, I'm not quite following. Is it something like this?

Sorry. Not much like this. It is more like a scenario like this (on time span): Training data (possibly used in a cross validation): 1980-2010 Validation Data 1: 2010-2020 Validation data 2 (the “backcasting scenario” ) : 1970-1980

I think you could use the same approach as in that guide, since it's predicting on a "different dataset", i.e. not training and not forecasting

Ok. And this would only be possible with the MLForecast, correct? Because the Statsforecast and neuralforecast would need some initialization that is dependent in temporal direction?

In neuralforecast is similar: guide

Ok. I will look carefully in all this documentation. Thanks a lot for bearing with me 🙂

Feel free to ask more questions if you get stuck