Stationarity and Autocorrelation
You'll learn: How stationarity and autocorrelation work together to reveal the predictable structure in time series data.
The Core Relationship
While you've learned about stationarity and autocorrelation separately, understanding their relationship is crucial for time series modeling. Stationarity ensures that statistical properties remain constant over time, while autocorrelation measures how values relate to their own past values at different lags.
Think of stationarity as the "stable stage" and autocorrelation as the "memory pattern." A stationary series might still have strong autocorrelation (today's value depends on yesterday's), but that relationship stays consistent over time. In contrast, non-stationary data has autocorrelation patterns that shift unpredictably.
Why They Matter Together
Many forecasting models (like ARIMA) assume both:
- Stationarity: So parameters learned from past data apply to future predictions
- Autocorrelation: So there's actually something to learn—patterns linking past to future
A non-stationary series often shows misleadingly high autocorrelation at all lags (this is called "spurious correlation"). After differencing to achieve stationarity, the true autocorrelation structure emerges—revealing which specific lags matter.
The Practical Workflow
When building forecasts:
- First check stationarity (you've learned this)
- If non-stationary, apply differencing
- Then examine autocorrelation/partial autocorrelation plots on the stationary series
- These plots tell you which lag features or model orders to use
Key Takeaway: Stationarity is the prerequisite that makes autocorrelation analysis meaningful—only on stationary data can you trust autocorrelation patterns to identify genuine predictive relationships rather than trending artifacts.