“Testing the long-run implications of the expectation hypothesis using cointegration

techniques with structural change”.

Emerson Fernandes Marçal
Universidade Presbiteriana Mackenzie e IE-UNICAMP
R. Conselheiro Brotero 1030, apt 33
01232-010, São Paulo, S.P.
Tel: (011) 3824-9205
Fax: (011) 3663-4345
[email protected]

Pedro Luiz Valls Pereira

EESP-FGV
Rua Itapeva 474 room 1202
01332-000, São Paulo. S.P.
Tel: (011) 3281-3726
Fax: (011) 3281-3357
[email protected]

Omar Abbara
IMECC-UNICAMP
Caixa Postal 6065
13083-859, Campinas, S.P.
[email protected]

Abstract:
This paper investigates whether or not multivariate cointegrated process with
structural change can describe the Brazilian term structure of interest rate data from 1995 to
2006. In this work the break point and the number of cointegrated vector are assumed to be
known. The estimated model has four regimes. Only three of them are statistically different.
The first starts at the beginning of the sample and goes until September of 1997. The
second starts at October of 1997 until December of 1998. The third starts at January of
1999 and goes until the end of the sample. It is used monthly data. Models that allows for
some similarities across the regimes are also estimated and tested. The models are
estimated using the Generalized Reduced-Rank Regressions developed by Hansen (2003).
All imposed restrictions can be tested using likelihood ratio test with standard asymptotic
qui-squared distribution. The results of the paper show evidence in favor of the long run
implications of the expectation hypothesis for Brazil.

Key-words: Term structure, cointegration, structural change

JEL codes: C32; C52; G14

1 Introduction:
A recent effort has been done to test the expectation hypothesis for Brazilian data.
Although this research is in its early stages for Brazil, there is a vast literature about this
theme for other countries. (Cuthbertson and Nitzsche (2005))
This paper aims to investigate whether or not multivariate cointegration models with
structural change can better describe the term structure of interest rate data for the period
from 1995 to 2006. The work uses the Generalized Reduced Rank Technique recently
developed by Hansen (2003) to estimate cointegrated process with structural change.
Under this framework the number of cointegration relations and the number and the
moments of the structural change are assumed to be known. This allows testing hypothesis
about the parameters in all regimes and evaluating whether or not there is structural change.
It’s also done an effort to control the possible non-stationarity of volatility by using the
results of Boswijk and Zu (2005). As far as the authors know this has not been done in the
literature for Brazil.
The paper is organized in the following questions. The first section discusses the
expectation hypothesis. The second section presents the Generalized Reduced Rank
technique and how to correct for heterocedasticity under this framework. The third section
the results of the estimated models are shown and discussed. The next section a comparison
with the literature is done. Finally the main conclusions are stated.

2 Expectation Hypothesis: What is?

Define Rtm as the logarithmic of annualized return paid by a m period long run bond

and by Rtn , the logarithmic of the annualized return paid by a n period short run bond with

m < n then the spread (m,n) Sm,n can be defined as Rtm − Rtn .
The basic equation to model the term structure is given by:

1
eq. 1:
Rtm = (1 k )E t [ Rtn + Rtn+in + Rtn+2 n + ... + Rtn+in ] + Tt m , n

where Et denotes the expectation formed in instant t, k = (m/n), i = 1,2,..,k and

