Time gap between the lookback period and the forecast period

[EdxDv]

Hi everyone ! we are interested in forecasting H days ahead after a gap of p days, actually what we want to forecast is y[t+p+1:t+p+H](we do not need to forecast y[t+1:t+p]), do you have any ideas on how to adjust the models ? Otherwise, do you think it's a good idea to forecast the total period and just ignore the first p forecasts ?

[kdgutier]

Hey @EdxDv,

• The simplest solution is to continue with a normal NeuralForecast model and filter the predictions to only keep them after the first p horizon steps.
• If you want to explore a simple model customization an interesting possibility would be to hack the horizon mask in the training_step method to make the loss function to only depend on the observations after p. This second solution can be achieved simply by overwriting a NeuralForecast model train_step.

Let us know your thoughts.

[EdxDv]

Thank you so much for your suggestions, we really appreciate your fast reply !

[EdxDv]

Sorry for bothering you again, do you think modifying the outsample_mask in the training_step method before feeding it to the loss function, namely by setting outsample_mask[:,0:p,:]=0 in the BaseRecurrent class and outsample_mask[:,0:p]=0 in the BaseWindows class will do the trick ?

[kdgutier]

Yes, outsample_mask[:,p:,:]=0=>[0,0,0,1,1,1,1,1] would specialize the network on the horizon after the first p steps.
It would be interesting to validate that such strategy helps the accuracy of the later steps.