In josephsdavid/tstest: tstest

AR(1) Models and Filtering

Algebra review

Be able to find the roots of a 2nd order polynomial

$$ x = \frac{ - b \pm \sqrt {b^2 - 4ac} }{2a} $$

$$ z = a + bi$$ $$ i = \sqrt{-1} $$ a is real b is imaginary $$ z^{} = a - bi $$ z is complex conjugate

It is basically a vector so absolute value is just the magnitude

Unit circle is the values in the complex plane in which the magnitude of z is equal to one. If Z is a complex number with magnitude greater than one, it is outside the unit circle.

quad_form <- function(a,b,c) {
    rad <- b^2 - 4 * a * c
    if(is.complex(rad) || all(rad >= 0)) {
        rad <- sqrt(rad)
    } else {
        rad <- sqrt(as.complex(rad))
    }
    round(cbind(-b - rad, -b + rad) / (2*a),3)
}

quad_form(c(1,4), c(1,-5), c(6,2))

Linear filters

A filter will turn $Z_t$ into $X_t$

for example:

$$\mathrm{difference:} X_t = Z_t - Z_{t-1}$$

Moving average also a filter (smoother)

Example

Consider $ Z_1 = 8 , Z_2 =14, Z_3 = 14, Z_4 = 7$

If we apply the difference filter to this, we have that: $$X_2 = 6, X_3 = 0, X_4 = 7$$

Note: the differenced data are a realization of length n-1

We can use the difference filter to remove the wandering behjaviior

5 point moving average

We can only get a realization of n-4. (n-nopoints-1). What does the 5 point moving average do (it averages a point and two points ahead and two behind). This filter can filter out some frequencies

Types of filters:

• Low pass filter
• Filters out high frequency
• Such as 5 point moving average, it smooths
• High pass filter
• Leaves high freq but removes low freq
• Such as differencing

An example in R

mafun <- function(xs, n) {
    stats::filter(xs,rep(1,n))/n
} 
dfun <- function(xs,n){
    diff(xs, lag = n)
}
th <- ggthemes::theme_few()
data(fig1.21a)
ma <- mafun(fig1.21a,5)
d <- dfun(fig1.21a,1)
fp <- tplot(fig1.21a) + th
mp <- tplot(ma)+th
dp <- tplot(d)+th

And here lets view the 5 point moving average

cowplot::plot_grid(fp, mp, nrow = 2,labels = c("Original Data","5-point MA"))

And now a look at the difference filter

cowplot::plot_grid(fp, dp, nrow = 2,labels = c("Original Data","1st order difference"))

An another example

rlz <- gen.sigplusnoise.wge(200,coef = c(5,3),freq = c(.1,.45),vara = 10,sn = 1, plot=F)

Some plots

pfun <- function(x,n,l) {
    fp <- tplot(x) + th
    mp <- tplot(mafun(x,n)) + th
    dp <- tplot(dfun(x,l)) + th
    list("original" = fp,"MA" = mp,"dif" = dp)
}
plts <- pfun(rlz,5,1)
plts$original
plts$MA
plts$dif

Some Parzen Windows

parzen.wge(rlz)
parzen.wge(na.omit(mafun(rlz,5)))
parzen.wge(na.omit(dfun(rlz,1)))

GLPs

a GLP is a linear filter with a white noise input

We take white noise input and put it through a filter and get something that isnt white noise.

$$\sum_{j=0}^\infty \psi_j a_{t-j} = X_t - \mu$$

GLP's are an infinite sum of white noise terms. This might be weird now but later on this concept will return.

The $\psi_j$'s are called psi weights, these will be useful again later. AR, MA, and ARMA are all special cases of GLP's, and will be useful when we study confidence intervals.

AR(1) Intro

AR(p) in which p = 1

We will go through well known forms of the AR(1) model.

$$X_t = \beta + \phi_1 X_{t-1} + a_t$$

$$\beta = (1-\phi_1)\mu$$

Beta is moving average constant

We are saying that the vluse of x depends on some constant, the previous value of x, and some random white noise. Looks a lot like regression except there is a wierd variable

we can rewirte tis as

$$ X_t = (1-\phi_1)\mu + \phi_1 X_{t-1} + a_t $$

An are 1 process is stationary iff the magnitude of $\phi_1$ is less than 1

We will deal with this a lot in the near future

AR(1) math zone

$$E\left[ X_t \right] = \mu ?$$ $$E \left[ X_t \right] =E \left[ 1-\phi_1 \mu \right] + E \left[ \phi_1 X_{t-1} \right] + E[a_t]$$

We can rewrite this as $$E \left[ X_t \right] = 1-\phi_1 \mu + \phi_1E \left[ X_{t} \right] + 0$$

$$E \left[ X_t \right] ( 1-\phi_1) = 1-\phi_1 \mu $$ $$E\left[ X_t \right] = \mu $$

Mean does not depend on T

The variance also does not, and if phi1 is less than one variance is finite

$$\sigma_X^2 = \frac{\sigma_a^2}{1-\phi_1^2}$$

and rhok is phi1 to the k

$$\rho_k = \phi_1^k, k \geq 0$$

Spectral Density of AR(1) also does not depend on time, it just monotonically increases or decreases depending on phi1:

$$S_X(f) = \frac{\sigma_a^2}{\sigma_X^2} \left( \frac{1}{\mid 1 - \phi_1 e^{-2\pi i f}\mid^2} \right)$$

the zero mean form of AR1

with zero mean, $$X_t = \phi_1 X_{t-1} + a_t$$ OR $$X_t- \phi_1 X_{t-1}$$A

AR1 with positive phi

remember we have $$\rho_k = \phi_1^k$$

With positive $\phi_1$, we have:

realizations are wandering and aperiodic

Autocorrelations are damped exponentials

Spectral density Peaks at zero.

Higher values = stronger versions of these characterstics

AR1 with negative phi

We have:

realization are oscillating

Autocorrelations are damped and alternating(negative to a power you fool)

Spectral density* has a peak at f = 0.5 (cycle length of 2)

Higher magnitude = stronger characteristics

Nonstationary

if phi = 1 or phi> 1, we have that realizations are from a nonstationary process. with it equal to one, it looks ok, and is actualy a special ARIMA model. WIth it > 1, we have crazy explosive realizations. Check this out:

nonstarma  <- function(n,phi){
    x <- rep(0,n)
    a <- rnorm(n)
    x[1:n]  <- 0
    for(k in 2:n){
        x[k] = phi*x[k-1] +a[k]
    }
    tplot(x)
}

nonstarma(n = 50, phi = 1.5) + th

josephsdavid/tstest documentation built on July 3, 2019, 4:48 p.m.