Linear Time Series (TS) Models

Financial TS: collection of a financial measurement over time

Example: log return rt

Data: {r1, r2, · · · , rT } (T data points)

Purpose: What information contained in {rt}?

Basic concepts

• Stationarity:

– Strict: distributions are time-invariant

– Weak: first 2 moments are time-invariant

What does weak stationarity mean in practice?

Past: time plot of {rt} varies around a fixed level within a
finite range!
Future: the first 2 moments of future rt are the same as those of
the data so that meaningful inferences can be made.

• Mean (or expectation) of returns:

µ = E(rt)

• Variance (variability) of returns:

Var(rt) = E[(rt − µ)2]

26
• Sample mean and sample variance are used to estimate the mean
and variance of returns.
1 T 1 X T
r̄ = rt & Var(rt) = (rt − r̄)2
X

T t=1 T − 1 t=1
• Test Ho : µ = 0 vs Ha : µ 6= 0. Compute
r̄ r̄
t= =r
std(r̄) Var(rt)/T
Compare t ratio with N (0, 1) dist.
Decision rule: Reject Ho of zero mean if |t| > Zα/2 or p-value
is less than α.

• Lag-k autocovariance:

γk = Cov(rt, rt−k ) = E[(rt − µ)(rt−k − µ)].

• Serial (or auto-) correlations:

cov(rt, rt−`)
ρ` =
var(rt)
Note: ρ0 = 1 and ρk = ρ−k for k 6= 0. Why?
Existence of serial correlations implies that the return is pre-
dictable, indicating market inefficiency.

• Sample autocorrelation function (ACF)

PT −`
t=1 (rt − r̄)(rt+` − r̄)
ρb` = PT 2
,
(r
t=1 t − r̄)
where r̄ is the sample mean & T is the sample size.
27
 • Test zero serial correlations (market efficiency)

– Individual test: for example,

Ho : ρ1 = 0 vs Ha : ρ1 6= 0
ρ̂1 √
t= r = T ρ̂1
1/T
Asym. N (0, 1).
Decision rule: Reject Ho if |t| > Zα/2 or p-value less than
α.
– Joint test (Ljung-Box statistics):
Ho : ρ1 = · · · = ρm = 0 vs Ha : ρi 6= 0
ρ̂2`
m
Q(m) = T (T + 2)
X

`=1 T − `

Asym. chi-squared dist with m degrees of freedom.

Decision rule: Reject Ho if Q(m) > χ2m(α) or p-value is
less than α.

• Sources of serial correlations in financial TS

– Nonsynchronous trading (ch. 5)

– Bid-ask bounce (ch. 5)
– Risk premium, etc. (ch. 3)

Thus, significant sample ACF does not necessarily imply market

inefficiency.

Example: Monthly returns of IBM stock from 1926 to 1997.

28
 • Rt: Q(5) = 5.4(0.37) and Q(10) = 14.1(0.17)

• rt: Q(5) = 5.8(0.33) and Q(10) = 13.7(0.19)

Remark: What is p-value? How to use it?

Implication: Monthly IBM stock returns do not have significant serial
correlations.
Example: Monthly returns of CRSP value-weighted index from
1926 to 1997.

• Rt: Q(5) = 27.8 and Q(10) = 36.0

• rt: Q(5) = 26.9 and Q(10) = 32.7

All highly significant. Implication: there exist significant serial corre-

lations in the equal-weighted index returns. (Nonsynchronous trad-
ing might explain the existence of the serial correlations, among other
reasons.)

R demonstration: Monthly IBM returns from 1926 to 1997.

> library(‘‘fSeries’’)
> ibm=[Link]("[Link]")
> acf(ibm,[Link]=15)
> x1=acf(ibm,[Link]=15)
> names(x1)
[1] "acf" "type" "[Link]" "lag" "series" "snames"
> x1$acf
, , 1
[,1]
[1,] 1.000000000
[2,] 0.074250386
[3,] 0.010515948
[4,] -0.023832213
[5,] -0.006133517
.... (editted)

29
[13,] 0.001143822
[14,] -0.057881530
[15,] -0.056303333
[16,] -0.021747045

> x2=pacf(ibm,[Link]=15)
> names(x2)
[1] "acf" "type" "[Link]" "lag" "series" "snames"
> x2$acf
, , 1
[,1]
[1,] 0.0742503857
[2,] 0.0050305627
[3,] -0.0251217508
[4,] -0.0025929619
[5,] -0.0061669432
... (editted)
[13,] -0.0562812983
[14,] -0.0467710983
[15,] -0.0124448122

> [Link](ibm,lag=5,type="Ljung")

Box-Ljung test

data: ibm
X-squared = 5.4474, df = 5, p-value = 0.3638

> [Link](log(ibm+1),lag=5,type="Ljung")

Box-Ljung test

data: log(ibm + 1)
X-squared = 5.7731, df = 5, p-value = 0.3289

Splus demonstration
> ibm=scan(file=’[Link]’) % Load data
> autocorTest(ibm,lag=5) % Perform Q(5) test

Test for Autocorrelation: Ljung-Box

Null Hypothesis: no autocorrelation

Test Statistics:

30
Test Stat 5.4474
[Link] 0.3638

Dist. under Null: chi-square with 5 degrees of freedom

Total Observ.: 864

> ibm=log(ibm+1) % Convert into log returns

> autocorTest(ibm,lag=5)

Test Statistics:
Test Stat 5.7731
[Link] 0.3289

Dist. under Null: chi-square with 5 degrees of freedom

SCA Demonstration: Output edited.

input ibm. file ’[Link]’ % Load data
--
acf ibm. maxl 10. % Compute 10 lags of ACF.

NAME OF THE SERIES . . . . . . . . . . IBM

TIME PERIOD ANALYZED . . . . . . . . . 1 TO 864
MEAN OF THE (DIFFERENCED) SERIES . . . 0.0142
STANDARD DEVIATION OF THE SERIES . . . 0.0670
T-VALUE OF MEAN (AGAINST ZERO) . . . . 6.2246

AUTOCORRELATIONS

1- 10 .07 .01 -.02 -.01 -.01 -.01 -.00 .07 .05 .04 % ACF
ST.E. .03 .03 .03 .03 .03 .03 .03 .03 .03 .03 % Stan. error
Q 4.8 4.9 5.4 5.4 5.4 5.5 5.5 10.2 12.6 14.1 % Ljung-Box Q
--
p=1-cdfc(5.4,5) % Calculate p-value
--
print p % Print p-value

.369

31
Back-shift (lag) operator
A useful notation in TS analysis.

• Definition: Brt = rt−1 or Lrt = rt−1

• B 2rt = B(Brt) = Brt−1 = rt−2.

