Autocorrelation

Autocorrelation is what happens when a time series “remembers” its own recent past. If today’s value tends to resemble yesterday’s (or tends to swing opposite), the series isn’t just random noise—it has a pattern of dependence over time.

What it means technically

Autocorrelation (also called serial correlation) is the correlation between a series and a lagged version of itself. For a lag of k, you compare values like (today) with (k days ago) across the whole series. Positive autocorrelation means high values tend to follow high values (and low follow low). Negative autocorrelation means the series tends to alternate: high values are followed by low values.

How to read it in practice

Autocorrelation is often summarized by the autocorrelation function (ACF), which reports correlation at lag 1, lag 2, and so on. Common real-world examples:

• House prices: monthly prices often show positive autocorrelation because markets move gradually.
• Electricity demand: strong autocorrelation at lag 24 hours (daily cycle) and 168 hours (weekly cycle).
• Exam scores over semesters: a student’s performance can be correlated with their previous term.

Why it matters in AI/ML

Many ML methods quietly assume observations are independent. Autocorrelation breaks that assumption, which can:

• Make train/test splits misleading if you randomly shuffle time-ordered data.
• Cause overly optimistic metrics because the model “sees” near-duplicates across time.
• Show up as patterns in residuals, signaling a model is missing time structure.

Time-series models like AR, ARIMA, and state-space models explicitly use autocorrelation. In practice, you’ll see ACF/PACF plots in statsmodels, and careful time-based validation in scikit-learn workflows.