Is it possible to use a different forecasthorizon without retraining

[ams015]

I am looking at models such as TFT, PatchTST, DeepAR, NBeats, etc. And it see that in fitting the model, h is somewhat we need

  nf = NeuralForecast(models=[TFT(h=20, hidden_size=int(row['config/hidden_size']),n_head =row['config/n_head'] 
            scaler_type=  row['config/scaler_type'] ,    learning_rate=row['config/learning_rate'], max_steps=row['config/max_steps'],  batch_size  =int(row['config/batch_size']   )       ,
            windows_batch_size=int(row['config/windows_batch_size']), loss=MAE(),random_seed =row['config/random_seed'],input_size=row['config/input_size']  ,    step_size=row['config/step_size']   ,
            valid_loss=MAE(),    val_check_steps=50, early_stop_patience_steps=10,enable_progress_bar=True

            )
  nf.fit(df=train_diff, val_size=20)
  Y_hat_df = nf.predict()

My question is, if I want to forecast with a different h, do I need to retrain the model with that new h? Or is there a way to forecast without retraining?

[jmoralez]

Hey @ams015, thanks for using neuralforecast. It's necessary to retrain because the model architecture is defined based on the horizon.

[tibeler]

why?in the recurssive way like lstm, we can obtain y_hat(t+1)based on y_hat(t). so when the horizon = 1 i can make another prediction based on the first prediction,so it no longer have to retrain?