Non-Stationary Models

Time Series

Ioannis Vrontos
Athens University of Economics and Business

Department of Statistics

Ioannis Vrontos, Athens University of Economics and Business Time Series

 Non-Stationary Models

Non-Stationary Models

I Introduction: Non-Stationary Time series

I examples
I the problem
I Co-integration
I Definition
I Co-integration tests
I Equilibrium
I Error Correction Models [ECM]
I error correction mechanism - term
I error correction models
I Methodology for practitioners

Ioannis Vrontos, Athens University of Economics and Business Time Series

 Non-Stationary Models

Introduction

I Consider yt and xt two stochastic processes [e.g. time series]

I Examples:
I Income and Expenses
I Stock price and Dividends
I 3-month and 6-month treasury bill rates
I Government Expenditures and taxes
I Money theory: money supply and prices
I Is there long term relationship between these variables?
I Problem: To examine if there exist real relation or not -
Spurious regression or not

Ioannis Vrontos, Athens University of Economics and Business Time Series

 Non-Stationary Models

Introduction

Consider yt and xt two non-stationary processes which are I(1)

i.e. become stationary after taking differences
Assume that yt and xt are Random walk processes with drift

xt = µ + xt−1 + εt

yt = λ + yt−1 + ut

which can be written as follows:

Pt
xt = x0 + tµ + s=1 εs
Pt
yt = y0 + tλ + s=1 us

Ioannis Vrontos, Athens University of Economics and Business Time Series

 Non-Stationary Models

Introduction

I Variables yt and xt depend on time t, they have trend

I If we estimate the regression model
yt = α + βxt + εt
coefficient β will be statistical significant, due to trend
I Use the residuals ε̂t
ε̂t = yt − α̂ − β̂xt
I If the residuals ε̂t are not stationary, then the regression is
Spurious

Ioannis Vrontos, Athens University of Economics and Business Time Series

 Non-Stationary Models

Introduction

I If the residuals ε̂t are not stationary

I their variance depends on time
I t-test and F-test are not valid
I statistical inference for the regression parameters is not valid
I yt and xt are non-stationary processes, but ∆yt and ∆xt are
stationary
I Why not regress ∆yt on ∆xt ?
I Because we lose information for the long term relationship
between yt and xt

Ioannis Vrontos, Athens University of Economics and Business Time Series

 Non-Stationary Models

Co-integration

I Aim: Find methods to investigate if the relation between the

random variables is real or not, is spurious or not
I Cointegration: if yt and xt are non-stationary processes, but
there exist a linear combination of them which is stationary,
then the random variables are co-integrated
I yt is not stationary
I xt is not stationary
I regression model and errors:
yt = α + βxt + εt ⇔ εt = yt − α − βxt
I If the errors εt are stationary, the linear combination is
stationary, and the random variables are co-integrated

Ioannis Vrontos, Athens University of Economics and Business Time Series

 Non-Stationary Models

Co-integration

I If the errors εt are stationary, the random variables yt and xt

are co-integrated. Then:
I The regression results are valid
I The regression is meaningful
I We do not lose any valuable information as if we used the
differences ∆yt and ∆xt

Ioannis Vrontos, Athens University of Economics and Business Time Series

 Non-Stationary Models

Co-integration tests: Different methods

Different co-integration testing approaches have been proposed in

the literature:
I Engle and Granger co-integration test
I Co-integrating regression Durbin-Watson test
I Johansen Co-integration test

Ioannis Vrontos, Athens University of Economics and Business Time Series

 Non-Stationary Models

Engle and Granger Co-integration test

Consider the regression model: yt = α + βxt + εt

I Estimate model parameters α̂ and β̂ using OLS
I Obtain the residuals ε̂t = yt − α̂ − β̂xt
I Perform stationarity test for the estimated residuals ε̂t
I Test the null hypothesis H0 : ε̂t are not stationary vs the
alternative hypothesis H1 : ε̂t are stationary
I Use a model without constant and trend, i.e. ε̂t = ρε̂t−1 + υt
or ∆ε̂t = δ ε̂t−1 + υt
I with appropriate critical values, since we estimate α and β

Ioannis Vrontos, Athens University of Economics and Business Time Series

 Non-Stationary Models

Co-integrating regression Durbin-Watson test

Consider the regression model: yt = α + βxt + εt

I Estimate model parameters α̂ and β̂ using OLS
I Obtain the residuals ε̂t = yt − α̂ − β̂xt
I Perform Durbin-Watson stationarity test for the estimated
residuals ε̂t
I Test the null hypothesis H0 : d = 0 (random walk, non
stationary process vs the alternative hypothesis H1 : d > 0,
stationary process P 2
t −ε̂t−1 )
I Use the test statistic DW = (ε̂P ε̂2t
I Critical values: α = 0.01, cv = 0.511, α = 0.05, cv = 0.386,
α = 0.1, cv = 0.322
I if DW > cv then reject H0 , stationary residuals, i.e. there is
co-integration relation
I if DW < cv then do not reject H0 , non stationary residuals,
i.e. there is not co-integration relation

