HyperTreeSTL default forecast fails for short horizons #10

Summary

HyperTreeSTL(type="default") fails during forecasting when the forecast horizon is shorter than the learned trend smoothing window / seasonal period. I hit this with fcst_h=3 on a monthly grocery-price example.

Reproduction

Train a default HyperTreeSTL model and forecast with a short horizon, for example:

from hypertrees.models import HyperTreeSTL

fcst_h = 3
model = HyperTreeSTL(period=12, num_seasonal_components=2, freq="M", fcst_h=fcst_h)
model.train(
    lgb_params={"learning_rate": 0.05},
    num_iterations=8,
    train_data=train_one,
    validation=False,
)
model.forecast(test_data=test_one)

where test_one has exactly 3 rows and includes the required series_id, date, time, and feature columns.

Observed behavior

Forecasting raises from the default STL forward pass:

RuntimeError: Forecasting not successful: Argument #4: Padding size should be less than the corresponding input dimension

The immediate source is torch.nn.functional.pad(..., mode="reflect") in _forward_default. With fcst_h=3, the input length is 3 but the smoothing window can produce padding such as (7, 7).

There is also a second short-horizon edge case in seasonal centering: when T < period, the existing reflected tail may not be long enough to pad to a full seasonal cycle before view(C, m, N).

Expected behavior

HyperTreeSTL(type="default") should support short forecast horizons such as fcst_h=3 and fcst_h=1, or fail earlier with a clear validation error. Since the model already accepts any positive fcst_h, handling short horizons seems preferable.

Suggested fix

A local patch fixes this by:

• capping the default smoothing window by the actual sequence length so reflect padding satisfies pad < T
• bypassing smoothing for T <= 1
• repeating reflected seasonal values enough times when padding T < period to a full seasonal cycle
• adding regression tests for 3-step and 1-step horizons

I can open a PR with that patch.