Partial Autocorrelation Plot (PACF) Calculator

Explore partial autocorrelation across configurable lags, transformations, confidence levels, and estimation methods while identifying autoregressive structure, seasonal signals, and significant dependencies instantly from data.

1. Time-Series Data

Use commas, spaces, semicolons, tabs, or new lines.
Use zero to keep all observations.

2. Editable Observation Table

Edit values, add timestamps, remove rows, reverse order, or sort observations.

#Date or labelValueAction
The editor displays up to 2,000 rows for responsive browser performance.

3. Synthetic Data Generator

4. Preprocessing Options

5. PACF Calculation Settings

6. Plot and Layout Options

Formula Used

Regression definition: PACF at lag k is the coefficient of yt−k after regressing yt on lags 1 through k.

Approximate confidence limits: ±zα/2 / √n.

Regular difference: Δyt = yt − yt−1.

Seasonal difference: Δsyt = yt − yt−s.

The Yule-Walker and Levinson-Durbin methods recover reflection coefficients recursively. OLS estimates each lag order with a separate regression. Burg estimation minimises forward and backward prediction errors.

How to Use

  1. Paste values, upload a file, load a preset, or generate data.
  2. Review and edit observations in the data table.
  3. Choose missing-value, transformation, differencing, and detrending options.
  4. Select the maximum lag, estimation method, and confidence level.
  5. Customise plot labels, dimensions, confidence shading, and display options.
  6. Calculate the PACF and inspect significant lags, diagnostics, and warnings.
  7. Export the chart, results, transformed data, JSON summary, or PDF report.

Example Interpretation Guide

Observed PACF patternPossible interpretationNext check
One strong spike at lag 1Possible AR(1) structureInspect ACF decay and residuals
Strong spikes through lag 2Possible AR(2) structureCompare candidate AR orders
Repeated spikes near lag 12Possible seasonal AR behaviourCheck seasonal differencing
Many slowly declining spikesPotential trend or nonstationarityApply stationarity diagnostics
No notable spikesSeries may resemble white noiseReview Ljung-Box result

Frequently Asked Questions

What does PACF measure?

PACF measures the direct relationship between observations separated by a lag. It removes effects transmitted through shorter lags.

How is PACF different from ACF?

ACF includes direct and indirect lag relationships. PACF isolates the remaining relationship after controlling intermediate lags.

How can PACF suggest an AR order?

An autoregressive process often shows notable PACF spikes through its order. Later spikes may fall inside confidence limits.

Why should the series be stationary?

Trend and changing variance can create misleading lag patterns. Differencing or transformation may improve interpretability.

What do confidence bands mean?

They provide approximate reference limits under weak serial dependence. Spikes outside them deserve further investigation.

Which PACF method should I use?

Yule-Walker is efficient for many stationary series. OLS and Burg provide useful sensitivity comparisons.

Can isolated significant spikes be trusted?

Multiple lags increase false-positive opportunities. Interpret isolated spikes with theory, seasonality, and residual checks.

What causes seasonal PACF spikes?

Repeated cycles can produce spikes at seasonal intervals. Seasonal AR terms or seasonal differencing may help.

Does PACF confirm an ARIMA model?

No. PACF only supports model identification. Compare alternatives using diagnostics and out-of-sample performance.

Why do methods give different values?

Methods use different finite-sample assumptions and estimators. Differences grow with short, noisy, or unstable series.

Related Calculators

Time-Series Line ChartRolling Mean PlotRolling Standard-Deviation PlotSeasonal Decomposition PlotAutocorrelation Plot (ACF)Lag PlotForecast-versus-Actual PlotForecast Confidence-Band PlotSeasonal Subseries PlotCalendar Heatmap

Important Note: All the Calculators listed in this site are for educational purpose only and we do not guarentee the accuracy of results. Please do consult with other sources as well.