B (or L) means time shift! Brt is the value of the series at time
t − 1.
Suppose that the daily log returns are
Day 1 2 3 4
rt 0.017 −0.005 −0.014 0.021
Answer the following questions:

• r2 =

• Br3 =

• B 2 r5 =

Question: What is B2?

What are the important statistics in practice?

Conditional quantities, not unconditional

A proper perspective: at a time point t

• Available data: {r1, r2, · · · , rt−1} ≡ Ft−1

32
 • The return is decomposed into two parts as

rt = predictiable part + not predic. part

= function of elements of Ft−1 + at

In other words, given information Ft−1

rt = µt + at
= E(rt|Ft−1) + σtt

– µt: conditional mean of rt

– at: shock or innovation at time t
– t: an iid sequence with mean zero and variance 1
– σt: conditional standard deviation (commonly called volatil-
ity in finance)

Traditional TS modeling is concerned with µt:

Model for µt: mean equation
Volatility modeling concerns σt.
Model for σt2: volatility equation

Univariate TS analysis serves two purposes

• a model for µt

• understanding models for σt2: properties, forecasting, etc.

Linear time series: rt is linear if

33
 • the predictiable part is a linear function of Ft−1

• {at} are indep. and have the same dist. (iid)

Mathematically, it means rt can be written as

∞
rt = µ + ψiat−i,
X

i=0

where µ is a constant, ψ0 = 1 and {at} is an iid sequence with mean

zero and well-defined distribution.
In the economic literature, at is the shock (or innovation) at time t
and {ψi} are the impulse responses of rt.

White noise: iid sequence (with finite variance), which is the build-
ing block of linear TS models.
White noise is not predictable, but has zero mean and finite variance.

Univariate linear time series models

1. autoregressive (AR) models

2. moving-average (MA) models

3. mixed ARMA models

4. seasonal models

5. regression models with time series errors

6. fractionally differenced models (long-memory)

34
Example Quarterly growth rate of U.S. real gross national product
(GNP), seasonally adjusted, from the second quarter of 1947 to the
first quarter of 1991.
An AR(3) model for the data is

rt = 0.005 + 0.35rt−1 + 0.18rt−2 − 0.14rt−3 + at, σ̂a = 0.01,

where {at} denotes a white noise with variance σa2. Given rn, rn−1 & rn−2,
we can predict rn+1 as

r̂n+1 = 0.005 + 0.35rn + 0.18rn−1 − 0.14rn−2.

Other implications of the model?

35
Example: Monthly simple return of CRSP equal-weighted index

Rt = 0.013 + at + 0.178at−1 − 0.13at−3 + 0.135at−9, σ̂a = 0.073

Checking: Q(10) = 11.4(0.122) for the residual series at.

Implications of the model?

Important properties of a model

• Stationarity condition

• Basic properties: mean, variance, serial dependence

• Empirical model building: specification, estimation, & checking

• Forecasting

36
 Lecture Notes of Bus 41202 (Spring 2007)
Analysis of Financial Time Series
Ruey S. Tsay

Simple AR models: (Regression with lagged variables.)

AR(1) model:

1. Form: rt = φ0 + φ1rt−1 + at, where φ0 and φ1 are real numbers,

which are referred to as “parameters” (to be estimated from the
data in an application). For example,

rt = 0.005 + 0.2rt−1 + at

2. Stationarity: necessary and sufficient condition |φ1| < 1. Why?

φ0
3. Mean: E(rt) = 1−φ1

σa2
4. Variance: Var(rt) = 1−φ21
.

5. Autocorrelations: ρ1 = φ1, ρ2 = φ21, etc. In general, ρk = φk1

and ACF ρk decays exponentially as k increases,

6. Forecast (minimum squared error):

(a) 1-step ahead forecast at time n, the forecast origin:

r̂n(1) = φ0 + φ1rn

(b) 1-step ahead forecast error:

en(1) = rn+1 − r̂n(1) = an+1

1
 Thus, an+1 is the un-predictable part of rn+1. It is the shock
at time n + 1!
(c) Variance of 1-step ahead forecast error:

Var[en(1)] = Var(an+1) = σa2.

(d) 2-step ahead forecast:

r̂n(2) = φ0 + φ1r̂n(1)

(e) 2-step ahead forecast error:

en(2) = rn+2 − r̂n(2) = an+2 + φ1an+1

(f) Variance of 2-step ahead forecast error:

Var[en(2)] = (1 + φ21)σa2

which is greater than or equal to Var[en(1)], implying that

uncertainty in forecasts increases as the number of steps in-
creases.
(g) Behavior of multi-step ahead forecasts.

7. A compact form: (1 − φ1B)rt = φ0 + at.

AR(2) model:

1. Form: rt = φ0 + φ1rt−1 + φ2rt−2 + at, or

(1 − φ1B − φ2B 2)rt = φ0 + at.

2. Stationarity condition: (factor of polynomial)

2
 3. Characteristic equation: (1 − φ1x − φ2x2) = 0
φ0
4. Mean: E(rt) = 1−φ1 −φ2
φ1
5. ACF: ρ0 = 1, ρ1 = 1−φ2 ,

ρ` = φ1ρ`−1 + φ2ρ`−1, ` ≥ 2.

6. Stochastic business cycle: if φ21 + 4φ2 < 0, then rt shows char-

acteristics of business cycles with average length
2π
k= √ ,
cos−1[φ1/(2 −φ2)]
where the cosine inverse is stated in radian. If we denote the
√
solutions of the polynomial as a ± bi, where i = −1, then we
have φ1 = 2a and φ2 = −(a2 + b2) so that
2π
k= √ .
−1
cos (a/ a2 + b2)
√
In R or S-Plus, one can obtain a2 + b2 using the command
Mod.

7. Forecasts: Similar to AR(1) models

Building an AR model

• Order specification

1. Partial ACF: (naive, but effective)

– Use consecutive fittings
– See Text (p. 40) for details
3
 – Key feature: PACF cuts off at lag p for an AR(p) model.
– Illustration
2. Akaike information criterion
2`
AIC(`) = ln(σ̃`2) + ,
T
for an AR(`) model, where σ̃`2 is the MLE of residual vari-
ance.
Find the AR order with minimum AIC for ` ∈ [0, · · · , P ].
3. BIC criterion:
` ln(T )
BIC = ln(σ̃`2) + .
T
• Needs a constant term? Check the sample mean.

