There are NaNs in y_pred (Pipeline.backtest)

[Darya031]

why there is mistake "There are NaNs in y_pred" while I'm doing Pipeline.backtest with NBeatsGenericModel model
(timeseries don't contain NaN values)

here is my code:

metrics = [SMAPE(), MAPE(), MAE()]
model_nbeats_generic = NBeatsGenericModel(
        input_size= HORIZON*2,
        output_size=HORIZON,
        loss="mape",
        stacks=30,
        layers=4,
        layer_size=256,
        trainer_params=dict(max_epochs=1000),
        lr=1e-3,
    )

pipeline_nbeats_generic = Pipeline(
        model=model_nbeats_generic,
        horizon=HORIZON,
        transforms=[])

metrics_nbeats_generic, forecast_nbeats_generic, _ = pipeline_nbeats_generic.backtest(
    ts, metrics=metrics, n_folds=3, n_jobs=1)```

[d-a-bunin]

Hi! How long is the timeseries (number of timestamps) and how long is HORIZON?

[Darya031]

The horizon is 30 days and there are 731 row in timeseries

[d-a-bunin]

• You could run get_historical_forecasts instead of backtest. This way you can inspect the backtest predictions without calculating metrics. Are there any NaNs?
• You could run fit, forecast and check the forecasts of the network. Are there any NaNs?

[Darya031]

There are NaNs if I'm doing get_historical_forecasts instead of backtest. There are also NaNs in predictions when I run fit, forecast and check the forecasts of the network

[d-a-bunin]

Let's dive into making fit, forecast.

• Are NaNs in all prediction timestamps or only in some of them?
• Are there any missing values in target history?
• Are there any missing values in exogenous variables (if there are any)?