Tt m , n the premium for the agents who decides for a long term strategy.
The investor can decide between two strategies. In the first he holds a bond of
maturity m and obtains an annualized return of Rtm . The other strategy consists in buying
bonds of maturity m during (m/n) successive periods. In equilibrium the equation (eq. 1)
must hold if the agents arbitrate the difference in two relations corrected for the risk
premium and if the expectations are rational. This is called the expectation hypothesis.
The discussion about the determinants of term premium started long ago. Authors
like Keynes (1930) and Hicks (1946) have discussed its determinants. A recent survey was
done by Shiller (1990).
Cuthbertson and Nitzsche (2005) pages 494 to 498 show the following typology to
model the term risk premium:
I) Pure Expectation Hypothesis: the term premium is assumed to be constant and
equals zeros for all maturities.
II) Constant Premium Expectation Hypothesis: The term premium differs from zero
and it is equal for all maturities;
III) Liquidity Preference or Growing Term premium: The term premium is constant
throughout the time and it’s bigger for longer spreads;
IV) Time Varying Risk Premium: the term premium varies throughout the time;
V) Market Segmentation: The value of the asset depends in some way of its stock level
and this influences the spreads;
VI) Preferred Habitat Theory: Bond that matures at the same date should be reasonably
close substitutes and, hence, have similar term premium.
The models I to III are very restrictive in terms of generality but are easily
implemented and tested. The model IV is more general and hard to test in this generic
formulation but it seems quite intuitive in the sense that it is possible to distinguish between
long-run premium and short-run premium. An interest application of this idea is done by
McCallum (1994). As a consequence of his model the spread can contain memory due to a
time varying risk premium and can help to predict the variations of the short run equations
2
under certain circumstances. The model V contains possible explanation for the failure of
the expectation hypothesis in its stronger versions. The model VI is a very skeptical
approach for the expectation hypothesis.
The expectation hypothesis can also be stated as the following form.
k −1
i
eq. 2: Rtm − Rtn = ∑ (1 − ) Et ( ∆n Rtn+in ) + Tt m ,n
i =1 k
The eq. 2 is the starting point for the most of the test of the so-called expectation
hypothesis. One can note that if the short interest rate following an integrated process of
order one and the term premium is time varying but stationary, then the spreads must
stationary.
3 Econometric Methodology:
In this section it will be discussed the econometric techniques used in the paper. The
generalized reduced rank regression is briefly discussed and it will be shown how it can be
used to estimate cointegrated models with structural change. This part of the paper is based
mainly in Hansen (2003).

3.1 Estimating a VAR with structural change:

Hansen (2003) generalizes the model proposed by Johansen (1988):
∆X t = Γ1 ( j ) ∆X t + ... + Γk ( j ) ∆X t − k + Φ ( j ) D t + α ( j ) β ' ( j ) X t −1 + ε t
t = 1,...,T
eq. 3
j = 1,...,m
m <T
where ε t are random errors with Ω( j ) as the covariance matrix and j denotes the regime.
The first regime starts at t=0 and ends at t=T1-1. The second regime starts at t=T1 and ends
at t=T2-1 and so on. It’s assumed that there are m different regimes.
The parameters of the model can be defined by the following equations:

3
 Γi ( j ) = Γ1,i 11,t + ... + Γm ,i 1m ,t
Φ ( j ) = Φ 1,i 11,t + ... + Φ m ,i 1m ,t
Ω( j ) = Ω1,i 11,t + ... + Ω m ,i 1m ,t
eq. 4 i = 1,..., k − 1
1 j ,t ≡ 1(T j −1 ≤ t ≤ T j − 1)
T0 ≡ 0
Tm ≡ T + 1

By using the following notation: Z ot = ∆X t , Z 1t = (11,t X t'−1 ,...,1m ,t X t' −1 )' ,

 β1 0 ... 0 0 
0 β2 ... ... 0 
~ 
Z 2t = ( ∆X t' −1 ,..., ∆X t'− k , Dt' )' , B= ... ...  A = (α 1 ,..., α m ) ,
 
0 0 ... β m−1 0 
 0 0 ... 0 β m 

C = (ψ 1 ,...,ψ 1 ) and ψ j = (Γ j ,1 ,..., Γ j , k −1 , Φ j ) .

The eq. 1 can be rewritten as

eq. 5 Z 0 t = AB' Z1t + CZ 2 t + ε t
The problem consists in finding an estimator for A, B and C. This can be done by
using the GRRR.

