[MFLES] Seasonality prediction incorrect when forecasting multi-seasonal time series beyond one period

[tehunter]

What happened + What you expected to happen

If you have a time series with multiple seasonality periods, MFLES currently constructs the combined seasonality pattern in an array sized to the largest of the seasonality periods. This leads to an incomplete combined pattern being stored (unless the largest seasonality period happens to be a multiple of all the smaller periods).

For example, in the following example, the data has seasonal periods of 2 and 5. 10 days are needed to correctly capture the combined seasonality, but the forecast uses only a 5-day pattern.

I believe it's inconsequential if the forecast horizon is less than a full seasonality period.

Versions / Dependencies

Bug exists in main branch

Reproducible example

from statsforecast import StatsForecast
from statsforecast.models import (
    MFLES
)
import pandas as pd
import numpy as np

season_1 = np.array([0, 100])
season_2 = np.array([0, -200, 0, 0, 0])
dates = pd.date_range("2024-01-01", "2024-01-30")
n = len(dates)

sample_ts = pd.DataFrame({
    "ds": dates,
    "y": np.resize(season_1, n) + np.resize(season_2, n),
    "unique_id": "1"
})

model = MFLES(
    season_length=[2, 5],
    trend_lr=0,
    residuals_lr=0,
)
sf = StatsForecast(
    models=[model],
    freq="D"
)
sf.fit(
    sample_ts
)

forecast_df = sf.predict(h=30)
sf.plot(sample_ts, forecast_df)

Issue Severity

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