Recursive Time Series Forecast, Multi-Step Forecasting

[jdalonsos]

Hello! Is there a way to perform multi-step forecasting? Specifically, recursive forecasting with features. For t+1, I use the features from the training data, but for t+2, in the rolling window, I include the forecast for t+1 along with the other lag features (w). Then, I iteratively recalculate the features for t+h

[nils-braun]

Hi @jdalonsos !
Interesting idea - there is no built-in feature for this and I do not know if this was tried before (at least not to my knowledge). Whether this works depends on your forecasting accuracy I guess. Also be careful how sensitive the feature extraction can be when the values are off - so that you do not propagate a small error to a larger one. Maybe compare the extracted features from your forecasted value with the extracted features from the original value. Even if the difference between the forecast and the actual value is small, some feature calculators might be very sensitive to this.

I would be interested in the results!