Theoretical questions regarding forecast and fable

Machine Learning and Modeling
,
1

Hi

Some theoretical questions:

Do the functions

forecast::Acf forecast::Pacf 
feasts::ACF feasts::PACF

use the Yule-Walker method or Burg's method?

Burg, J. (1967), "Maximum Entropy Spectral Analysis", Proceedings of the 37th Meeting of the Society of Exploration Geophysicists.

Also, for the following process:

y(t) = mu + phi1*y(t-1) + beta*trend + eps(t)

y = series
mu = constant
phi1 = coefficient
beta = coefficient
trend = time-trend
eps = white noise process

Using recursive back-substitution I can calculate the E(y(t)) and Var(y(t)) for phi1=1. What is the E(y(t)) and Var(y(t)) for |phi1|<1 ?

thanks,
Amarjit

2

All four functions call stats::acf() or stats::pacf() to do the calculations. They are simply wrappers to provide more convenient output or improved plots.

The help file for stats::acf() and stats::pacf() gives a brief explanation. For ACF:

For type = "correlation" and "covariance" , the estimates are based on the sample covariance. (The lag 0 autocorrelation is fixed at 1 by convention.)

and for PACF

The partial correlation coefficient is estimated by fitting autoregressive models of successively higher orders up to lag.max.

Later it gives the reference:

Venables, W. N. and Ripley, B. D. (2002) Modern Applied Statistics with S. Fourth Edition. Springer-Verlag.
(This contains the exact definitions used.)

If you look at the reference, you will find the standard covariance estimates of the ACF (p390), and the PACF defined using the Yule-Walker equations (on p398).

If you want to see precisely how it is done, you can always look at the code. After all, R is open source. The code for stats::acf() includes

acf <- .Call(C_acf, x, lag.max, type == "correlation")

which tells you that it calls C to do the calculation. If you look at the C sources, you will find the relevant function here:

Similarly, the PACF uses the C code here:

The last question looks like a homework problem. To solve moment problems for problems like this, backshift notation is helpful for re-arranging the equation into a suitable form -- you want y_t on the left and only lagged error terms on the right. Then you need a couple of standard results for infinite power series.

3

Thanks.

The reason I ask regarding autocorrelations is ESTIMA use Burg as default https://estima.com/webhelp/index.html?context=1290 see Correlate - method used to compute correlations - autocorrelations - Yule–Walker vs Burg Methods.

My second question, the process is: y(t) = mu + phi1*y(t-1) + beta*(t) + eps(t) , as from ADF test, third model form.

And I have

E[y(t)] = mu*(1 + phi1 + phi1^2 + ... + phi1^(n-1)) + 
          (phi1^(n-1))*E[y(t-n)] +
          E[beta*(n) + phi1*beta*(n-1) + phi1^2*beta*(n-2) + ... + phi1^(n-1)*beta*(n)]

As E[eps(t)]=0 by definition.

With |phi1|<1, as n->infinity, the middle-term cancels to zero.

So the first term is

E[y(t)] = mu/(1-phi1) via infinite power series result: from i=0 ∑ to i=(n-1) phi^i = 1/(1-phi1) as per 'textbook(s)'.

Same reason for the beta term and therefore,

E[y(t)] = (mu + beta*(t))/(1-phi1)

Correct?

I can then use Var[y(t)] = (E[y(t)^2] - (E[y(t)])^2).

I am not certain how backshift notation makes things easier?

4

This topic was automatically closed 90 days after the last reply. New replies are no longer allowed.

If you have a query related to it or one of the replies, start a new topic and refer back with a link.