Seasonality

Some patterns in data don’t happen just once—they repeat on a schedule. If ice cream sales rise every summer or electricity demand spikes every weekday morning, you’re seeing seasonality: a predictable rhythm tied to the calendar.

What seasonality means
In time series data, seasonality is a repeating pattern that occurs at a fixed, known frequency: daily, weekly, monthly, quarterly, yearly, etc. The key idea is regularity: the “shape” of the ups and downs tends to recur at the same positions in each cycle. Seasonality is different from a long-term trend (overall increase/decrease over years) and from random noise (unpredictable wiggles). It can be additive (the seasonal swing stays about the same size) or multiplicative (the swing grows as the series level grows).

Everyday examples
Common seasonal patterns include:

• Retail sales: higher in November–December, lower in January.
• Web traffic: weekday vs. weekend cycles.
• Hospital admissions: flu season peaks in winter.
• Energy demand: daily peaks in the evening; summer cooling vs. winter heating.

Why it matters in AI/ML
If a model ignores seasonality, it may mistake predictable cycles for “surprises,” leading to biased forecasts and poor anomaly detection (e.g., flagging every December sales surge as abnormal). Many forecasting methods model seasonality explicitly, such as SARIMA, ETS, and Prophet. In ML workflows, seasonality often becomes engineered features (day-of-week, month, holidays) or is handled via decomposition (trend + seasonal + residual) before training.

How it’s handled in practice
Typical approaches include:

• Seasonal decomposition (classical or STL) to separate components.
• Seasonal differencing (subtracting the value from one season ago).
• Using models with built-in seasonal terms or Fourier features.