Scribd
Upload
125 views5 pages

Partial Autocorrelation and The PACF First Examples

The document discusses the partial autocorrelation function (PACF) and how it can help identify the order of an autoregressive (AR) time series process. The PACF cuts off after the order of an AR process, similar to how the autocorrelation function (ACF) cuts off for a moving average (MA) process. Code examples are provided to simulate AR processes and examine their ACFs and PACFs. The Beveridge wheat price data is also analyzed, and its PACF suggests it follows a second-order AR process.

Uploaded by

Srikanth Shenoy
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF, TXT or read online on Scribd
125 views5 pages

Partial Autocorrelation and The PACF First Examples

The document discusses the partial autocorrelation function (PACF) and how it can help identify the order of an autoregressive (AR) time series process. The PACF cuts off after the order of an AR process, similar to how the autocorrelation function (ACF) cuts off for a moving average (MA) process. Code examples are provided to simulate AR processes and examine their ACFs and PACFs. The Beveridge wheat price data is also analyzed, and its PACF suggests it follows a second-order AR process.

Uploaded by

Srikanth Shenoy
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF, TXT or read online on Scribd
You are on page 1/ 5

Thistleton and Sadigov Partial Autocorrelation and the PACF Week 3

The Partial Autocorrelation Function


Let’s suppose you have time series data in front of you, and you have plotted the series as well as
plotted the ACF. Unless you have created the data yourself, say with arima.sim(), or if you have
good knowledge of the physical process that generated your time series, it’s not easy to tell from
the time series data alone just what the generating process is.

Now the ACF is one of our primary tools for characterizing an autoregressive or moving average
process. We’ve already seen how to find the ACF as a formula for MA(q) and AR(p) processes and
learned to predict the characteristic shapes which occur in the ACF for these simple processes.
First, the “good news”:

A moving average process of order q has an ACF that cuts off after q lags.

So, if you have an MA() process with enough terms that you believe the ACF is well estimated,
and the ACF cuts off after 4 lags, you can be reasonably sure you have an MA(4) process.

How about an AR(p) process? That is, if you know you have an AR(p) process, can you tell what
the order of the process is? It would be terrific if we had a plot that would function for an AR(p)
process the way the ACF does for the MA(q) process.

Generating Data and Observing the Partial Autocorrelation Function


More “good news”! As we develop in this lecture, the Partial Autocorrelation Function, or PACF
will help us to find the order of an AR(p) process. We will develop this graph in the lecture below.
First, let’s create some data and look at some pictures.

rm( list=ls( all = TRUE ) )


par(mfrow=c(3,1))

phi.1 = .6; phi.2 = .2; data.ts = arima.sim(n = 500, list(ar = c(phi.1, phi.2)))

plot(data.ts, main=
paste("Autoregressive Process with phi1=",phi.1," phi2=",phi.2) )
acf(data.ts, main="Autocorrelation Function")
acf(data.ts, type="partial", main="Partial Autocorrelation Function")

Practical Time Series Analysis Page 1


Thistleton and Sadigov Partial Autocorrelation and the PACF Week 3

You should run the code several times. You will observe that, while the actual time series itself
changes from simulation to simulation, the ACF and PACF are relatively constant. Now run the
code again, this time with the following two lines substituted into the obvious place:

phi.1 = .9; phi.2 = -.6; phi.3 = .3;


data.ts = arima.sim(n = 500, list(ar = c(phi.1, phi.2, phi.3)))
plot(data.ts, main= paste("Autoregressive Process with phi1=",
phi.1," phi2=",phi.2," phi3=",phi.3) )

Practical Time Series Analysis Page 2


Thistleton and Sadigov Partial Autocorrelation and the PACF Week 3

You can continue to play with the code, changing the number of terms and the coefficients used
in your simulations. Do you have any conjectures?

Partial Autocorrelation Function and the Beveridge Wheat Price Data Set
Here is a classic data set in Time Series Analysis, the good old Beveridge Wheat Price Index. You
can get the numbers at the Time Series Data Library…just visit the website:
http://datamarket.com/data/list/?q=provider:tsdl

The data set originally appeared in a paper called “Weather and Harvest Cycles” (Beveridge, 1921)
and has also been discussed in subsequent papers critiquing the original analysis, for example
(Sargan, 1953) and (Granger & Hughes, 1971). While we note that there are issues with
Beveridge’s analysis, we present the data as a nice illustration of the topic at hand.

Can you follow along on this code? I have downloaded the data into a text file, removing some
header information. Next, I create a time series with the data in the second column (the actual
prices, starting in the year 1500) and create a “filter” with the simple moving average that uses 15
data points on either side of a given year to introduce some smoothing. We plot the original series
and the smoothed series on the same axes.

beveridge = read.table("beveridge.txt", header=TRUE)


beveridge.ts = ts(beveridge[,2], start=1500)
plot( beveridge.ts, ylab="price", main="Beveridge Wheat Price Data")
beveridge.MA = filter(beveridge.ts, rep(1/31, 31), sides = 2)
lines(beveridge.MA, col="red")

Practical Time Series Analysis Page 3


Thistleton and Sadigov Partial Autocorrelation and the PACF Week 3

Now Beveridge transformed his data by scaling each data point by its corresponding smoothed
value. I’ve done this with the following lines, and plotted the usual graphs.

par(mfrow=c(3,1))
Y = beveridge.ts/beveridge.MA
plot( Y, ylab="scaled price", main="Transformed Beveridge Wheat Price Data")
acf(na.omit(Y),
main="Autocorrelation Function of Transformed Beveridge Data")
acf(na.omit(Y), type="partial",
main="Partial Autocorrelation Function of Transformed Beveridge Data")
The acf() function doesn’t like missing data, so the first and last 15 numbers are ignored or omitted
with the clean-up function na.omit(). What do you notice about the Partial Autocorrelation
function?

Practical Time Series Analysis Page 4


Thistleton and Sadigov Partial Autocorrelation and the PACF Week 3

We will use the routine ar() to estimate the coefficients of our model. In other lectures we describe
how to do this; for now, just think of this as similar to a call to lm() when we are doing a regression
curve fitting. We believe we have a model

𝐴𝑅(𝑝) 𝑝𝑟𝑜𝑐𝑒𝑠𝑠: 𝑋𝑡 = 𝑍𝑡 + 𝜙1 𝑋𝑡−1 + ⋯ + 𝜙𝑝 𝑋𝑡−𝑝

We let the ar() function find the coefficients for us. We discuss selecting a model using quality
criteria such as the Akaike Information Criterion (AIC) in other lectures. As R will tell you, the
ar() routine will “Fit an autoregressive time series model to the data, by default selecting the
complexity by AIC.” By complexity we mean how many terms to take, or what the value of p is.
If we will allow up to 5 terms in our model (a reasonable number) we call

ar(na.omit(Y), order.max = 5)

And obtain a second order model

Coefficients:
1 2
0.7239 -0.2957

Order selected 2 sigma^2 estimated as 0.02692

Just to state the obvious at this point:

An autoregressive process of order p, an AR(p), has a PACF that cuts off after p lags.

That’s handy!

Now that we know one of the ways we use a PACF, let’s discuss the intuition behind the PACF
and how to calculate it for a stochastic process and how to estimate it for some given time series
data.

Practical Time Series Analysis Page 5

You might also like