Course contentsShow
← Data Science
Lesson 741 of 2,14519. Time Series FoundationsFree lesson

Testing Stationarity After Transformation

Verify that differencing or detrending achieved stationarity using ADF tests and visual diagnostics.

Testing Stationarity After Transformation

You'll learn: How to confirm that your differencing or detrending actually made your time series stationary using statistical tests and visual checks.

The Core Idea

After you've applied transformations like first-order differencing, seasonal differencing, or detrending to your time series, you need to verify that the transformation worked. It's like checking your work in math class—you've done the surgery on the data, but did the patient actually get better?

Testing stationarity after transformation involves two complementary approaches:

1. Statistical Tests

Use the Augmented Dickey-Fuller (ADF) test on your transformed series. Remember that the ADF test checks the null hypothesis that the series has a unit root (is non-stationary). After transformation, you want:

  • A low p-value (typically < 0.05) to reject the null hypothesis
  • Confirmation that your series is now stationary

You might also use the KPSS test as a backup, which has the opposite null hypothesis (assumes stationarity).

2. Visual Diagnostics

Examine plots of your transformed series:

  • Time series plot: Should show constant mean and variance over time
  • ACF plot: Should show rapid decay to zero, not slow tailing off
  • Rolling statistics: Mean and variance shouldn't drift systematically

Why This Matters

If your transformation didn't achieve stationarity, you may need to apply additional transformations (like second-order differencing) or try a different approach (switching from differencing to detrending). Many forecasting models assume stationarity, so confirming it prevents garbage-in-garbage-out scenarios.

Key Takeaway: Always test both statistically (ADF/KPSS) and visually (plots, ACF) that your transformed series is stationary before proceeding to model fitting—transforming alone isn't enough without verification.