1) Definition

  • Seasonality = a repeating, predictable pattern in data that occurs at regular intervals (daily, weekly, yearly, etc.).
  • Different from random fluctuations — it’s systematic and periodic.

Example: Retail sales peak every December due to the holiday season.

2) Characteristics

  • Frequency: how often the cycle repeats (daily, weekly, yearly).
  • Amplitude: size of the seasonal effect (strong vs weak seasonality).
  • Phase: timing of peaks/troughs (e.g., summer vs winter).

3) Examples

  • Retail: More purchases on weekends, big spikes in holiday seasons.
  • Web traffic: Higher usage during weekdays, lower at night.
  • Finance: “January effect” in stock returns.
  • Healthcare: Flu cases peak in winter.
  • Energy: Electricity demand higher in summer (air conditioning).

4) Seasonality vs Trend vs Noise

  • Trend = long-term increase/decrease (e.g., rising EV adoption).
  • Seasonality = repeating cycle (e.g., EV sales spike every December).
  • Noise = irregular, unpredictable variations.

5) Detection Methods

  • Visualization: line plots, seasonal decomposition.
  • Statistical tests: autocorrelation, Fourier transforms.
  • Time series decomposition:
    • STL (Seasonal-Trend decomposition using Loess)
    • Classical decomposition (additive/multiplicative models).

6) Handling Seasonality in Models

  • Classical time-series models:
    • ARIMA → extended to SARIMA for seasonal patterns.
  • ML models:
    • Add seasonal features (day-of-week, month, holiday flag).
    • Fourier features for periodic cycles.
  • Forecasting libraries: Prophet, GluonTS, statsmodels handle seasonality directly.

7) Connection to Drift

  • Seasonality can look like drift if not accounted for.
  • Example: Sales model trained in summer performs poorly in winter → not true drift, just seasonality.
  • Guardrail implication: distinguish natural seasonal cycles from unexpected distribution shifts.

Summary

  • Seasonality = repeating, predictable cycles in data.
  • Common in retail, finance, healthcare, energy, web traffic.
  • Must separate from trend and noise.
  • Handled via decomposition, SARIMA, Prophet, or feature engineering.
  • Important for ML monitoring: avoid confusing seasonality with data drift.