• Estimation: least squares method or maximum likelihood method

• Model checking:

1. Residual: obs minus the fit, i.e. 1-step ahead forecast errors
at each time point.
2. Residual should be close to white noise if the model is ade-
quate. Use Ljung-Box statistics of residuals, but degrees of
freedom is m − g, where g is the number of AR coefficients
used in the model.

• Many software packages available, e.g. SCA, Splus, Eviews, SAS,

SPSS, etc.

4
Example: US GNP growth rate series revisited.
R demonstration:
> setwd("C:/teaching/bs41202")
> library(fSeries)
> da=[Link]("[Link]")
> x=da[,1]
> plot(x,type=’l’) % Plot not shown in this handout.
> title(main=’Growth rate of U.S. GNP: 1947-1991’) % title of plot.
> acf(x,[Link]=12) % Compute ACF (not shown in this handout)
> pacf(x,[Link]=12) % Compute PACF (not shown in this handout)
> [Link](x,lag=10,type=’Ljung’)

Box-Ljung test

data: x
X-squared = 43.2345, df = 10, p-value = 4.515e-06

> m1=ar(x,method=’mle’) % Automatic AR fitting using AIC criterion.

> m1

Call:
ar(x = x, method = "mle")

Coefficients:
1 2 3 % An AR(3) is specified.
0.3480 0.1793 -0.1423

Order selected 3 sigma^2 estimated as 9.427e-05

> names(m1)
[1] "order" "ar" "[Link]" "[Link]" "aic"
[6] "[Link]" "[Link]" "partialacf" "resid" "method"
[11] "series" "frequency" "call" "[Link]"

> plot(m1$resid,type=’l’) % Plot residuals of the fitted model (not shown)

> [Link](m1$resid,lag=10,type=’Ljung’) % Model checking

Box-Ljung test

data: m1$resid
X-squared = 7.0808, df = 10, p-value = 0.7178

> m2=arima(x,order=c(3,0,0)) % Another approach with order given.

> m2

5
Call:
arima(x = x, order = c(3, 0, 0))

Coefficients:
ar1 ar2 ar3 intercept % Fitted model is
0.3480 0.1793 -0.1423 0.0077 % y(t)=0.348y(t-1)+0.179y(t-2)
s.e. 0.0745 0.0778 0.0745 0.0012 % -0.142y(t-3)+a(t),
% where y(t) = x(t)-0.0077

sigma^2 estimated as 9.427e-05: log likelihood = 565.84, aic = -1121.68

> names(m2)
[1] "coef" "sigma2" "[Link]" "mask" "loglik" "aic"
[7] "arma" "residuals" "call" "series" "code" "[Link]"
[13] "model"

> [Link](m2$residuals,lag=10,type=’Ljung’)

Box-Ljung test

data: m2$residuals
X-squared = 7.0169, df = 10, p-value = 0.7239

> plot(m2$residuals,type=’l’) % Residual plot

>tsdiag(m2) % obtain 3 plots of model checking (not shown in handout).

>
> p1=c(1,-m2$coef[1:3]) % Further analysis of the fitted model.
> roots=polyroot(p1)
> roots
[1] 1.590253+1.063882e+00i -1.920152-3.530887e-17i 1.590253-1.063882e+00i
> Mod(roots)
[1] 1.913308 1.920152 1.913308

> k=2*pi/acos(1.590253/1.913308)
> k
[1] 10.65638

> predict(m2,8) % Prediction 1-step to 8-step ahead.

$pred
Time Series:
Start = 177
End = 184
Frequency = 1
[1] 0.001236254 0.004555519 0.007454906 0.007958518
[5] 0.008181442 0.007936845 0.007820046 0.007703826

6
$se
Time Series:
Start = 177
End = 184
Frequency = 1
[1] 0.009709322 0.010280510 0.010686305 0.010688994
[5] 0.010689733 0.010694771 0.010695511 0.010696190

S-Plus demonstration
> module(finmetrics)
> gnp=scan(file=’[Link]’)

> plot(gnp,type=’l’)
> acf(gnp,[Link]=12)
Call: acf(x = gnp, [Link] = 12) % Plot not shown in the handout.

Autocorrelation matrix:
lag gnp
1 0 1.0000
2 1 0.3769
3 2 0.2539
4 3 0.0125
5 4 -0.0859
6 5 -0.1071
7 6 -0.0575
8 7 -0.0182
9 8 -0.0772
10 9 -0.0702
11 10 0.0104
12 11 -0.0230
13 12 -0.0967
> acf(gnp,[Link]=12,type=’partial’) % Compute PACF
Call: acf(x = gnp, [Link] = 12, type = "partial")

Partial Correlation matrix:

lag gnp
1 1 0.3769
2 2 0.1304
3 3 -0.1421
4 4 -0.0988
5 5 -0.0199
6 6 0.0325
7 7 0.0120
8 8 -0.1106
9 9 -0.0415

7
10 10 0.0981
11 11 -0.0370
12 12 -0.1533
> ord=ar(gnp,[Link]=10) % Perform order selection via AIC
> ord$aic
[1] 27.5691310 2.6081086 1.5895550 0.0000000 0.2734771 2.2034466
[7] 4.0171066 5.9916210 5.8264833 7.5230025 7.8223499
> ord$order
[1] 3

> m1=[Link](gnp,model=list(order=c(3,0,0))) %This fit misses the mean.

> summary(m1)
Call: [Link](x = gnp, model = list(order = c(3, 0, 0)))
Method: Maximum Likelihood with likelihood conditional on 3 observati
ons

ARIMA order: 3 0 0

Value Std. Error t-value % No intercept because the program assumes it is zero.
ar(1) 0.45420 0.07597 5.9780
ar(2) 0.26680 0.08095 3.2960
ar(3) -0.03817 0.07597 -0.5024

Variance-Covariance Matrix:
ar(1) ar(2) ar(3)
ar(1) 0.005771926 -0.002566306 -0.001441892
ar(2) -0.002566306 0.006552753 -0.002566306
ar(3) -0.001441892 -0.002566306 0.005771926

Estimated innovations variance: 0.0001

Optimizer has converged

Convergence Type: relative function convergence
AIC: -1085.0397

> x=gnp-mean(gnp) % Remove sample mean.

