Why are ACF and PACF equal at lag one?

No intermediate lag needs to be removed.

Does zero PACF imply no relationship at all?

It indicates no remaining linear association in the specified adjustment, not necessarily absence of nonlinear dependence.

Can software methods disagree?

Yes, because finite-sample estimators and conventions differ.

AI Glossary / Statistics and probability / Intermediate

Partial autocorrelation function

Definition

The partial autocorrelation function, or PACF, describes the linear association between a time series and its lagged values after accounting for the intervening lags. At lag k, it measures the correlation between the current value and the value k periods earlier after their linear relationships with lags 1 through k minus one have been removed.

What it means

Ordinary autocorrelation includes connections transmitted through intermediate observations. Today's value can be associated with the value two periods ago because both relate to yesterday's value.
Partial autocorrelation isolates the remaining linear association for a specified lag after adjusting for those intervening observations. It is a tool for understanding time dependence and suggesting forecasting models. It is not a measure of causal influence, and removing linear relationships does not eliminate every possible nonlinear dependence.

How it works

Start with a clearly defined, regularly spaced series. Inspect trends, seasonal structure, missing periods, and changes in measurement. The conventional stationary PACF interpretation requires an appropriate stable covariance structure, often after justified transformation or differencing.
At lag one, there are no intervening lags, so partial autocorrelation equals ordinary lag-one autocorrelation. At lag two, adjust both current and two-period-lagged values for the one-period lag, then correlate their residuals.
For a covariance-stationary process, the population PACF at lag k also equals the last coefficient in the best linear autoregression using lags one through k. This connection makes it useful for studying autoregressive order.
A pure stationary autoregressive process of order p has population partial autocorrelations equal to zero beyond lag p, assuming the usual nondegenerate model conditions. Sample estimates will fluctuate around those zeros rather than exhibit a perfectly clean cutoff.
Several estimators are available, including regression, Yule-Walker, Burg, and recursion-based methods. They can differ in finite samples, especially at longer lags. Record the estimator and lag range instead of expecting every software default to agree.
Plots often display approximate pointwise reference bands. They are not simultaneous guarantees across all inspected lags, and their validity depends on the assumptions used. A few isolated spikes can arise by chance.
Use PACF together with autocorrelation plots, model diagnostics, information criteria, and forecasting evaluation. Seasonal patterns and mixed autoregressive-moving-average processes can make simple visual order selection unreliable. Fit and validate plausible candidates rather than treating the plot as an automatic specification rule.

Three practical examples

Demand forecasting: an analyst checks whether earlier demand contains linear information beyond the most recent periods.
AI service monitoring: a team examines dependence in a regularly sampled performance measure before choosing a forecasting baseline.
Model diagnosis: researchers compare candidate autoregressive structures while checking whether residual dependence remains after fitting.

Think of it this way

Think of a message passed through several people. The first and third people may share information because the second relayed it. Partial autocorrelation asks what linear association remains between the endpoints after accounting for the people in between. It does not establish who caused the original message.

Worked example

For a stationary series with lag-one autocorrelation rho_1 and lag-two autocorrelation rho_2, the lag-two partial autocorrelation is (rho_2 - rho_1 squared)/(1 - rho_1 squared), provided the denominator is nonzero.
Consider an original theoretical example with rho_1 = 0.60 and rho_2 = 0.36. The numerator is 0.36 - 0.60 squared = 0, and the denominator is 1 - 0.36 = 0.64. The lag-two PACF is therefore zero, despite ordinary lag-two autocorrelation being 0.36.
This pattern is consistent with a stationary AR(1) process whose coefficient is 0.60: its autocorrelations are 0.60^k, while its population PACF vanishes after lag one.
For comparison, keep rho_1 = 0.60 but set rho_2 = 0.50. The lag-two partial autocorrelation is (0.50 - 0.36)/0.64 = 0.21875. A positive linear association remains after lag-one adjustment.
These are stipulated population-correlation illustrations, not estimates from an undisclosed dataset. A finite sample would add estimation uncertainty, and two autocorrelation values alone would not fully identify an entire time-series model.

Illustrative case study

Illustrative fictional case: Willow Forecasts uses PACF plots to explore daily support demand. An analyst sees a long sequence of large values and proposes a high-order autoregression.
The team first examines the time plot and discovers a changing level and weekly pattern. It revises the preprocessing and compares plausible seasonal models using time-ordered validation. It treats the PACF as one diagnostic among several. This fictional case illustrates why interpretation depends on the series being modeled.

Common misconceptions and mistakes

Do not confuse PACF with ordinary autocorrelation or with cross-correlation between different series. Do not interpret a partial correlation as proof of causality or complete conditional independence without stronger assumptions.
Avoid selecting every lag that crosses a reference band. Do not assume sample estimates reproduce theoretical cutoffs exactly. Large-lag estimates can be unstable, and differencing or seasonal adjustment changes the object whose PACF is being studied.

Business use

PACF helps businesses build transparent time-series baselines and understand whether additional lags may improve linear prediction. Its practical value comes from combining it with domain knowledge and out-of-sample performance. Reports should state the time interval, preprocessing, estimator, lag range, and uncertainty interpretation so that patterns can be reproduced and assessed.

Check your understanding

Why is lag-two PACF zero in the first example?

Answer: The ordinary lag-two association equals the part explained through lag one under the correlation calculation.

What is the second example's value?

Answer: 0.21875.

Does either calculation establish an observed process's true order?

Answer: No.

Related terms

Autocorrelation function
Partial correlation
Stationarity
autoregressive model
ARIMA
lag
Yule-Walker equations
time-series diagnostics

Learn next

Next compare theoretical ACF and PACF patterns for autoregressive and moving-average models. Then examine estimation methods, seasonal lags, and time-ordered validation of competing forecasting specifications.

Moving average

Sources

Hyndman and Athanasopoulos, Forecasting: Principles and Practice, ARIMA models, ACF and PACF plots: https://otexts.com/fpp3/arima.html — author explanation of partial autocorrelation and autoregressive interpretation. statsmodels, pacf: https://www.statsmodels.org/stable/generated/statsmodels.tsa.stattools.pacf.html — official estimator options and interval conventions. Relevant passages were read; the two correlation calculations and fictional case are original.

Previous: Partial autocorrelation Next: Partial correlation