Overview of Time Series Models and Tests

Dr Merwan Roudane
July 12, 2024

Unit Root Tests KSS Test (Nonlinear) Kapetanios et al. (2005) Test

Assumptions: Assumptions:
ADF Test • ESTAR nonlinearity • Nonlinear alternative hypothesis
• Globally stationary process • Globally stationary process
Assumptions:
Use: When suspecting smooth Use: When suspecting smooth
• Series follows AR(p) process
transition nonlinearity transition nonlinearity
• Errors are i.i.d.
Don’t use: With other forms of Don’t use: With other forms of
Use: Testing for unit root in time
nonlinearity nonlinearity
series
Software: R (nlar package), Software: Gretl
Don’t use: With structural
GAUSS
breaks or nonlinearity
Software: EViews, Stata, R,
MATLAB, Gretl Quantile Unit Root Test

Assumptions:
Sollis Test (Nonlinear) • Unit root properties may differ
Phillips-Perron (PP) Test across quantiles
Assumptions: • Robust to heteroskedasticity
Assumptions: • Asymmetric STAR-type nonlin- Use: When suspecting different
• Non-parametric correction for earity unit root properties at different
serial correlation • Different speeds of adjustment quantiles
• Robust to heteroskedasticity Use: When asymmetric adjust- Don’t use: With homogeneous
Use: When concerned about serial ment suspected unit root properties
correlation or heteroskedasticity Don’t use: When symmetry is Software: GAUSS, Python
Don’t use: With small samples reasonable
(size distortions) Software: R, GAUSS
Software: EViews, Stata, R,
MATLAB, Gretl

Cointegration Tests
KPSS Test Zivot-Andrews Test
Engle-Granger Test
Assumptions: Assumptions:
• Null hypothesis is stationarity • Single endogenous break Assumptions:
• Series is I(0) or I(1) • Break in intercept, trend, or • Variables are I(1)
Use: Confirmatory analysis with both • Single cointegrating relationship
ADF or PP Use: When suspecting a single Use: Testing for bivariate cointe-
Don’t use: When series poten- structural break gration
tially I(2) or higher Don’t use: When multiple breaks Don’t use: With multiple cointe-
Software: EViews, Stata, R, are likely grating relationships
MATLAB, Gretl Software: EViews, Stata, R Software: EViews, Stata, R,
MATLAB, Gretl

DF-GLS Test
Johansen Test
Assumptions: Lee-Strazicich Test
• Detrended series follows AR(p) Assumptions:
process Assumptions: • Variables are I(1)
• More powerful than standard • Two endogenous breaks • VAR framework
ADF • Breaks in level and/or trend Use: Testing for multiple cointe-
Use: When more power needed, Use: When suspecting two struc- grating relationships
especially with trend tural breaks Don’t use: With small samples or
Don’t use: With structural Don’t use: When series too short near-unit roots
breaks for multiple breaks Software: EViews, Stata, R,
Software: EViews, Stata, R Software: GAUSS, R, MATLAB MATLAB, Gretl
Kapetanios et al. Test (Non- Quantile Cointegration Test ARDL Models and Variants
linear)
Assumptions: Standard ARDL
Assumptions: • Cointegration properties may Assumptions:
• ESTAR-type nonlinearity in differ across quantiles • Variables are I(0) or I(1)
residuals • Robust to heteroskedasticity • No I(2) variables
• Globally stationary residuals Use: When suspecting different • Single long-run relationship
Use: When suspecting smooth cointegration properties at differ- Use: Testing cointegration with
transition in long-run relationship ent quantiles mixed order of integration
Don’t use: With other forms of Don’t use: With homogeneous Don’t use: With I(2) variables or
nonlinearity cointegration properties multiple long-run relationships
Software: GAUSS, R Software: Matlab Software: EViews, Stata, R,
MATLAB

Error Correction Models NARDL (Nonlinear ARDL)

Arai-Kurozumi Test
Linear ECM
Assumptions:
Assumptions: • Asymmetric effects in short
Assumptions:
• Null of cointegration
• Variables are cointegrated and/or long run
• Allows for structural breaks • Variables decomposed into posi-
• Linear adjustment to equilib-
Use: Testing stability of cointe- tive and negative changes
rium
grating relationship Use: When suspecting asymmet-
Use: Modeling short-run dynam-
Don’t use: Without prior belief in ric relationships
ics with long-run relationship
cointegration Don’t use: When asymmetry is
Don’t use: When nonlinear ad-
Software: R not theoretically justified
justment suspected
Software: EViews, Stata, R, Software: EViews, GAUSS, R
MATLAB, Gretl