> m1=[Link](x,model=list(order=c(3,0,0)))
> summary(m1)
Call: [Link](x = x, model = list(order = c(3, 0, 0)))
Method: Maximum Likelihood with likelihood conditional on 3 observati
ons

ARIMA order: 3 0 0

Value Std. Error t-value

ar(1) 0.3509 0.07523 4.664 % Fitted model is

8
ar(2) 0.1809 0.07863 2.301 % x(t)=0.351x(t-1)+0.181x(t-2)-0.144x(t-3)+a(t).
ar(3) -0.1443 0.07523 -1.919

Variance-Covariance Matrix:
ar(1) ar(2) ar(3)
ar(1) 0.0056599161 -0.001877448 -0.0007529176
ar(2) -0.0018774480 0.006182526 -0.0018774480
ar(3) -0.0007529176 -0.001877448 0.0056599161

Estimated innovations variance: 0.0001

Optimizer has converged

Convergence Type: relative function convergence
AIC: -1104.1574
> names(m1)
[1] "model" "[Link]" "method" "series" "aic"
[6] "loglik" "sigma2" "[Link]" "[Link]" "converged"
[11] "[Link]" "call"
> names(m1$model)
[1] "order" "ar" "ndiff"

> m1$model$ar
[1] 0.3509107 0.1809056 -0.1443412
>
> [Link](m1) % Model checking, plots not shown.

> p1=c(1,-m1$model$ar) % Further analysis of the fitted model.

> roots=polyroot(p1)
> roots
[1] 1.582837+1.057071e+000i -1.912355-6.609277e-017i
[3] 1.582837-1.057071e+000i
> Mod(roots)
[1] 1.903359 1.912355 1.903359
> k=2*pi/acos(1.582837/1.903359)
> k
[1] 10.67098
>

> [Link](x,m1$model,8) % prediction

$mean:
[1] -0.00651901645 -0.00317061250 -0.00023632985 0.00028445018
[5] 0.00051471315 0.00026618912 0.00014546524 0.00002490612

$[Link]:
[1] 0.009779314 0.010363943 0.010782026 0.010784985 0.010785783
[6] 0.010791060 0.010791857 0.010792592

9
Another example: Quarterly U.S. unemployment rate from 1948
to 2006. Focus on the change in umeployment rate.

Moving-average (MA) model

Model with finite time lags of memory!
Some daily stock returns have minor serial correlations. Can be
modeled as MA or AR models.
MA(1) model

• Form: rt = µ + at − θat−1

• Stationarity: always stationary.

• Mean (or expectation): E(rt) = µ

• Variance: Var(rt) = (1 + θ2)σa2.

• Autocovariance:

1. Lag 1: Cov(rt, rt−1) = −θσa2

2. Lag `: Cov(rt, rt−`) = 0 for ` > 1.

Thus, rt is not related to rt−2, rt−3, · · ·.

−θ
• ACF: ρ1 = 1+θ2
, ρ` = 0 for ` > 1.
Finite memory! MA(1) models do not remember what happen
two time periods ago.

• Forecast (at origin t = n):

10
 Changes in quarterly unemployment rates: 1948−2006
0.4
0.2
qq

0.0
−0.2

0 50 100 150

Index

Figure 1: Changes in U.S. quarterly unemployment rate from 1948 to 2006.

11
 1. 1-step ahead: r̂n(1) = µ − θan. Why? Because at time n,
an is known, but an+1 is not.
2. 1-step ahead forecast error: en(1) = an+1 with variance σa2.
3. Multi-step ahead: r̂n(`) = µ for ` > 1.
Thus, for an MA(1) model, the multi-step ahead forecasts
are just the mean of the series. Why? Because the model
has memory of 1 time period.
4. Multi-step ahead forecast error:

en(`) = an+` − θan+`−1

5. Variance of multi-step ahead forecast error:

(1 + θ2)σa2 = variance of rt.

• Invertibility:

– Concept: rt is a proper linear combination of at and the past

observations {rt−1, rt−2, · · ·}.
– Why is it important? It provides a simple way to obtain the
shock at.
For an invertible model, the dependence of rt on rt−` con-
verges to zero as ` increases.
– Condition: |θ| < 1.
– Invertibility of MA models is the dual property of stationarity
for AR models.

12
MA(2) model

• Form: rt = µ + at − θ1at−1 − θ2at−2. or

rt = µ + (1 − θ1B − θ2B 2)at.

• Stationary with E(rt) = µ.

• Variance: Var(rt) = (1 + θ12 + θ22)σa2.

• ACF: ρ2 6= 0,but ρ` = 0 for ` > 2.

• Forecasts go the the mean after 2 periods.

Building an MA model

• Specification: Use sample ACF

Sample ACFs are all small after lag q for an MA(q) series. (See
test of ACF.)

• Constant term? Check the sample mean.

• Estimation: use maximum likelihood method

– Conditional: Assume at = 0 for t ≤ 0

– Exact: Treat at with t ≤ 0 as parameters, estimate them to
obtain the likelihood function.

Exact method is preferred, but it is more computing intensive.

• Model checking: examine residuals (to be white noise)

13
 • Forecast: use the residuals as {at} (which can be obtained from
the data and fitted parameters) to perform forecasts.

Example:Daily log return of the value-weighted index

R demonstration
> setwd("C:/teaching/bs41202")
> library(fSeries)
> da=[Link]("[Link]")
> dim(da)
[1] 10194 4

> vw=log(1+da[,3])*100 % Compute percentage log returns of the vw index.

> acf(vw,[Link]=10) % ACF plot is not shon in this handout.
> m1=arima(vw,order=c(0,0,1)) % fits an MA(1) model
> m1

Call:
arima(x = vw, order = c(0, 0, 1))

Coefficients:
ma1 intercept
0.1465 0.0396 % The model is vw(t) = 0.0396+a(t)+0.1465a(t-1).
s.e. 0.0099 0.0100

sigma^2 estimated as 0.7785: log likelihood = -13188.48, aic = 26382.96

> tsdiag(m1)
> predict(m1,5)
$pred
Time Series:
Start = 10195
End = 10199
Frequency = 1
[1] 0.05036298 0.03960887 0.03960887 0.03960887 0.03960887

$se
Time Series:
Start = 10195
End = 10199
Frequency = 1
[1] 0.8823290 0.8917523 0.8917523 0.8917523 0.8917523

S-Plus demonstration
> vw=d6202[,3] % Identify the vw-index returns.
> lnvw=log(1+vw) % compute log returns.

