Course contentsShow
← Data Science
Lesson 718 of 2,14519. Time Series FoundationsFree lesson

Interpreting Stationarity Test Results

Combine multiple tests and visual evidence to make confident conclusions about whether a series requires transformation.

Interpreting Stationarity Test Results

You'll learn how to combine multiple statistical tests and visual evidence to confidently decide whether your time series needs transformation before modeling.

The Challenge of Multiple Signals

You've learned the Augmented Dickey-Fuller (ADF) test and the KPSS test, but they sometimes give contradictory signals. ADF tests the null hypothesis of non-stationarity (you want low p-values), while KPSS tests the null hypothesis of stationarity (you want high p-values). It's like having two doctors giving opposite opinions—you need a strategy to reconcile them.

A Practical Decision Framework

Think of stationarity testing like diagnosing a car problem: you don't rely on just one gauge. Here's how to triangulate:

  1. Visual inspection first: Does your plot show obvious trend or changing variance? Your eyes are powerful diagnostic tools.

  2. ADF test: Low p-value (< 0.05) suggests stationarity. The series likely doesn't need differencing.

  3. KPSS test: High p-value (> 0.05) suggests stationarity. The series is likely stable.

  4. Agreement is confidence: When both tests agree (ADF rejects + KPSS doesn't reject), you can confidently call the series stationary.

  5. Disagreement requires judgment: If tests conflict, lean on your visual evidence and domain knowledge. Often, KPSS is more sensitive to trend-stationarity issues.

The Four Possible Outcomes

  • Both say stationary: Proceed with modeling—no transformation needed.
  • Both say non-stationary: Apply differencing or detrending.
  • ADF says non-stationary, KPSS says stationary: Likely difference-stationary; try first differencing.
  • ADF says stationary, KPSS says non-stationary: Trend-stationary; consider detrending instead of differencing.

Key Takeaway: Never rely on a single stationarity test in isolation—combine ADF results, KPSS results, and visual inspection to make robust decisions about whether your time series requires transformation before modeling.