Gregory-Hansen Test Quantile ARDL

Nonlinear ECM
Assumptions: Assumptions:
• Single unknown break in cointe- Assumptions: • Relationships vary across quan-
grating relationship • Nonlinear adjustment to equilib- tiles of dependent variable
• Break in intercept, trend, or rium • Allows for heterogeneous cointe-
regime shift • Often uses smooth transition or gration
Use: When suspecting a break in threshold models Use: To examine effects across dif-
long-run relationship Use: With asymmetric or state- ferent quantiles
Don’t use: When multiple breaks dependent adjustment Don’t use: When relationships
are likely Don’t use: When linear adjust- are homogeneous across distribu-
Software: Stata, WinRATS, ment is sufficient tion
GAUSS Software: EViews, R, MATLAB Software: GAUSS

Standard VECM Bootstrapping ARDL

Assumptions: Assumptions:
Hatemi-J Test
• Multiple variables are cointe- • Resampling preserves data struc-
Assumptions: grated ture
• Two unknown breaks in cointe- • Linear adjustment to equilib- • Improves small sample proper-
grating relationship rium ties
• Breaks in level and/or slope Use: Modeling multivariate sys- Use: To obtain robust inference in
Use: When suspecting two breaks tems with cointegration small samples
in long-run relationship Don’t use: Without cointegration Don’t use: With very large sam-
Don’t use: When series too short or with nonlinear adjustment ples where asymptotics work well
for multiple breaks Software: EViews, Stata, R, Software: EViews, GAUSS, R,
Software: GAUSS MATLAB, Gretl Gretl

Markov-Switching VECM Dynamic Simulation ARDL

Maki Cointegration Test Assumptions: Assumptions:

• Regime-switching behavior in • Model is correctly specified
Assumptions: VECM • Simulations capture uncertainty
• Multiple endogenous breaks • Discrete state changes in estimates
• Breaks in intercept and/or trend Use: When suspecting regime Use: For forecasting and policy
Use: When suspecting multiple changes in adjustment process analysis
structural breaks Don’t use: Without clear regime Don’t use: When model misspec-
Don’t use: With small samples shifts or with small sample ification is suspected
Software: GAUSS Software: EViews, R, MATLAB Software: Stata, R
 Fourier ARDL • Interpretation of results can be complex,
especially for nonlinear models
Assumptions: • Bootstrap methods may be necessary
• Smooth structural changes for reliable inference in small samples
• Fourier approximation captures
unknown form of breaks
Use: When suspecting gradual Additional Considerations
structural changes • Check for appropriate lag length selec-
Don’t use: With abrupt struc- tion
tural breaks • Consider panel versions for panel data
Software: EViews, WinRATS, R, • Be aware of size distortions and power
GAUSS issues in small samples
• Seasonal unit roots may require special-
ized tests (e.g., HEGY)
ARDL Model Selection and Di- • For high-frequency data, consider frac-
agnostics tional integration and cointegration
• Bootstrapping can improve inference,
• Use information criteria (AIC, SIC, HQ)
especially for nonlinear models and for
for lag selection
small sample size
• Check for serial correlation in residuals
• Always validate model assumptions and
(e.g., LM test)
conduct robustness checks
• Test for heteroskedasticity (e.g., ARCH
• Consider economic theory when specify-
test)
ing models and interpreting results
• Examine stability of parameters (e.g.,
CUSUM, CUSUMSQ tests)
• Ensure no I(2) variables using unit root
tests
• Validate exogeneity assumptions when
necessary
• Consider model averaging for robust in-
ference
• Use bounds test for cointegration in
standard ARDL
• For NARDL, test for asymmetry using
Wald tests
• In Quantile ARDL, examine coefficient
stability across quantiles
• For Fourier ARDL, select appropriate
number of Fourier terms
• Always conduct robustness checks with
alternative specifications

Advantages of ARDL Ap-

proach
• Allows for mixed order of integration
(I(0) and I(1))
• Provides both short-run and long-run
estimates
• Can be used with relatively small sam-
ple sizes
• Allows for different optimal lags for each
variable
• Easily extended to account for struc-
tural breaks
• Can incorporate nonlinear and asym-
metric relationships
• Useful for both time series and panel
data analysis

Limitations and Cautions

• Not suitable for I(2) variables
• Assumes a single long-run relationship
• May suffer from endogeneity issues in
some cases
• Sensitive to lag length selection
• Can be computationally intensive, espe-
cially advanced variants
Compiled by Dr Merwan Roudane
Comparison of ARDL Family Models

Model Assumptions Software

Standard ARDL Variables are I(0) or I(1); EViews, Stata, R, MATLAB
No I(2) variables; Single
long-run relationship
NARDL (Nonlinear Asymmetric effects in EViews, GAUSS, R
ARDL) short and/or long run;
Variables decomposed
into positive and negative
changes
Quantile ARDL Relationships vary across GAUSS
quantiles of dependent
variable; Allows for het-
erogeneous cointegration
Bootstrapping Resampling preserves EViews, GAUSS, R, Gretl
ARDL data structure; Improves
small sample properties
Dynamic Simula- Model is correctly speci- Stata, R
tion ARDL fied; Simulations capture
uncertainty in estimates
Fourier ARDL Smooth structural EViews, WinRATS, R, GAUSS
changes; Fourier ap-
proximation captures
unknown form of breaks