Financial Econometrics

Simon Trimborn

Introduction to Time-Series Analysis

Executive MSc in International Finance (MIF)

Simon Trimborn (Introduction to Time-Series Analysis) Financial Econometrics Executive MSc in International Finance (MIF) 1 / 50
Outline for Week 4

Introduction
Stationarity and Autocorrelation
AR Processes
MA and ARMA Processes
Integrated Processes
Forecasting
Unit root testing
Exercises

Simon Trimborn (Introduction to Time-Series Analysis) Financial Econometrics Executive MSc in International Finance (MIF) 2 / 50
Please register your attendance

Simon Trimborn (Introduction to Time-Series Analysis) Financial Econometrics Executive MSc in International Finance (MIF) 3 / 50
Time Series Analysis

Financial time series analysis (TSA) is concerned with modelling, estimating, analyzing and
forecasting the behaviour of financial variables over time.
Distinguish univariate time series analysis, which tries to forecast individual time series
from its own past, from multivariate time series analysis, where different time series are
explained by their own and each other’s past.
A time series {yt , t = 1, 2, . . . , T } is a collection of subsequent observations on a particu-
lar variable.
We will focus on univariate TSA

Simon Trimborn (Introduction to Time-Series Analysis) Financial Econometrics Executive MSc in International Finance (MIF) 4 / 50
Purpose

TSA is regression but with time series data

Example: Say a time series is found to depend on the last observation:

yt = ϕyt −1 + ϵt ;

this is the first-order autoregressive model (AR(1)).

The purpose of the analysis may be:
▶ Forecasting
▶ Analysis of the effect of “shocks” (impulse responses): ∂ yt +s /∂ϵt ;

Simon Trimborn (Introduction to Time-Series Analysis) Financial Econometrics Executive MSc in International Finance (MIF) 5 / 50
Example: Dow Jones Index
Time period: 02.01.2000 until 26.07.2022

Dow Jones Index

DJI
35000

30000

25000

20000

15000

10000

0 4 8 2 6 0
200 200 200 201 201 202
Date

Simon Trimborn (Introduction to Time-Series Analysis) Financial Econometrics Executive MSc in International Finance (MIF) 6 / 50
Log return of Dow Jones Index

Return of Dow Jones Index

0.10 RDJI

0.05

0.00

0.05

0.10
0 4 8 2 6 0
200 200 200 201 201 202
Date

Why do we consider the log return?

Simon Trimborn (Introduction to Time-Series Analysis) Financial Econometrics Executive MSc in International Finance (MIF) 7 / 50
Stationarity and the ACF
Clearly: the CLRM assumptions are no longer valid
Hence: We need new assumptions / properties
Saying something about yT +s based on {y1 , . . . , yT } requires some kind of time homo-
geneity ⇒ stationarity.
A process is weakly (second-order / covariance-) stationary if its mean, variance and auto-
covariances of yt do not change over time:

E (yt ) = µ,
var(yt ) = γ0 = σ 2 ,
cov(yt , yt −k ) = γk ,

where µ, σ 2 and γk do not depend on t. Note that γ−k = γk .

The autocorrelation function (acf) is the normalized version

cov(yt , yt −k ) γk
τk = p = , k = 0, 1, . . .
var(yt )var(yt −k ) γ0

with τ−k = τk . (Usually denoted by ρk instead of τk .)

Simon Trimborn (Introduction to Time-Series Analysis) Financial Econometrics Executive MSc in International Finance (MIF) 8 / 50
Mind the similarity with CLRM assumptions!

1 The model is linear in the parameter vector β .

2 E [εi ] = 0 ∀i
 
3 Var [εi ] = E ε2i = σ 2 ∀i
4 Cov [εi , εj ] = E [εi εj ] = 0, ∀i ̸= j
5 On X :
▶ X is deterministic with rk (X ) = k < T
▶ limT →∞ T1 X ′ X ≡ Q , where Q is a finite, nonsingular matrix.
6 ϵt ∼ N (0, σ 2 ) (normality)

Simon Trimborn (Introduction to Time-Series Analysis) Financial Econometrics Executive MSc in International Finance (MIF) 9 / 50
White Noise I

Important example of a stationary process: white noise, with zero mean and zero autoco-
variances:

E (ϵt ) = 0,
var(ϵt ) = σ2 ,
cov(ϵt , ϵt −k ) = 0, k = 1, 2, . . .

