Seasonality

Seasonality and Holiday Effects

Modelling Seasonality

Seasonal Periods

Given a Polars offset alias freq, use functime.offsets.freq_to_sp to return a list of seasonal periods.

seasonal_periods = {
    "1s": [60, 3_600, 86_400, 604_800, 31_557_600],
    "1m": [60, 1_440, 10_080, 525_960],
    "30m": [48, 336, 17_532],
    "1h": [24, 168, 8_766],
    "1d": [7, 365],
    "1w": [52],
    "1mo": [12],
    "3mo": [4],
    "1y": [1],
}

Method 1. Dummy Variables / Categorical

Use add_calendar_effects to generate datetime and calendar effects. functime supports two strategies to model seasonality as discrete features: though a categorical column (useful for forecasters with native categorical features support e.g. lightgbm) or multiple binary columns (i.e. one-hot encoding). Check out Chapter 7.4: Seasonal dummy variables for a quick primer.

If you choose the dummy variable strategy, beware of the "dummy variable trap" (i.e. remember to set fit_intercept=False if you decide to include all dummy columns).

• minute: 1, 2, ..., 60 (in a day)
• hour: 1, 2, ..., 24 (in a day)
• day: 1, 2, ..., 31 (in a month)
• weekday: 1, 2, ..., 7 (in a week)
• week: 1, 2,..., 52 (in a year)
• quarter: 1, 2, ..., 4 (in a year)
• year: 1999, 2000, ..., 2023 (any year)

from functime.feature_extraction import add_calendar_effects

# Returns X with one categorical column "month" with values 1,2,...,12
X_new = X.pipe(add_calendar_effects(["month"])).collect()

# Returns X with one-hot encoded calendar effects
# i.e. binary columns "month_1", "month_2", ..., "month_12"
X_new = X.pipe(add_calendar_effects(["month"]), as_dummies=True).collect()

Method 2. Fourier Terms

Fourier terms are a common way to model multiple seasonal periods and complex seasonality (e.g. long seasonal periods 365.25 / 7 ≈ 52.179 for weekly time series). For every seasonal period sp and Fourier term k=1,..,K pair, there are 2 fourier terms sin_sp_k and cos_sp_k.

Fourier terms can be used to approximate a continuous periodic signal, which can then be used as exogenous regressors to model seasonality. Chapter 12.1: Complex Seasonality from Hyndman's textbook "Forecasting: Principles and Practice" contains a great practical introduction to this topic.

add_fourier_terms returns the original X DataFrame along with the Fourier terms as additional columns. For example, if sp=12 and K=3, X_new would contain the columns sin_12_1, cos_12_1, sin_12_2, cos_12_2, sin_12_3, and cos_12_3.

from functime.offsets import freq_to_sp
from functime.feature_extraction import add_fourier_terms

sp = freq_to_sp["1mo"][0]
X_new = X.pipe(add_fourier_terms(sp=sp, K=3)).collect()

Modelling Holidays / Special Events

functime has a wrapper function around the holidays Python package to generate categorical features for special events. Dates without a holiday are filled with nulls.

from functime.feature_extraction import add_holiday_effects

# Returns X with two categorical columns "holiday__US" and "holiday__CA"
north_america_holidays = add_holiday_effects(country_codes=["US", "CA"])
X_new = X.pipe(north_america_holidays).collect()

# Returns X with one-hot encoded holidays (e.g. "holiday__US_christmas)
north_america_holidays = add_holiday_effects(country_codes=["US", "CA"], as_dummies=True)
X_new = X.pipe(north_america_holidays).collect()

Custom Events

If you have your own custom special events (e.g. special promotions), you can always create your own dummy variables as Polars boolean series.