Why is the first PACF equal to the first ACF?

There are no intervening lags to adjust for.

Does zero partial autocorrelation imply full independence?

Not in general; it describes a linear relationship, and stronger conclusions require additional assumptions.

Can PACF be estimated in different ways?

Yes, including regression-based and Yule-Walker approaches, which can differ in finite samples.

Should a trend be ignored when reading the PACF?

No. Consider whether the stationarity assumptions and data preparation are appropriate first.

AI Glossary / Time series and signal processing / Intermediate

Partial autocorrelation

Definition

Partial autocorrelation at lag k measures the linear relationship between a time-series value and its value k periods earlier after accounting for the intervening lags. The collection across lags is the partial autocorrelation function, or PACF. It helps distinguish direct linear lag relationships from associations transmitted through intermediate observations.

What it means

If today's workload resembles yesterday's and yesterday's resembles the day before, today's workload can correlate with the day before even without an additional lag-two relationship. Partial autocorrelation asks what linear association remains after the intermediate observation is accounted for.
The word partial refers to that adjustment. It does not mean using part of the dataset, and it does not establish a causal effect. In the usual time-series definition, lag k adjusts for lags one through k minus one, rather than every other lag in the past and future.

How it works

For a covariance-stationary process, remove the linear projections of both the current value and the lag-k value on the intervening observations. The correlation between their remaining components is the population partial autocorrelation, provided the relevant variances and covariance matrices are nondegenerate.
Equivalently, fit a population autoregression using the first k lags; its final coefficient gives the lag-k PACF under the standard stationary formulation. The lag-one PACF equals the lag-one autocorrelation because there are no intervening observations to remove.
For a stationary autoregressive process of order p, population partial autocorrelations beyond p are zero under the usual assumptions. Estimated values from a finite sample will fluctuate around those population values. Estimation methods, sample size, nonstationarity, and unusual observations can change the appearance of the plot.

Three practical examples

An analyst examines a PACF after suitable preparation of a demand series to identify plausible autoregressive orders for further testing.
A forecasting team compares ACF and PACF plots rather than mistaking a gradual ACF decay for evidence that every lag needs a separate coefficient.
A model reviewer investigates remaining lag relationships in residuals as part of a broader diagnostic process, while recognizing that one isolated spike can occur by chance.

Think of it this way

Think of a message passed along a chain. Two people far apart may share information because the person between them relays it. Partial autocorrelation asks whether an additional linear association remains after accounting for the intermediate links. The analogy describes statistical dependence, not proof of who caused whom to act.

Worked example

Consider an original stationary AR(1) process X_t = 0.6 X_(t-1) + e_t, with independent mean-zero innovations and a stationary initialization. Its population autocorrelation at lag k is 0.6 raised to k. Thus rho_1 = 0.6 and rho_2 = 0.36.
At lag two, the partial autocorrelation is (rho_2 - rho_1 squared)/(1 - rho_1 squared). Substituting gives (0.36 - 0.36)/(1 - 0.36) = 0/0.64 = 0.
The ordinary lag-two correlation is therefore 0.36, while the lag-two partial correlation is zero. Their difference is exactly the point: the lag-two linear association is accounted for through lag one in this model. Lag one's partial autocorrelation remains 0.6.
The calculation uses population correlations from a specified process, not empirical values guessed from a short series. A sample PACF need not be exactly zero beyond lag one. Independent fraction arithmetic confirms the numerator and denominator, and the AR(1) recurrence confirms the correlation values.

Illustrative case study

Fictional case: Amber Scheduling sees demand correlations extending across several weeks and initially proposes a model with many lag terms. Its analyst notices that the PACF is more consistent with a short autoregressive structure and compares that candidate with alternatives.
The team checks stationarity assumptions, residual behavior, and later forecasts before choosing a model. It does not treat the plotted cutoff as definitive evidence. This invented case illustrates informed model identification; it does not claim that a PACF alone selects the best forecasting model or establishes a business mechanism.

Common misconceptions and mistakes

Do not confuse PACF with the ordinary ACF or interpret a PACF coefficient as a causal effect. Adjustment for intervening lags does not remove every possible confounder.
Do not read approximate confidence bands as a guarantee that every spike outside them is meaningful. Inspecting many lags creates opportunities for chance exceedances, and standard bands rely on assumptions.
Do not force an exact cutoff on noisy sample estimates. Trends, seasonal structure, structural changes, and inappropriate preparation can produce misleading patterns that require separate investigation.

Business use

PACF plots can narrow the model candidates that a business forecasting team needs to fit, helping it avoid unnecessary lag terms. They are particularly useful alongside domain knowledge and transparent statistical models.
Record the series transformation, frequency, estimation method, and lag range. Compare resulting candidates on relevant forecast horizons and inspect residuals. For seasonal data, consider seasonal lags explicitly rather than treating every pattern as ordinary short-run dependence. The plot is a diagnostic aid within a modeling workflow, not an automated approval of a model.

Check your understanding

What is the example's ordinary lag-two correlation?

Answer: It is 0.36.

What is its lag-two partial autocorrelation?

Answer: Zero.

Which lag is adjusted for at lag two?

Answer: Lag one.

Would a sample estimate of 0.04 disprove the specified population zero by itself?

Answer: No; sampling variability and the estimation context must be assessed.

Related terms

Autocorrelation
Stationarity
Partial correlation
ACF
PACF
autoregression
ARIMA
Yule-Walker equations
residual diagnostics

Learn next

simulate a stationary AR(1) process and compare its sample ACF and PACF with the population functions. Repeat with different sample lengths to see how apparent cutoffs and isolated spikes change through sampling variation.

Sources

Sources: Hyndman and Athanasopoulos, Forecasting: Principles and Practice, ACF and PACF in ARIMA modeling: https://otexts.com/fpp3/arima.html . Hyndman, Glossary of Forecasting Terms: https://robjhyndman.com/mwh3/FG4.pdf . Relevant primary textbook and glossary passages consulted. Population example and fictional case are original.

Previous: Partial assignment Next: Partial autocorrelation function