Notation ϵt emphasizes similarity to regression errors.

White noise is unpredictable; building block for other processes that are predictable.

Simon Trimborn (Introduction to Time-Series Analysis) Financial Econometrics Executive MSc in International Finance (MIF) 10 / 50
White Noise II

White noise
3 4

3
2

2
1
1
0
0

1 1

2 2

3
3 0 200 400 600 800 1000
0 200 400 600 800 1000

Simon Trimborn (Introduction to Time-Series Analysis) Financial Econometrics Executive MSc in International Finance (MIF) 11 / 50
The Autocorrelation Function (ACF)

From a given time series {y1 , . . . , yT }, we may estimate (under the assumption of station-
arity) µ, γk , and τk by

T T
1 X 1 X γ
bk
µ
b= yt , γ
bk = (yt − ȳ )(yt −k − ȳ ), τk =
b .
T
t =1
T
t =k +1
γ
b0

The sequence {b τ k , k = 1, 2, . . . , m}, with m < T , is called the autocorrelation function

(acf) or correlogram.
For an uncorrelated and homoskedastic stationary series (τk = 0, k ̸= 0), one can show
that for large T , √
τ k ∼ N (0, 1),
Tb k = 1, 2, . . .

This implies√that a sample autocorrelation is significantly different from zero if it lies outside
the ±1.96/ T interval.

Simon Trimborn (Introduction to Time-Series Analysis) Financial Econometrics Executive MSc in International Finance (MIF) 12 / 50
Q-Statistics

Test for autocorrelation based on Ljung-Box Q-statistic:

m
∗
X τ 2k
b
Qm = T (T + 2) .
T −k
k =1

∗
Qm ∼ χ2m under H0 : τ1 = . . . = τm = 0.
▶ Will reject under H1 : at least one τk , 1 ≤ k ≤ m is non-zero.

Simon Trimborn (Introduction to Time-Series Analysis) Financial Econometrics Executive MSc in International Finance (MIF) 13 / 50
Partial ACF (PACF)

Software packages also give the partial autocorrelation function (pacf) b τ kk , k = 1, 2, . . .,

measuring correlation between yt and yt −k , controlling for effect of intermediate lags (esti-
mated coefficient of yt −k in AR(k ), see below).
For a stationary process, theoretically one can show that the acf and pacf converge to zero
at a geometric (exponential) rate, as k increases.
If the acf does not seem to converge at all, or too slowly (linearly), then this is an indication
of possible nonstationarity.

Simon Trimborn (Introduction to Time-Series Analysis) Financial Econometrics Executive MSc in International Finance (MIF) 14 / 50
ACF / PACF for Dow Jones Index
Autocorrelation function of DJI
1.0

0.5

0.0
ACF

0.5

1.0
0 5 10 15 20 25 30 35

Partial autocorrelation function of DJI

1.0

0.5
PACF

0.0

0.5

1.0
0 5 10 15 20 25 30 35
Lags

Simon Trimborn (Introduction to Time-Series Analysis) Financial Econometrics Executive MSc in International Finance (MIF) 15 / 50
ACF / PACF for log returns Dow Jones Index I
Autocorrelation function of RDJI
1.0

0.5

0.0
ACF

0.5

1.0
0 5 10 15 20 25 30 35

Partial autocorrelation function of RDJI

1.0

0.5
PACF

0.0

0.5

1.0
0 5 10 15 20 25 30 35
Lags

Simon Trimborn (Introduction to Time-Series Analysis) Financial Econometrics Executive MSc in International Finance (MIF) 16 / 50
ACF / PACF for log returns Dow Jones Index II
Autocorrelation function of RDJI
0.10

0.05

0.00
ACF
0.05

0.10
0 5 10 15 20 25 30 35

Partial autocorrelation function of RDJI

0.10

0.05
PACF

0.00

0.05

0.10
0 5 10 15 20 25 30 35
Lags

How are ACF / PACF useful?