14
> acf(lnvw,[Link]=10) % ACF plot i snot shown in this handout.
Call: acf(x = lnvw, [Link] = 10)

Autocorrelation matrix:
lag lnvw
1 0 1.0000
2 1 0.1402
3 2 -0.0120
4 3 -0.0027
5 4 0.0029
6 5 0.0075
7 6 -0.0149
8 7 -0.0066
9 8 -0.0034
10 9 -0.0085
11 10 -0.0074
> length(lnvw)
[1] 10194
> x1=rep(1,10194) % Create a constant to handle non-zero mean
> m1=[Link](lnvw,xreg=x1,model=list(order=c(0,0,1)))
> summary(m1)
Call: [Link](x = lnvw, model = list(order = c(0, 0, 1)), xreg = x1)
Method: Maximum Likelihood with likelihood conditional on 0 observati
ons

ARIMA order: 0 0 1

Value Std. Error t-value

ma(1) -0.1465000 0.009797 -14.96
x1 0.0003962 NA NA % Model is vw = .000396+a(t)+0.1465a(t-1)

Variance-Covariance Matrix:
ma(1)
ma(1) 0.00009599039

Estimated innovations variance: 0.0001

Optimizer has converged

Convergence Type: relative function convergence
AIC: -67509.2476
> [Link](m1) % Plots not shown in this handout.

> [Link](lnvw,model=m1$model,6)
$mean:
[1] 0.0001581654 0.0000000000 0.0000000000 0.0000000000 0.0000000000
[6] 0.0000000000 % Need to add the constant 0.000396 to the forecast.

15
$[Link]:
[1] 0.008830056 0.008924361 0.008924361 0.008924361 0.008924361
[6] 0.008924361

Mixed ARMA model: A compact form for flexible models.

Focus on the ARMA(1,1) model for
1. simplicity

2. useful for understanding GARCH models in Ch. 3 for volatility

modeling.
ARMA(1,1) model
• Form: (1 − φ1B)rt = φ0 + (1 − θB)at or

rt = φ1rt−1 + φ0 + at − θ1at−1.

A combination of an AR(1) on the LHS and an MA(1) on the

RHS.

• Stationarity: same as AR(1)

• Invertibility: same as MA(1)

φ0
• Mean: as AR(1), i.e. E(rt) = 1−φ1

• Variance: given in the text

• ACF: Satisfies ρk = φ1ρk−1 for k > 1, but

ρ1 = φ1 − [θ1σa2/Var(rt)] 6= φ1.

This is the difference between AR(1) and ARMA(1,1) models.

16
 • PACF: does not cut off at finite lags.

Building an ARMA(1,1) model

• Specification: use EACF or AIC

• What is EACF? How to use it? [See text].

• Estimation: cond. or exact likelihood method

• Model checking: as before

• Forecast: MA(1) affects the 1-step ahead forecast. Others are

similar to those of AR(1) models.

Three model representations:

• ARMA form: compact, useful in estimation and forecasting

• AR representation: (by long division)

rt = φ0 + at + π1rt−1 + π2rt−2 + · · ·

It tells how rt depends on its past values.

• MA representation: (by long division)

rt = µ + at + ψ1at−1 + ψ2at−2 + · · ·

It tells how rt depends on the past shocks.

For a stationary series, ψi converges to zero as i → ∞. Thus, the

effect of any shock is transitory.
17
The MA representation is particularly useful in computing variances
of forecast errors.
For a `-step ahead forecast, the forecast error is

en(`) = an+` + ψ1an+`−1 + · · · + ψ`−1an+1.

The variance of forecast error is

Var[en(`)] = (1 + ψ12 + · · · + ψ`−1

2
)σa2.

Unit-root Nonstationarity
Random walk

• Form pt = pt−1 + at

• Unit root? It is an AR(1) model with coefficient φ1 = 1.

• Nonstationary: Why? Because the variance of rt diverges to

infinity as t increases.

• Strong memory: sample ACF approaches 1 for any finite lag.

• Repeated substitution shows

∞ ∞
pt = at−i = ψiat−i
X X

i=0 i=0

where ψi = 1 for all i. Thus, ψi does not converge to zero. The

effect of any shock is permanent.

Random walk with drift

• Form: pt = µ + pt−1 + at, µ 6= 0.

18
 • Has a unit root

• Nonstationary

• Strong memory

• Has a time trend with slope µ. Why?

differencing

• 1st difference: rt = pt − pt−1

If pt is the log price, then the 1st difference is simply the log
return. Typically, 1st difference means the “change” or “incre-
ment” of the original series.

• Seasonal difference: yt = pt − pt−s, where s is the periodicity,

e.g. s = 4 for quarterly series and s = 12 for monthly series.
If pt denotes quarterly earnings, then yt is the change in earning
from the same quarter one year before.

Meaning of the constant term in a model

• MA model: mean

• AR model: related to mean

• 1st differenced: time slope, etc.

Practical implication in financial time series

Example: Monthly log returns of General Electrics (GE) from 1926
to 1999 (74 years)
19
Sample mean: 1.04%, std(µ̂) = 0.26
Very significant!
is about 12.45% a year
$1 investment in the beginning of 1926 is worth

• annual compounded payment: $5907

• quarterly compounded payment: $8720

• monthly compounded payment: $9570

• Continuously compounded?

Unit-root test (for self study only)

Let pt be the log price of an asset. To test that pt is not predictable
(i.e. has a unit root), two models are commonly employed:

pt = φ1pt−1 + et
pt = φ0 + φ1pt−1 + et.

The hypothesis of interest is Ho : φ1 = 1 vs Ha : φ1 < 1.

Dickey-Fuller test is the usual t-ratio of the OLS estimate of φ1 being
1. This is the DF unit-root test. The t-ratio, however, has a non-
standard limiting distribution.
Let ∆pt = pt − pt−1. Then, the augmented DF unit-root test for an
AR(p) model is based on
p−1
∆pt = ct + βpt−1 + φi∆pt−i + et.
X

i=1

20
The t-ratio of the OLS estimate of β is the ADF unit-root test sta-
tistic. Again, the statistic has a non-standard limiting distribution.
Example: Consider the log series of U.S. quaterly real GDP series
from 1947.I to [Link]. (Federal Reserve Bank of St. Louis).
R demonstration
> help(UnitrootTests) % See the tests available

> adfTest(gdp,lags=4,type=c("c")) % Assume an AR(4) model.

Title:
Augmented Dickey-Fuller Test

Test Results:
PARAMETER:
Lag Order: 4
STATISTIC:
Dickey-Fuller: -1.1199
P VALUE:
0.6397 % Cannot reject a unit root.

