backtest N-HITS on test dataset

[jimbow973]

Hi,
I made a N-HITS forecaster with historical exogenous series, no problem to do insample predictions and futur_df prediction.

horizon = 1 # minutes forecast
models = [NHITS(h = horizon,
                input_size = 3,
                exclude_insample_y=True,
                loss=HuberLoss(delta=0.5),
                valid_loss=MAE(),
                scaler_type = 'robust',
                hist_exog_list = ['r1','r2','r3','r4','r5','r6','a1','a2','a3','a4','a5','a6'],
                learning_rate=0.001,
                max_steps=2000,
                early_stop_patience_steps=10)] 
nf = NeuralForecast(models=models, freq='T')
nf.fit(df=dftrain,val_size=horizon)

The issue I face is that I am not able to do a backtest: I have a test dataset where exogenous features are known, and I cannot do a prediction for more than "h" rows. How can I do it please?
Or said differently, I'd like to be able to use my model on new data for several horizons before considering further retraining.

I tried this:

forecast = nf.predict(df=dftrain.tail(3),futr_df=dftest.iloc[[0]],verbose=True); works fine for the first horizon.

Then using my test dataframe with all historic exogenous known to forecast 100 steps with h=1, without autoregression on y.

result_df = pd.DataFrame();
for j in range(1,100):
    forecast = nf.predict(df=dftrain.tail(3).combine_first(dftest.iloc[j-1:j]),futr_df=dftest.iloc[[j]],verbose=True);
    if j < 2:
        result_df = forecast;
    else:
        result_df  = pd.concat([result_df,forecast], axis=0);
result_df.describe()

return all the time the near same value (check issue #449, very similar)