Answer: They help to determine the structure of the time series models. So, we need to talk
time series models next
Simon Trimborn (Introduction to Time-Series Analysis) Financial Econometrics Executive MSc in International Finance (MIF) 17 / 50
The AR(1) Process
The AR(1) model is
yt = µ + ϕ1 yt −1 + ϵt ,
where ϵt is a white noise process.
Process is only stationary if −1 < ϕ1 < 1. Why?
▶ Mean of yt should be solved from

E (yt ) = µ + ϕ1 E (yt −1 ) + E (ut ) = µ + ϕ1 E (yt ),

which implies, provided that ϕ1 ̸= 1,

µ
E (yt ) = .
1 − ϕ1
▶ Because cov(ut , yt −1 ) = 0, the variance is solved from
2 2 2
var(yt ) = ϕ1 var(yt −1 ) + var(ut ) + 2ϕ1 var(yt −1 , ut ) = ϕ1 var(yt ) + σ ,

which has a positive solution only if |ϕ1 | < 1:

σ2
var(yt ) = .
1 − ϕ21
Simon Trimborn (Introduction to Time-Series Analysis) Financial Econometrics Executive MSc in International Finance (MIF) 18 / 50
ACF / PACF AR(1)
Autocorrelation function of AR(1) process
1.0

0.5

0.0
ACF

0.5

1.0
0 5 10 15 20 25 30 35

Partial autocorrelation function of AR(1) process

1.0

0.5
PACF

0.0

0.5

1.0
0 5 10 15 20 25 30 35
Lags

Simon Trimborn (Introduction to Time-Series Analysis) Financial Econometrics Executive MSc in International Finance (MIF) 19 / 50
ACF of Stationary AR(1)

yt = µ + ϕ1 yt −1 + ϵt .

First-order autocovariance:

γ1 = cov(yt , yt −1 ) = ϕ1 cov(yt −1 , yt −1 ) + cov(µ + ϵt , yt −1 ) = ϕ1 γ0 ,

so τ1 = γ1 /γ0 = ϕ1 .
For second-order autocovariance, use backward substitution:

yt = ϕ1 yt −1 + µ + ϵt
= ϕ21 yt −2 + ϕ1 (µ + ϵt −1 ) + µ + ϵt ,

which implies γ2 = cov(yt , yt −2 ) = ϕ21 γ0 , and hence τ2 = ϕ21 .

Analogously,
γk
τk = = ϕk1 , k = 1, 2, . . .
γ0

Simon Trimborn (Introduction to Time-Series Analysis) Financial Econometrics Executive MSc in International Finance (MIF) 20 / 50
Random Walk and Explosive Process

The AR(1) process is non-stationary if ϕ1 = 1:

▶ y = µ + y
t t −1 + ϵt , called a random walk with drift. It satisfies

E (yt ) = y 0 + µt ,
var(yt ) = σ2 t ,
p
corr(yt , yt −k ) = (t − k )/t .

▶ Correlogram stays close to 1, and decreases slowly, approximately linearly.

▶ Example of an I (1) process: yt is non-stationary, ∆yt = µ + ut is stationary.
Also non-stationary if ϕ1 > 1. Implies explosive process, with mean and variance increas-
ing very fast (exponentially). May be used for shorter periods to describe stock market
bubble or hyperinflation.

Simon Trimborn (Introduction to Time-Series Analysis) Financial Econometrics Executive MSc in International Finance (MIF) 21 / 50
ACF / PACF Random Walk
Autocorrelation function of random walk
1.0

0.5

0.0
ACF

0.5

1.0
0 5 10 15 20 25 30 35

Partial autocorrelation function of random walk

1.0

0.5
PACF

0.0

0.5

1.0
0 5 10 15 20 25 30 35
Lags

Simon Trimborn (Introduction to Time-Series Analysis) Financial Econometrics Executive MSc in International Finance (MIF) 22 / 50
AR(p) Models

The AR(1) model can be extended to the AR(p) model:

yt = µ + ϕ1 yt −1 + . . . + ϕp yt −p + ϵt .

Stationary under complicated conditions on ϕi ; necessary (not sufficient) condition:

ϕ1 + . . . + ϕp < 1.