*** A more careful analysis

> x=diff(gdp) % Take the first difference
> ord=ar(x) % Find AR models for x series.
> ord

Call:
ar(x = x)

Coefficients:
1 2 3 4
0.3021 0.1311 -0.0856 -0.1060

Order selected 4 sigma^2 estimated as 8.592e-05

> adfTest(gdp,lags=5,type=c("c"))

Title:
Augmented Dickey-Fuller Test

Test Results:
PARAMETER:
Lag Order: 5
STATISTIC:

21
 Dickey-Fuller: -1.1339
P VALUE:
0.6345

S-Plus demonstration
> da=[Link]("[Link]")
> dim(da)
[1] 236 4
> plot(da[,4],type=’l’)
> module(finmetrics)

> gdp=log(da[,4])
> plot(gdp,type=’l’)

> x=diff(gdp) % take the first difference

> ord=ar(x)
> ord
$order:
[1] 4

> adf=unitroot(gdp,trend=’c’,lags=5,method=’adf’)
> adf

Test for Unit Root: Augmented DF Test

Null Hypothesis: there is a unit root

Type of Test: t-test
Test Statistic: -1.12
P-value: 0.7083

Coefficients:
lag1 lag2 lag3 lag4 lag5 constant
-0.0012 0.2954 0.1358 -0.0864 -0.1108 0.0168

Degrees of freedom: 231 total; 225 residual

Residual standard error: 0.009283

22
 Lecture Note: Analysis of Financial Time Series
Spring 2007, Ruey S. Tsay

Seasonal Time Series

• TS with periodic patterns, e.g., quarterly earnings

• Useful in pricing weather-related derivatives

• Useful in analysis of transactions data (high-frequency data),

e.g., U-shaped pattern in intraday data
Example Demand of electricity of a manufacturing sector of U.S.
from 1972 to 1993. The data are logged usage on the 15th day of
each month. See Figure 1.
Example Quarterly earnings of Johnson & Johnson
See the time plot, Figures 2 and 3, and sample ACFs
Multiplicative model
Airline model (for quarterly series)
• Form:

rt − rt−1 − rt−4 + rt−5 = at − θ1at−1 − θ4at−4 + θ1θ4at−5

or
(1 − B)(1 − B 4)rt = (1 − θ1B)(1 − θ4B 4)at

• Define the differenced series wt as

wt = rt − rt−1 − rt−4 + rt−5 = (rt − rt−1) − (rt−4 − rt−5).

1
 Demand of Electricity: 15th day of every month, 1972−1993
••
•
10.0

••
•
•
••
• •• ••
•• • • ••
•• •• • • •
•• ••• •••
• • •••
• • • • • •
9.8

• ••
•• • •• •• ••
• • •• ••
• ••
••• • • •• ••• • •
•
• • • • ••
•• •• • •• • •
• •• ••• •
•• ••• •• • •• •••• ••
• • • • ••
9.6

• • •• • • ••
• • • • ••
• ••
•• • •• • • • ••
•
• ••• • •• • • • •
• •• • • • •• •••• •
• •• •
• • • • ••••• •••• •
••
• •
•
•• • •
• •
9.4

• • •
•• • •
• •• • • •• •• ••••
• • • •
• • • • • •• ••
• •••• • • • • •• ••
• ••• •
••• ••• •
•

1975 1980 1985 1990

Figure 1: Time plot of electricity demand of an industrial sector: 15th day of each month
from1972 to 1993.

Quarterly earnings of JNJ: 1960−1980

15
10
x
5
0

1960 1965 1970 1975 1980

Time

Figure 2: Time plot of quarterly earnings of Johnson and Johnson: 1960-1980

2
 1 2
y
0

1960 1965 1970 1975 1980

Time

Figure 3: Time plot of quarterly logged earnings of Johnson and Johnson: 1960-1980

It is called regular and seasonal differenced series.

• ACF of wt has a nice symmetric structure (see the text), i.e.

ρs−1 = ρs+1 = ρ1ρs. Also, ρ` = 0 for ` > s + 1.

• This model is widely applicable to many many seasonal time

series.

• Multiplicative model means that the regular and seasonal depen-

dences are roughly orthogonal to each other.

• Forecasts: exhibit the same pattern as the observed series. See

Figure 4.

Example Detailed analysis of J&J earnings.

R Demonstration: output edited.
3
 3
2
1
0
−1

0 20 40 60 80

Figure 4: Forecast plot for the quarterly earnings of Johnson and Johnson. Data: 1960-1980,
Forecasts: 1981-82.

> library(fSeries)
> setwd("C:/teaching/bs41202")
> x=ts(scan("[Link]"),frequency=4,start=c(1960,1)) % Load data into a time series object.
> plot(x,type=’l’) % Plot data with calendar time
> y=log(x) % Natural log transformation
> plot(y,type=’l’) % plot data
> points(y) % put circles on data points.
> par(mfcol=c(2,1)) % two plots per page
> acf(y,[Link]=16)
> y1=[Link](y) % Creates a sequence of data in R
> acf(y1,[Link]=16)
> dy1=diff(y1) % regular difference
> acf(dy1,[Link]=16)
> sdy1=diff(dy1,4) % seasonal difference
> acf(sdy1,[Link]=12)
> m1=arima(y1,order=c(0,1,1),seasonal=list(order=c(0,1,1),period=4)) % Airline
% model in R.
> m1

Call:
arima(x = y1, order = c(0, 1, 1), seasonal = list(order = c(0, 1, 1), period = 4))

4
Coefficients:
ma1 sma1
-0.6809 -0.3146 % The fitted model is (1-B^4)(1-B)R(t) =
s.e. 0.0982 0.1070 % (1-0.68B)(1-0.31B^4)a(t), var[a(t)] = 0.00793.

sigma^2 estimated as 0.00793: log likelihood = 78.38, aic = -150.75

> par(mfcol=c(1,1)) % One plot per page
> tsdiag(m1) % Model checking
> f1=predict(m1,8) % prediction
> names(f1)
[1] "pred" "se"
> f1
$pred % Point forecasts
Time Series:
Start = 85
End = 92
Frequency = 1
[1] 2.905343 2.823891 2.912148 2.581085 3.036450 2.954999 3.043255 2.712193

$se % standard errors of point forecasts

Time Series:
Start = 85
End = 92
Frequency = 1
[1] 0.08905414 0.09347895 0.09770358 0.10175295 0.13548765 0.14370550
[7] 0.15147817 0.15887102

> s1=c(y1,f1$pred) % Join data with forecasts

> lcl=c(y1,f1$pred-2*f1$se) % Lower limit for 95% interval
> ucl=c(y1,f1$pred+2*f1$se) % Upper limit for 95% interval
> max(ucl)
[1] 3.346211
> min(y1)
[1] -0.8209806
> plot(s1,type=’l’,ylim=c(-1,3.5)) % Forecast plot
> lines(1:92,ucl,lty=2)
> lines(1:92,lcl,lty=2)

S-Plus Demonstration: output edited.

