Course contentsShow
← Data Science
Lesson 719 of 2,14519. Time Series FoundationsFree lesson

What is Autocorrelation?

Understanding correlation between a time series and lagged versions of itself.

What is Autocorrelation?

You'll learn: How time series observations can be correlated with their own past values, and why this matters for forecasting.

The Core Idea

Autocorrelation (also called serial correlation) measures the relationship between a time series and a lagged version of itself. In simple terms, it tells you whether today's value is predictably related to yesterday's value, last week's value, or any earlier observation.

Think of it like this: If you know it rained today, can you predict tomorrow's weather better? That's autocorrelation in action — the past influences the present in a measurable way.

Why This Matters for Time Series

Unlike regular data where observations are independent, time series data often exhibits dependence over time. Stock prices don't jump randomly — they tend to drift. Monthly sales often follow seasonal patterns. Temperature today is usually similar to temperature yesterday.

Autocorrelation quantifies this dependence by calculating the correlation coefficient between the series and itself at different time lags:

  • Lag 1: Correlation between consecutive observations (today vs. yesterday)
  • Lag 2: Correlation between observations two steps apart (today vs. two days ago)
  • Lag k: Correlation at any time lag k

What It Reveals

High positive autocorrelation at lag 1 means values cluster together — if today is high, tomorrow is likely high too. This violates the independence assumption of standard regression and signals that time-based patterns exist that we can potentially exploit for forecasting.

Negative autocorrelation means values tend to alternate — high followed by low, creating an oscillating pattern.

Key Takeaway: Autocorrelation measures how correlated a time series is with its own past values at different lags, revealing temporal dependencies that standard statistical methods assume don't exist.