Latest Results
Index the per-season alpha in the seasonal ES models
`SeasonalExponentialSmoothing` and its optimized variant fit one alpha
per season, so `model_["alpha"]` has shape `(season_length,)`. Two places
copied from the scalar `SimpleExponentialSmoothing` use it unindexed, and
both raise:
SeasonalExponentialSmoothing(season_length=12, alpha=0.1).fit(y)
.predict(h=6, level=[80])
ValueError: operands could not be broadcast together with shapes (6,) (12,)
.simulate(h=6, n_paths=4)
ValueError: operands could not be broadcast together with shapes (12,) (4,)
`predict` needs the alpha of the season each step falls in, and the
existing `k = ((steps - 1) // m) + 1` already gives the cycle, so the
season is `(steps - 1) % m`. `simulate` already computes `s_idx = i % m`
for `levels` on the line above and needs the same index for `alpha`.
The scalar models are untouched - their `alpha` is 0-d and their two
call sites read differently, so nothing that currently works changes.
Point forecasts are identical before and after; only the paths that
raised now return values.VenishPaneliya:seasonal-es-per-season-alpha Latest Branches
0%
VenishPaneliya:seasonal-es-per-season-alpha 0%
DresdenGman:fix/conformal-interval-h-validation 0%
© 2026 CodSpeed Technology