[Bug]: NaN-filled predict from ETSModel operation on metric evaluation #1285
Description
Expected Behavior
Passed integration test:
FEDOT/test/integration/real_applications/test_examples.py
Lines 86 to 88 in 0bdece1
Current Behavior
Predictions on the metric evaluation process of a ETSModel end up being NaN-containing:
Possible Solution
Maybe with a multiplicative trend these lines in statsmodels are raising some kind of exception (e.g. zero-devision) with a small enough endog values.
Steps to Reproduce
You can add the following code in api_forecasting.py:
def run_ts_forecasting_example(dataset='australia', horizon: int = 30, timeout: float = None,
visualization=False, validation_blocks=2, with_tuning=True):
train_data, test_data, label = get_ts_data(dataset, horizon, validation_blocks=validation_blocks)
# init model for the time series forecasting
pipeline = Pipeline().load('<PATH_TO_PIPELINE>')
model = Fedot(problem='ts_forecasting',
task_params=Task(TaskTypesEnum.ts_forecasting,
TsForecastingParams(forecast_length=horizon)).task_params,
timeout=timeout,
n_jobs=-1,
metric='mae',
with_tuning=with_tuning.
initial_assumption=pipeline)
...Here is one of a troubled pipelines: 0_pipeline_saved.zip