> x=ts(scan(’[Link]’),frequency=4,start=c(1960,1)) % Load data into Splus
> plot(x,type=’l’) % Plot the data
> title(main=’Quarterly earnings of JNJ: 1960-1980’) % title of the plot
> y=log(x) % natural log transformation
> plot(y,type=’l’)
> par(mfcol=c(2,1)) % put two plots on a page
> acf(y,[Link]=16) % 16 lags of ACF

5
> acf(diff(y),[Link]=16)

> y1=[Link](y) % creates a sequence in Splus, not a time-series object.

> acf(y1,[Link]=16)

Autocorrelation matrix:
lag y1
1 0 1.0000
2 1 0.9566
3 2 0.9260 % Indicates 1st difference is needed.
4 3 0.8978
5 4 0.8723
6 5 0.8285
....
17 16 0.4578
> acf(diff(y1),[Link]=16)

> dy1=diff(y1)
> sdy1=diff(dy1,4)
> acf(sdy1,[Link]=12)
> tra=mean(sdy1)/sqrt(var(sdy1)/length(sdy1)) % Compute t-ratio of the mean.
> tra
[1] 0.3101582

> air=list(list(order=c(0,1,1)),list(order=c(0,1,1),period=4)) % Define the

% airline model.
> m1=[Link](y1,model=air) % estimation

> summary(m1)
Call: [Link](x = y1, model = air)
Method: Maximum Likelihood with likelihood conditional on 5 observations

Multiplicative ARIMA model --

Model component 1
ARIMA order: 0 1 1

Model component 2
ARIMA order: 0 1 1
Period: 4

Value Std. Error t-value % Fitted model (1-B^4)(1-B)R(t) =

ma(1) 0.6809 0.08582 7.934 % (1-0.68B)(1-0.31B^4)a(t),
ma(4) 0.3146 0.11120 2.828 % with var[a(t)] = 0.0079.

Variance-Covariance Matrix:
ma(1) ma(4)

6
ma(1) 0.007364341 -0.002665339
ma(4) -0.002665339 0.012370336

Estimated innovations variance: 0.0079

Optimizer has converged

Convergence Type: relative function convergence
AIC: -152.7529

> [Link](m1) % Model checking

> f1=[Link](y1,model=m1$model,8) % Forecasts of the next two years

> names(f1)
[1] "mean" "[Link]"
> f1
$mean:
[1] 2.905343 2.823891 2.912148 2.581085 3.036450 2.954999 3.043255 2.712193

$[Link]:
[1] 0.08905406 0.09347891 0.09770357 0.10175299 0.13548748 0.14370536
[7] 0.15147807 0.15887096

% The following commands create a forecast plot.

> s1=c(y1,f1$mean) % Join forecasts with data
> lcl=c(y1,f1$mean-2*f1$[Link]) % Join data with lower limit
> ucl=c(y1,f1$mean+2*f1$[Link]) % Join data with upper limit
> par(mfcol=c(1,1)) % One plot per page
> length(s1)
[1] 92
> max(ucl) % Maximum and minimum are used to set limits on plot.
[1] 3.346211
> min(y1)
[1] -0.8209806
> plot(s1,type=’l’,ylim=c(-1,3.5))
> lines(1:92,ucl,lty=2)
> lines(1:92,lcl,lty=2)

Consider monthly series with period 12. Airline model becomes

(1 − B)(1 − B 12)rt = (1 − θ1B)(1 − θ12B 12)at.

What is the pattern of ACF?

7
Regression Models with Time Series Errors

• Has many applications

• Impact of serial correlations in regression is often overlooked.

It may introduce biases in estimates and in standard errors, re-
sulting in unreliable t-ratios.

• Detecting residual serial correlation: Use Q-stat instead of DW-

statistic, which is not sufficient!

• Joint estimation of all parameters is preferred.

• Proper analysis: see the illustration below.

Example. U.S. weekly interest rate data: 1-year and 3-year con-
stant maturity rates. Data are shown in Figure 5.
R Demonstration: output edited.
> library(fSeries)
> setwd("C:/teaching/bs41202")
> da=[Link]("[Link]") % load the data
> r1=da[,1] % 1-year rate
> r3=da[,2] % 3-year rate
> plot(r1,type=’l’) % Plot the data
> lines(1:1967,r3,lty=2)
> plot(r1,r3) % scatter plot of the two series

> m1=lm(r3~r1) % Fit a regression model with likelihood method.

> summary(m1)
Call:
lm(formula = r3 ~ r1)

Residuals:
Min 1Q Median 3Q Max
-1.812147 -0.402280 0.003097 0.402588 1.338746

Coefficients:

8
 15
percent
105

1970 1980 1990 2000

year

Figure 5: Time plots of U.S. weekly interest rates: 1-year constant maturity rate (solid line)
and 3-year rate (dashed line).

Estimate Std. Error t value Pr(>|t|)

(Intercept) 0.910687 0.032250 28.24 <2e-16 ***
r1 0.923854 0.004389 210.51 <2e-16 ***
---
Signif. codes: 0 ’***’ 0.001 ’**’ 0.01 ’*’ 0.05 ’.’ 0.1 ’ ’ 1

Residual standard error: 0.538 on 1965 degrees of freedom

Multiple R-Squared: 0.9575, Adjusted R-squared: 0.9575
F-statistic: 4.431e+04 on 1 and 1965 DF, p-value: < 2.2e-16

> acf(m1$residuals)
> c3=diff(r3)
> c1=diff(r1)
> plot(c1,c3)

> m2=lm(c3~c1) % Fit a regression with likelihood method.