The acf gradually approaches zero, not necessarily with a clear pattern.
Partial autocorrelation function (pacf): τkk = 0, k > p. Can be used to select p!
All autoregressive models are linear regressions, so can be estimated by OLS.
After estimation, residuals should be tested for residual autocorrelation (Breusch-Godfrey
test, Q-tests). If significant, more lags should be added.

Simon Trimborn (Introduction to Time-Series Analysis) Financial Econometrics Executive MSc in International Finance (MIF) 23 / 50
Where do we stand?

How to model past returns? ✓ - AR model

How to detect order of AR model? ✓ - ACF, PACF
Now:
▶ How to model past errors

Simon Trimborn (Introduction to Time-Series Analysis) Financial Econometrics Executive MSc in International Finance (MIF) 24 / 50
MA and ARMA Models

Alternative time series models are the moving average model of order q, MA(q):

yt = µ + ϵt + θ1 ϵt −1 + . . . + θq ϵt −q ,

and the mixed autoregressive-moving average model, ARMA(p,q):

yt = µ + ϕ1 yt −1 + . . . + ϕq yt −p + ϵt + θ1 ϵt −1 + . . . + θq ϵt −q .

These provide approximations to AR(p) with p large: can be expressed as AR(∞).

▶ To be considered if pacf τkk does not have a clear cut-off point;

▶ MA(q) = ARMA(0,q) models will be useful if acf τk = 0 for k > q.

Estimation is more complicated, to be done via non-linear least-squares.

Simon Trimborn (Introduction to Time-Series Analysis) Financial Econometrics Executive MSc in International Finance (MIF) 25 / 50
ACF / PACF MA(1)
Autocorrelation function of MA(1) process
1.0

0.5

0.0
ACF

0.5

1.0
0 5 10 15 20 25 30 35

Partial autocorrelation function of MA(1) process

1.0

0.5
PACF

0.0

0.5

1.0
0 5 10 15 20 25 30 35
Lags

Simon Trimborn (Introduction to Time-Series Analysis) Financial Econometrics Executive MSc in International Finance (MIF) 26 / 50
ACF / PACF White Noise
Autocorrelation function of white noise
1.0

0.5

0.0
ACF

0.5

1.0
0 5 10 15 20 25 30 35

Partial autocorrelation function of white noise

1.0

0.5
PACF

0.0

0.5

1.0
0 5 10 15 20 25 30 35
Lags

Simon Trimborn (Introduction to Time-Series Analysis) Financial Econometrics Executive MSc in International Finance (MIF) 27 / 50
Integrated Processes

Many financial time series do not seem stationary, displaying a trend in mean, and a vari-
ance increasing with the level. The latter is usually solved by a log-transformation, but then
the series may still be non-stationary.
yt is called integrated of order 1, or I (1), if it is non-stationary, while ∆yt = yt − yt −1 is
stationary. Such processes have acfs close to one, which die out very slowly.
Why integrated?: prior values are ‘integrated’ into other series
Note that if yt = log Pt , with Pt an asset price, then ∆yt is a return. It seems reasonable to
assume stationarity of financial returns and growth rates. This implies that prices are I (1)
and hence nonstationary.
Informal way to check stationarity: inspect the plot and the correlogram of the series. If
the plot displays a tendency to revert to a constant mean, with a more or less constant
variance, and the correlogram converges to zero “exponentially fast”, then stationarity may
be assumed.
A formal test will be introduced later.

Simon Trimborn (Introduction to Time-Series Analysis) Financial Econometrics Executive MSc in International Finance (MIF) 28 / 50
Properties of Integrated Series

No mean-reversion. Some financial theories (purchasing power parity, term structure mod-
els) imply that certain (combinations of) series should revert to a constant mean.
Persistence of shocks. For I (1) processes, the effect of shocks ϵt on yt +s does not die out
as s increases; for stationary series it will decay exponentially.
Increasing forecast intervals. For I (0) time series, 95% forecast interval for yt +s converges
with horizon s; for an I (1) process, forecasts intervals keep increasing.
Spurious regressions. When regressing two integrated time series onto each other, the R 2
and t-statistic may become large even if they are totally independent. This is avoided if we
regress ∆yt on ∆xt .
Asymptotic properties of estimators and tests. In time-series regressions, t- and F -statistics
have asymptotic t- or F -distributions under the null for stationary time series. In regressions
with I (1) series, different distributions may occur.

