AutoETS Exception no model able to be fitted

[webert6]

What happened + What you expected to happen

I have am getting this error unexpectedly. This data doesn't look to have any issues with it that would lead me to believe ETS would have issues.

Traceback (most recent call last):
File "/local_disk0/.ephemeral_nfs/cluster_libraries/python/lib/python3.10/site-packages/fugue_spark/execution_engine.py", line 228, in _udf_pandas
output_df = map_func(cursor, input_df)
File "/local_disk0/.ephemeral_nfs/cluster_libraries/python/lib/python3.10/site-packages/fugue/extensions/_builtins/processors.py", line 333, in run
return self.transformer.transform(df)
File "/local_disk0/.ephemeral_nfs/cluster_libraries/python/lib/python3.10/site-packages/fugue/extensions/transformer/convert.py", line 346, in transform
return self._wrapper.run(
File "/local_disk0/.ephemeral_nfs/cluster_libraries/python/lib/python3.10/site-packages/fugue/dataframe/function_wrapper.py", line 103, in run
rt = self._func(**rargs)
File "/local_disk0/.ephemeral_nfs/cluster_libraries/python/lib/python3.10/site-packages/statsforecast/distributed/fugue.py", line 139, in _forecast_noX
_, result = self._forecast(
File "/local_disk0/.ephemeral_nfs/cluster_libraries/python/lib/python3.10/site-packages/statsforecast/distributed/fugue.py", line 110, in _forecast
result = model.forecast(
File "/local_disk0/.ephemeral_nfs/cluster_libraries/python/lib/python3.10/site-packages/statsforecast/core.py", line 939, in forecast
res_fcsts = self.ga.forecast(
File "/local_disk0/.ephemeral_nfs/cluster_libraries/python/lib/python3.10/site-packages/statsforecast/core.py", line 204, in forecast
raise error
File "/local_disk0/.ephemeral_nfs/cluster_libraries/python/lib/python3.10/site-packages/statsforecast/core.py", line 204, in forecast
raise error
File "/local_disk0/.ephemeral_nfs/cluster_libraries/python/lib/python3.10/site-packages/statsforecast/models.py", line 799, in forecast
res = self._add_conformal_intervals(fcst=res, y=y, X=X, level=level)
File "/local_disk0/.ephemeral_nfs/cluster_libraries/python/lib/python3.10/site-packages/statsforecast/models.py", line 167, in _add_conformal_intervals
cs = self._conformity_scores(y, X) if y is not None else self._cs
File "/local_disk0/.ephemeral_nfs/cluster_libraries/python/lib/python3.10/site-packages/statsforecast/models.py", line 153, in _conformity_scores
fcst_window = self.forecast(h=h, y=y_train, X=X_train, X_future=X_test) # type: ignore[attr-defined]
File "/local_disk0/.ephemeral_nfs/cluster_libraries/python/lib/python3.10/site-packages/statsforecast/models.py", line 788, in forecast
mod = ets_f(
File "/local_disk0/.ephemeral_nfs/cluster_libraries/python/lib/python3.10/site-packages/statsforecast/ets.py", line 922, in ets_f
raise Exception("no model able to be fitted")
Exception: no model able to be fitted

Versions / Dependencies

Click to expand

Dependencies: statsforecast==1.7.6 and greater

Reproducible example

data = [
  (1,   "2024-01-01", 1),
  (1, "2023-12-01", 2),
  (1, "2023-11-01", 2),
  (1, "2023-10-01", 7),
  (1, "2023-09-01", 6),
  (1, "2023-08-01", 7),
  (1, "2023-07-01", 12),
  (1, "2023-06-01", 4),
  (1, "2023-05-01", 9),
  (1, "2023-04-01", 2),
  (1, "2023-03-01", 2),
  (1, "2023-02-01", 0),
  (1, "2023-01-01", 1),
  (1, "2022-12-01", 2),
  (1, "2022-11-01", 5),
  (1, "2022-10-01", 6),
  (1, "2022-09-01", 4),
  (1, "2022-08-01", 5),
  (1, "2022-07-01", 4),
  (1, "2022-06-01", 10),
  (1, "2022-05-01", 2),
  (1, "2022-04-01", 2),
  (1, "2022-03-01", 1),
  (1, "2022-02-01", 1),
  (1, "2022-01-01", 1),
  (1, "2021-12-01", 0),
  (1, "2021-11-01", 2),
  (1, "2021-10-01", 4),
  (1, "2021-09-01", 1),
  (1, "2021-08-01", 1),
  (1, "2021-07-01", 1),
  (1, "2021-06-01", 2),
  (1, "2021-05-01", 1),
  (1, "2021-04-01", 3),
  (1, "2021-03-01", 5),
  (1, "2021-02-01", 3),
]


from pyspark.sql import SparkSession
from statsforecast.models import *
from statsforecast.utils import ConformalIntervals
from statsforecast import StatsForecast

# Initialize Spark session
spark = SparkSession.builder \
    .appName("Create DataFrame from Array") \
    .getOrCreate()

# Define the schema
columns = ["unique_id", "ds", "y"]

# Create the DataFrame
df = spark.createDataFrame(data, schema=columns)

horizon = 2

model_params = [AutoETS(season_length= 12, model='AAA', damped=True, prediction_intervals=ConformalIntervals(h=horizon, n_windows=9), alias='AUTO_ETS_AAdA')]

sf = StatsForecast(
    models=model_params,
    freq="MS"
)

forecasts = sf.forecast(
    df=df , 
    h=horizon, 
    level=[50,60,70,80,90,95,99]
)

# to invoke the actual model
forecasts.toPandas()

Issue Severity

Medium: It is a significant difficulty but I can work around it.

[webert6]

by removing the parameter prediction_intervals=ConformalIntervals(h=horizon, n_windows=9) entirely the model works as expected. Conformal intervals are required for my use case tho.

[webert6]

If I change the n_windows parameter in ConformalIntervals() from 9 to something less, it works. From my understanding, the training data needs to be at least h * n_windows + 2 rows long so this must be different for ETS models?

[jmoralez]

Hey. Seasonal models require at least 2 seasonal periods as history (24 samples in your case) and you're forcing it to be a seasonal model, which is most likely why it fails. Can you try setting the seasonal part as auto (model='AAZ')?

[webert6]

Great insight @jmoralez that makes sense. 9 was too many windows. Closing