fix: compute the seasonal ETS interval variance per step

[Max Gorbuk (mkzung)]

_compute_pred_intervals computes the seasonal count once, from the whole horizon:

steps = steps = np.arange(1, h + 1)
hm = np.floor((h - 1) / season_length)

hm then sits next to steps in the ANA, AAA and AAdA variances, so every step gets the count of the last one. In Hyndman, Koehler, Ord and Snyder (2008), ch. 6, it is the count for the step being forecast. The visible effect is that the step-1 interval depends on how far ahead you forecast:

from statsforecast.models import AutoETS
from statsforecast.utils import AirPassengers as ap

m = AutoETS(season_length=12, model="ANA").fit(ap.astype(float))
for h in (1, 36):
    p = m.predict(h=h, level=[95])
    print(h, p["hi-95"][0] - p["lo-95"][0])

On main:

1 62.02033057382857
36 103.207115866462

With this change both lines read 62.02033057382857.

The fix computes hm from steps. The added test runs all six additive-error models through _compute_pred_intervals and checks the variance against sigma2 * (1 + sum of c_j**2), with c_j = alpha + beta * (phi + ... + phi**j) + gamma * [j % m == 0]. The three non-seasonal models already matched; the three seasonal ones fail on main and pass here. tests/test_ets.py is 84 passed and tests/test_models.py 206 passed.

[CLAassistant]

All committers have signed the CLA.