Simon Trimborn (Introduction to Time-Series Analysis) Financial Econometrics Executive MSc in International Finance (MIF) 29 / 50
ARIMA Models

Following patterns for the acf and pacf of AR, MA and ARMA processes:
▶ AR(p): geometrically decaying acf, pacf is zero after p lags;
▶ MA(q): acf is zero after q lags, geometrically decaying pacf;
▶ ARMA(p, q): geometrically decaying acf and pacf.
If yt ∼ I (1), and ∆yt has an ARMA structure, then yt follows an autoregressive-integrated-
moving average or ARIMA(p, d , q) model with d = 1.
General d would imply repeated differencing ∆d yt = ∆ · · · ∆yt to get stationarity, but we
usually only consider d = 1 or d = 0; ARIMA(p, 0, q) = ARMA(p, q).
It is often hard to infer the orders p and q from the correlogram. Therefore, try out model
with low orders (e.g., ARMA(1,1)), and test whether it needs to be extended.

Simon Trimborn (Introduction to Time-Series Analysis) Financial Econometrics Executive MSc in International Finance (MIF) 30 / 50
Where do we stand?

How to model past errors? ✓ - MA model

How to detect order of MA model? ✓ - ACF, PACF
What if data are integrated? ✓ - Take differences until they are no longer integrated
Now:
▶ How to select appropriate model

Simon Trimborn (Introduction to Time-Series Analysis) Financial Econometrics Executive MSc in International Finance (MIF) 31 / 50
Model Selection Criteria

If more than one model passes diagnostic tests (e.g., both AR(1) and AR(3)), choice is
often based on Akaike information criterion (AIC) or Bayesian information criterion (BIC),
with k the number of parameters:

T
!
1 X 2k
AIC = ln bϵ2t + ,
T T
t =1
T
!
1 X k ln T
BIC = ln bϵ2t + .
T T
t =1

▶ In some packages 1 + ln 2π is added to all criteria; irrelevant for comparison.

Choose model with smallest AIC or BIC (AIC leads to higher orders).
Similar idea as adjusted R 2 : second term is penalty for including too many parameters.
Trade-off between goodness of fit and parsimony.
▶ Smaller models often forecast better — less overfitting.

Simon Trimborn (Introduction to Time-Series Analysis) Financial Econometrics Executive MSc in International Finance (MIF) 32 / 50
AIC & BIC Differentiation I

AIC asymp. optimal

▶ Overfitting
▶ Bias-Variance trade off
▶ Target model depends on sample size
BIC consistent
▶ Underfitting
▶ BIC achieves true model if part of set
▶ True model existed, independent of T
▶ Choice of quasi-true model

Simon Trimborn (Introduction to Time-Series Analysis) Financial Econometrics Executive MSc in International Finance (MIF) 33 / 50
AIC & BIC Differentiation II

Choice depends on philosophy

More complex models decrease ‘Likelihood’ (L) by a lot: AIC
L constant for more complex ones: BIC
L decreases strongly until one model, then slightly: unclear
Tapering effects: AIC

Simon Trimborn (Introduction to Time-Series Analysis) Financial Econometrics Executive MSc in International Finance (MIF) 34 / 50
AIC & BIC Differentiation III

Search for true model: BIC

Search for best model (under squared loss function): AIC
Search for
▶ simple model: BIC
▶ complex model: AIC
Model in set: BIC

Simon Trimborn (Introduction to Time-Series Analysis) Financial Econometrics Executive MSc in International Finance (MIF) 35 / 50
AIC/BIC for ARMA models for RDJI

Simon Trimborn (Introduction to Time-Series Analysis) Financial Econometrics Executive MSc in International Finance (MIF) 36 / 50
Forecasting Terminology

The main purpose of AR / MA / ARMA models is forecasting.