Ioannis Vrontos, Athens University of Economics and Business Time Series

 Non-Stationary Models

Co-integrating regression Durbin-Watson test

Test the null hypothesis H0 : d = 0 , rather than d = 2, because

d = 2(1 − ρ), so if there is a unit root ρ will be about one, which
implies that d = 0
ε̂2t + ε̂2t−1 −2
P P P
(ε̂t −ε̂t−1 )2 ε̂t ε̂t−1
P
dˆ = ε̂2t ≈ ε̂2t−1 ]
P P
P 2
ε̂t
= P 2
ε̂t
,[

ε̂2t −2
P P P P
2 ε̂t ε̂t−1 ε̂ ε̂ ε̂ ε̂
dˆ ≈ P
ε̂2t
= 2(1 − Pt t−1
ε̂2t
) , [ρ̂ = Pt t−1
ε̂2t
]

dˆ = 2(1 − ρ̂)

Ioannis Vrontos, Athens University of Economics and Business Time Series

 Non-Stationary Models

Co-integration and Error Correction Models

Equilibrium
I If there is a stable equilibrium yt = α + βxt , the discrepancy
yt − α − βxt contains information since on average the system
will move towards equilibrium, if it is not there
I The term yt−1 − α − βxt−1 represents the deviation from
equilibrium at previous time period t − 1
I The discrepancy or error yt−1 − α − βxt−1 should be used as
explanatory variable for the next direction of movements of yt .
If yt−1 − α − βxt−1 is positive, yt−1 is too high relative to
xt−1 and on average we might accept a fall in y in future
periods relative to its trend growth
I The term yt−1 − α − βxt−1 is called error correction
mechanism

Ioannis Vrontos, Athens University of Economics and Business Time Series

 Non-Stationary Models

Co-integration and Error Correction Models

Correction Mechanism
I If yt and xt are co-integrated there is a long term equilibrium
relationship between them
I In the short term however, there may be disequilibrium
I Therefore, the error term εt of the regression
yt = α + βxt + εt or εt = yt − α − βxt can be seen as
equilibrium error
I We can use this equilibrium error to find the short-run
behaviour of y . This is done by the error correction models

Ioannis Vrontos, Athens University of Economics and Business Time Series

 Non-Stationary Models

Error Correction Models

Simple Error Correction Model:

∆yt = α0 + α1 ∆xt + α2 ε̂t−1 + υt

I ε̂t−1 = yt−1 − α̂ − β̂xt−1 is the empirical estimate of the

equilibrium error term, i.e. are the one period lagged values of
the residuals of the co-integrated equation
I The Error Correction model relates the change in y to the
change in x and in the equilibrium error in the previous
period. The term ∆xt captures the short-run disturbances in
x, while the error correction term ε̂t−1 captures the correction
to the long run equilibrium. If the parameter α2 is statistical
significant, shows what amount from disequilibrium in y at
one period is corrected for the next period

Ioannis Vrontos, Athens University of Economics and Business Time Series

 Non-Stationary Models

Methodology for practinioners

I Use unit root tests to determine the order of integration of

the initial (row) data series
I Run the co-integration regression suggested by the economic -
financial theory
I Apply appropriate unit root tests to the residuals from this
regression to test for co-integration. If the residuals are
stationary we have co-integration
I If co-integration is accepted, use lagged residuals from the
co-integrating regression as an error correction term in the
Error Correction Model
I if more than two explanatory variables are involved, there could
be more than one co-integration relationship
I need for simultaneous equation formulation
I the case of more co-integration relationships can be addressed
by Johansen procedure
Ioannis Vrontos, Athens University of Economics and Business Time Series
 Non-Stationary Models

Application to financial and economic series

I Example: Co-integration test and Error Correction models

Ioannis Vrontos, Athens University of Economics and Business Time Series

 Non-Stationary Models

References

Hamilton, James D. Time Series Analysis. Princeton, New Jersey: Princeton University Press, 1994.
Enders, Walter. Applied Econometric Time Series. New York: Wiley, 2010.
Cowpertwait, Paul S.P., and Metcalfe V. Andrew. Introductory Time Series with R. New York: Springer Texts in
Statistics, 2009.
Cryer, Jonathan D., and Chan Kung-Sik. Time Series Analysis with Applications in R. Springer Texts in Statistics,
2010.
Gujarati, Damodar N. Basic Econometrics. New York: McGraw-Hill, 2008.
Harvey, Andrew. Time Series Models. Cambridge: MIT Press, 1993.
Hendry, David F. Dynamic Econometrics. Oxford: Oxford University Press, 1995.
Shumway, Robert H. and David S. Stoffer. Time Series Analysis and Its Applications with R Examples. New York:
Springer Texts in Statistics, 2011.

Ioannis Vrontos, Athens University of Economics and Business Time Series