Types of Non-Stationarity
You'll learn how to recognize the three main types of non-stationarity that break the stationarity assumption in time series data.
The Three Flavors of Non-Stationarity
Remember that stationarity means your time series has constant statistical properties over time. When this breaks down, it typically happens in three distinct ways:
1. Trend Non-Stationarity
The mean changes systematically over time. Imagine tracking a company's revenue over years—it steadily climbs upward. The average revenue in 2020 is fundamentally different from 2010. This upward (or downward) drift means the series has no stable mean to return to.
Visual clue: The series consistently moves in one direction, like climbing stairs or sliding down a hill.
2. Seasonal Non-Stationarity
The seasonal pattern itself changes over time. Consider ice cream sales: the summer peak might grow larger each year, or the winter dip might become more pronounced. The shape of the seasonality isn't stable—what happens in summer 2020 differs structurally from summer 2010.
Visual clue: The peaks and valleys in your seasonal pattern get progressively bigger, smaller, or shift in timing.
3. Variance Non-Stationarity (Heteroskedasticity)
The spread or volatility changes over time. Think of stock prices: early in a company's history, daily price swings might be 10–20. The average might stay similar, but the uncertainty has grown.
Visual clue: The series shows a "megaphone" or "funnel" shape—wider fluctuations in some periods than others.
Why This Matters
Each type requires different fixes: detrending for trend non-stationarity, seasonal differencing for seasonal issues, and transformations (like logarithms) for variance problems.
Key Takeaway: Non-stationarity comes in three main forms—trend (changing mean), seasonal (changing pattern), and variance (changing spread)—and identifying which type you face guides your modeling approach.