Denote information available at time t as Ωt . The s-period-ahead forecast of yt +s based on
Ωt is denoted b
yt ,s
Distinguish in-sample and out-of-sample forecasts:
▶ In-sample means forecasting observations already used for estimation.
▶ Out-of sample means that the data is divided into two groups, say t = 1, . . . , T1 (estimation
sample) and t = T1 + 1, . . . , T (holdout sample). The former is used for model selection and
estimation. Based on this, we forecast yT1 +1 , . . . , yT .
Alternatively, one may roll the estimation window forward: always use most recent informa-
tion Ωt for forecasting yt +1 . Relevant for taking model for a test drive.

Simon Trimborn (Introduction to Time-Series Analysis) Financial Econometrics Executive MSc in International Finance (MIF) 37 / 50
One-Step-Ahead and Multi-Step Forecasts

Consider stationary AR(p) model,

yt +1 = µ + ϕ1 yt + . . . + ϕp yt −p+1 + ϵt +1 ,

then the one-step-ahead forecast is

yt ,1 = µ + ϕ1 yt + . . . + ϕp yt −p+1 .
b

When forecasting yt +s with s > 1 based on Ωt , future yt +1 , . . . , yt +s−1 are unknown.

Solution: forecast yt +1 first, then use this to forecast yt +2 , etc. All future errors are replaced
by their conditional expectation zero.
Procedure can be generalized to ARMA(p,q).
In practice, parameters must be replaced by their estimates.

Simon Trimborn (Introduction to Time-Series Analysis) Financial Econometrics Executive MSc in International Finance (MIF) 38 / 50
Remarks

p
Possible to construct forecast standard errors SEt ,s = var(yt +s |Ωt ), hence 95% forecast
intervals
yt ,s ± 1.96 × SEt ,s .
b
▶ For stationary processes, forecast interval converges to µ ± 1.96 × σ as s → ∞.
For integrated processes, first forecast stationary ∆yt +s , and then use

E (yt +s |Ωt ) = yt + E (∆yt +1 |Ωt ) + . . . + E (∆yt +s |Ωt ).

Thus forecasts of ∆yt +i are accumulated;

▶ as are the errors: SEt ,s keeps on increasing with s.

Simon Trimborn (Introduction to Time-Series Analysis) Financial Econometrics Executive MSc in International Finance (MIF) 39 / 50
Forecast log returns Dow Jones Index

Dynamic forecasts of RDJI

0.10 RDJI
forecast
95% confidence interval

0.05

0.00

0.05

0.10
0 4 8 2 6 0
200 200 200 201 201 202
Date

Simon Trimborn (Introduction to Time-Series Analysis) Financial Econometrics Executive MSc in International Finance (MIF) 40 / 50
Forecast Dow Jones Index

Dynamic forecasts of DJI

DJI
35000 forecast
95% confidence interval

30000

25000

20000

15000

10000

0 4 8 2 6 0
200 200 200 201 201 202
Date

Simon Trimborn (Introduction to Time-Series Analysis) Financial Econometrics Executive MSc in International Finance (MIF) 41 / 50
Forecast Evaluation

Forecast accuracy measured by mean squared error (MSE) or mean absolute error (MAE):

T T
1 X 1 X
MSE = bϵ2t , MAE = |bϵt | ,
T − T1 T − T1
t =T1 +1 t =T1 +1

ϵt = yt − b
with b yt .
2 2
Forecasts from two competing models can be compared by regressing b ϵ1t − bϵ2t on a con-
stant. Under H0 of equal predictive ability, its t statistic (using HAC standard errors) is
asymptotically N (0, 1) (Diebold-Mariano test).

Simon Trimborn (Introduction to Time-Series Analysis) Financial Econometrics Executive MSc in International Finance (MIF) 42 / 50
Where do we stand?

Models ✓
Choice of models ✓
Forecast ✓
Now:
▶ Unit root (integrated time series) and how to test for it

Simon Trimborn (Introduction to Time-Series Analysis) Financial Econometrics Executive MSc in International Finance (MIF) 43 / 50
Unit Root Tests

Recall that if yt ∼ I (1), it must be must be differenced to become stationary. Such a

process is said to have a unit root.
So far, decision to take differences was based on correlogram: if autocorrelations decay
slowly, approximately linearly, then the series may be I (1).
This procedure is subjective and unreliable: a trend-stationary series, such as yt = α +
β t + ϵt , will also have large and slowly decaying autocorrelations.
Hence the need for formal tests to distinguish integrated from (trend-) stationary time series:
unit root tests. Best-known is the (Augmented) Dickey-Fuller test.