> summary(m2)
Call:
lm(formula = c3 ~ c1)

Residuals:
Min 1Q Median 3Q Max

9
-0.3806040 -0.0333840 -0.0005428 0.0343681 0.4741822

Coefficients:
Estimate Std. Error t value Pr(>|t|)
(Intercept) 0.0002475 0.0015380 0.161 0.872
c1 0.7810590 0.0074651 104.628 <2e-16 ***
---
Signif. codes: 0 ’***’ 0.001 ’**’ 0.01 ’*’ 0.05 ’.’ 0.1 ’ ’ 1

Residual standard error: 0.06819 on 1964 degrees of freedom

Multiple R-Squared: 0.8479, Adjusted R-squared: 0.8478
F-statistic: 1.095e+04 on 1 and 1964 DF, p-value: < 2.2e-16

> acf(m2$residuals)
> plot(m2$residuals,type=’l’)

> m3=arima(c3,xreg=c1,order=c(0,0,1)) % Residuals follow an MA(1) model

> m3

Call:
arima(x = c3, order = c(0, 0, 1), xreg = c1)

Coefficients:
ma1 intercept c1 % Fitted model is
0.2115 0.0002 0.7824 % c3 = 0.0002+0.782c1 + a(t)+0.212a(t-1)
s.e. 0.0224 0.0018 0.0077 % with var[a(t)] = 0.00446.

sigma^2 estimated as 0.004456: log likelihood = 2531.84, aic = -5055.69

> acf(m3$residuals)
> tsdiag(m3)

> m4=arima(c3,xreg=c1,order=c(1,0,0)) % Residuals follow an AR(1) model.

> m4

Call:
arima(x = c3, order = c(1, 0, 0), xreg = c1)

Coefficients:
ar1 intercept c1 % Fitted model is
0.1922 0.0003 0.7829 % c3 = 0.0003 + 0.783c1 + a(t),
s.e. 0.0221 0.0019 0.0077 % a(t) = 0.192a(t-1)+e(t).

sigma^2 estimated as 0.004474: log likelihood = 2527.86, aic = -5047.72

S-Plus Demonstration
> module(finmetrics)

10
> da=[Link]("[Link]")
> dim(da)
[1] 1967 3
> r3=da[,2]
> r1=da[,1]
> plot(r1,type=’l’) % plot the data
> lines(1:1967,r3,lty=2)
> plot(r1,r3)
>
> m1=OLS(r3~r1) % Least-square regression
> summary(m1)

Call:
OLS(formula = r3 ~ r1)

Residuals:
Min 1Q Median 3Q Max
-1.8121 -0.4023 0.0031 0.4026 1.3387

Coefficients:
Value Std. Error t value Pr(>|t|)
(Intercept) 0.9107 0.0323 28.2380 0.0000 % Fitted model is
r1 0.9239 0.0044 210.5084 0.0000 % r3=0.911+0.924r1 + e

Regression Diagnostics:

R-Squared 0.9575 % R-square is 96%!!! Any good?

Adjusted R-Squared 0.9575
Durbin-Watson Stat 0.0190 % What is the ‘‘ideal’’ value of DW?

Residual Diagnostics:
Stat P-Value
Jarque-Bera 9.0032 0.0111
Ljung-Box 42303.0824 0.0000

Residual standard error: 0.538 on 1965 degrees of freedom

F-statistic: 44310 on 1 and 1965 degrees of freedom, the p-value is 0
> names(m1)
[1] "R" "coef" "[Link]" "fitted" "residuals" "assign"
[7] "contrasts" "[Link]" "terms" "call"

> acf(m1$residuals) % ACF of residuals

> c3=diff(r3) % Take the first difference

> c1=diff(r1)
> m2=OLS(c3~c1) % LS regression of the differenced series

11
> summary(m2)

Call:
OLS(formula = c3 ~ c1)

Residuals:
Min 1Q Median 3Q Max
-0.3806 -0.0334 -0.0005 0.0344 0.4742

Coefficients:
Value Std. Error t value Pr(>|t|)
(Intercept) 0.0002 0.0015 0.1609 0.8722 % c3 = 0.002+0.781c1+e
c1 0.7811 0.0075 104.6283 0.0000

Regression Diagnostics:

R-Squared 0.8479
Adjusted R-Squared 0.8478
Durbin-Watson Stat 1.6158

Residual Diagnostics:
Stat P-Value
Jarque-Bera 1508.0683 0.0000
Ljung-Box 230.5767 0.0000

Residual standard error: 0.06819 on 1964 degrees of freedom

F-statistic: 10950 on 1 and 1964 degrees of freedom, the p-value is 0
> acf(m2$residuals) % Plot not shown

> m3=[Link](c3,xreg=c1,model=list(order=c(0,0,1))) % Regression model with

% time-series errors
> summary(m3)
Call: [Link](x = c3, model = list(order = c(0, 0, 1)), xreg = c1)
Method: Maximum Likelihood with likelihood conditional on 0 observations

ARIMA order: 0 0 1

Value Std. Error t-value

ma(1) -0.2115 0.02204 -9.594 % Fitted model
c1 0.7824 NA NA % c3 = 0.782c1 + a(t)+0.212a(t-1).
% Because [Link] assumes mean of c3
Variance-Covariance Matrix: % is zero, there is no intercept.
ma(1)
ma(1) 0.0004858961

Estimated innovations variance: 0.0045

12
Optimizer has converged
Convergence Type: relative function convergence
AIC: -5059.6702
> [Link](m3) & model checking

Long-memory models

• Meaning? ACF decays to zero very slowly!

• Example: ACF of squared or absolute log returns

ACFs are small, but decay very slowly.

• How to model long memory? Use “fractional” difference: namely,

(1 − B)drt, where −0.5 < d < 0.5.

• Importance? In theory, Yes. In practice, yet to be determined.

Summary of the chapter

• Sample ACF ⇒ MA order

• Sample PACF ⇒ AR order

• Some packages have “automatic” procedure to select a simple

model for “conditional mean” of a FTS, e.g., SCA uses “iarima”
for ARIMA models and R uses “ar” for AR models.

• Check a fitted model before forecasting, e.g. residual ACF and

hetroscedasticity (chapter 3)

• Interpretation of a model, e.g. constant term &

13
 For an AR(1) with coefficient φ1, the speed of mean reverting as
measured by half-life is
ln(0.5)
k= .
ln(φ1)
For an MA(q) model, forecasts revert to the mean in q steps.

• Make proper use of regression models with time series errors, e.g.
regression with AR(1) residuals
Perform a joint estimation instead of using any two-step proce-
dure, e.g. Cochrane-Orcutt (1949).
Example: Is there a Friday effect on asset returns?
If a daily market index is used, serial correlation may exist.
See Exercise 8 of Chapter 2.

• Basic properties of a random-walk model

• Multiplicative seasonal models, especially the so-called airline

model.

14