[BUG] AutoETS not working on some time series

[config-i1]

Describe the bug

I ran your implementation on M1, M3 and Tourism competition dataset, and found several cases, where the function fails or produces unreasonable forecasts.

Steps/Code to reproduce the bug

Install the package with the datasets

pip install fcompdata

Load the data

from fcompdata import M1, M3

Here is where it fails:

series = M1[83]

model = AeonAutoETS()
model.fit(series.x)
forecast = model.iterative_forecast(series.x, prediction_horizon=series.h)

print(forecast)

[nan nan nan nan nan nan]

And here I get extremely high point forecasts:

series = M3[113]

model = AeonAutoETS()
model.fit(series.x)
forecast = model.iterative_forecast(series.x, prediction_horizon=series.h)

print(forecast)

[7.82501581e+04 3.39696794e+05 2.72884577e+06 5.25038119e+07
3.48929606e+09 1.34675105e+12]

The latter could be due to the multiplicative trend (if it is selected). If that's the case, I would suggest switching it off by default (all ets implementations do that). But even in case of multiplicative trend, other implementations produce more reasonable forecasts. e.g.:

[8819.883 9853.339 10978.959 12197.934 13594.791 15163.362]

Expected results

The first example should produce something like this:

[735.8964 756.7602 777.6240 798.4879 819.3517 840.2156]

The second one:

[8819.883 9853.339 10978.959 12197.934 13594.791 15163.362]

Actual results

In case of the first example, I get:

[nan nan nan nan nan nan]

For the second one:

[7.82501581e+04 3.39696794e+05 2.72884577e+06 5.25038119e+07
3.48929606e+09 1.34675105e+12]

Versions

Details

System: python: 3.13.7 (main, Jan 22 2026, 20:15:57) [GCC 15.2.0] executable: [/home/config/Python/Libraries/smooth/python/.venv/bin/python](https://file+.vscode-resource.vscode-cdn.net/home/config/Python/Libraries/smooth/python/.venv/bin/python) machine: Linux-6.17.0-12-generic-x86_64-with-glibc2.42 Python dependencies: aeon: 1.3.0 pip: 25.1.1 setuptools: 80.10.2 scikit-learn: 1.7.2 numpy: 2.2.6 numba: 0.61.2 scipy: 1.15.3 pandas: 2.3.3

[MatthewMiddlehurst]

@alexbanwell1 @TonyBagnall

[TonyBagnall]

hi @config-i1 sorry for the delay, I'll look at this now

[TonyBagnall]

I can reproduce this

f = AutoETS()
f.fit(y)
print("Horizon =",series.h)
print(f.iterative_forecast(y, prediction_horizon=series.h))

[417 451 360 525 392 375 326 594 519 570 416 556 272 391 732 890 855 737
470]
Horizon = 6
[nan nan nan nan nan nan]

[TonyBagnall]

So I think I might know whats going on. Currently we search meta parameters then refit with chosen parameters. For really short series like this, given the nature of nelder mead, the refit could result in a poor model, and NaNs indcate actual failure. Also if we have already fit the model we are doing more work than we need. However, extracting model during optimisation presents some challenges, may need to specialise the optimiser. Will look further .....