Simon Trimborn (Introduction to Time-Series Analysis) Financial Econometrics Executive MSc in International Finance (MIF) 44 / 50
The Dickey-Fuller Test

In AR(1) model
yt = ϕyt −1 + ϵt ,

we wish to test the unit root null hypothesis against a one-sided alternative:

H0 : ϕ = 1, H1 : ϕ < 1 .

Under H0 , yt is a random walk, hence I (1); under the alternative, yt is stationary (provided
ϕ > −1).
Hence, we test for H0 : yt ∼ I (1) against H1 : yt ∼ I (0).
Dickey-Fuller test is based on t-statistic for ψ = ϕ − 1 = 0 in (rewritten) AR(1):

∆yt = ψ yt −1 + ϵt .

Note that ψ < 0 under alternative hypothesis.

Simon Trimborn (Introduction to Time-Series Analysis) Financial Econometrics Executive MSc in International Finance (MIF) 45 / 50
The Dickey-Fuller Distribution

We reject H0 if τ = tψ is less than the (negative) critical value.

In stationary regression, 5% one-sided critical value is −1.645. Because regressor yt −1 is
I (1) under H0 , distribution changes, with 5% critical value τ0.05 = −1.95.
This critical value changes to τµ,0.05 = −2.86 if we include an intercept:

∆yt = ψ yt −1 + µ + ϵt ,

in which case we call the t-statistic τµ . Relevant if we want to allow for a constant non-zero
mean E (yt ) = −µ/ψ under H1 .
If we want to test a random walk with drift against a trend-stationary alternative, the relevant
regression is
∆yt = ψ yt −1 + µ + λt + ϵt ,
and the t-statistic is called ττ , with 5% critical value ττ,0.05 = −3.41.

Simon Trimborn (Introduction to Time-Series Analysis) Financial Econometrics Executive MSc in International Finance (MIF) 46 / 50
Choice of Model

Three Dickey-Fuller tests, based on τ , τµ and ττ ; which one to use in practice?

▶ The τ test without both an intercept or trend is almost never applicable.

▶ The τ
µ test is relevant for series such as interest rates and real exchange rates, where we do
not expect a linear trend (under H0 or H1 ).
▶ Many other financial time series, including (log-) asset prices, display a trend ⇒ ττ .

Simon Trimborn (Introduction to Time-Series Analysis) Financial Econometrics Executive MSc in International Finance (MIF) 47 / 50
The Augmented Dickey-Fuller Test

The augmented Dickey-Fuller (ADF) test is an extension to the AR(p) model:

yt = ϕ1 yt −1 + . . . + ϕp yt −p + ϵt .

yt is I (1) under H0 : ϕ1 + . . . + ϕp = 1, and stationary under H1 : ϕ1 + . . . + ϕp < 1.

Hence test H0 : ψ = 0 against H1 : ψ < 0 in the regression

∆yt = ψ yt −1 + α1 ∆yt −1 + . . . + αp∗ ∆yt −p∗ + ϵt ,

where ψ = ϕ1 + . . . + ϕp − 1, and p∗ = p − 1.
Interpretation of H0 and H1 , role of the constant and trend, and critical values are the same
as AR(1) model.
p is unknown:
▶ must be large enough to avoid autocorrelation, but not too large (test power);
▶ chosen based on AIC, BIC; usually packages have built-in options (also p-values).

Simon Trimborn (Introduction to Time-Series Analysis) Financial Econometrics Executive MSc in International Finance (MIF) 48 / 50
ADF unit root tests on RDJI

Simon Trimborn (Introduction to Time-Series Analysis) Financial Econometrics Executive MSc in International Finance (MIF) 49 / 50
Exercises

Questions 6.3, 6.8, 6.9, 8.2, 8.3 of Brooks (2019).

Simon Trimborn (Introduction to Time-Series Analysis) Financial Econometrics Executive MSc in International Finance (MIF) 50 / 50