3.1.1 Generalized Reduced Rank Regression (GRRR):

Hansen (2003) show how to estimate the process described in eq. 5. In order to do
this he uses the vec operator and defines two matrices H and G::
eq. 6 vec( B) = Hϕ + h
where H is known, φ, contains the free parameters and h is a tool to normalize or
identifies the parameters A and B.
eq. 7 vec( A, C ) = Gψ
where G is known and, ψ, contains the free parameters.
By using the eq. 6 and eq. 7, it’s possible to estimate the parameters of the model by
using the following equations:

4
 −1
 T  Bˆ ' Z Z ' Bˆ Bˆ ' Z Z '   
vec( A, C ) = G G ' ∑ 
ˆ ˆ 1t 1t 1t 2 t
 ˆ (t ) −1  G  ×
⊗Ω
 t =1  Z 2 t Z1't Bˆ Z 2 t Z 2' t   
eq. 8 
T
G ' ∑ vec(Ω
ˆ (t ) −1 Z ( Z ' Bˆ , Z ' ))
ot 1t 2t
t =1

−1
 T
ˆ (t )−1 Aˆ ⊗ Z Z ' ) H  ×
vec( Bˆ ) = H  H ' ∑ ( Aˆ ' Ω 1t 1t 
eq. 9  t =1 
 T T
ˆ (t ) −1 Aˆ ⊗ Z Z ' )h  + h
H '  H ' ∑ vec( Z1t ( Z 0t − CZ
ˆ ) 'Ω
2t
ˆ (t ) −1 Aˆ ) −∑ ( Aˆ ' Ω
1t 1t 
 t =1 t =1 
Tj
ˆ ( j ) = (T − T ) −1
eq. 10 Ω j j −1 ∑ εˆ εˆ'
t =T j −1 +1
t t

eq. 11 εˆt = Z 0 t − Aˆ Bˆ ' Z1t + Cˆ Z 2 t

This equation must be used iteratively. The analysts starts the routine with a good
guess for A, C and Ω(t ) and then obtains an estimator for B. Then another estimator for B
and Ω(t ) can be obtained. The process must continue until convergence and the maximum
likelihood estimator are obtained (see Hansen (2003) for details).
The algorithm can also be used to estimate the model if one assumes that variance
and covariance of the errors are known (or estimated previously). The algorithm uses only
equations 8 and 9. In this case Ω(t ) is estimated using the Dynamic Conditional
Correlation Model of Tse & Tsui (2002) and it is this approach that is used in order to
control the heteroscedasticity in the model.

3.2 Description of the estimate models:

In this section it is discussed the models that are going to be used in the empirical

5
part of this paper. It is used models with 4 regimes. With the exception of the variance and
covariance matrix of the errors, all other parameters can be made different within the
regimes.1
The starting point of the analysis is the following VECM:
eq. 12 ∆X t = Γ1 ∆X t −1 + Γ2 ∆X t −2 + α (t )( β ' (t ) X t −1 + ρ (t )) + ε t 2

ε t ~ N (0, Ω(t )) X t = [i 30 i 60 i 90 i180 i 360 ]'

3.3 Data base description:

The data were collected from Risktech web page (www.risktech.com.br) for
different vertices. The frequency of the data is monthly and the data corresponds to last
working day of the month. The vertices were chosen due to liquidity constrains for 30, 60,
90, 180 e 360 days.
Figure 1 shows the evolution of all interest rate for the whole sample. Some periods,
particular in the beginning of the sample has huge instability. This is due to the effects of
financial crisis in Asia, Russia. Brazilian economy had worked with an almost fixed
exchange rate regime that generates great volatility in interest rates. After 1999 the interest
rates have smoother movements. There is also a tendency of interest rate falling in long run.

Figure 1: Taxa de Juros de 30, 60, 90, 180 e 360 dias.

1
It is possible to work with different error variances, allowing for heteroscedasticity. Since the sample
is not big enough to accommodate all this heavy structure. It must be noted that a time varying short structure
in the models allow for different unconditional variances.
2
where ρ (t ) denotes the deterministic terms as intercept and time trends.
6
 Li30 Li60
Li90 Li180
Li360
0.35

0.30

0.25

0.20

0.15

1996 1997 1998 1999 2000 2001 2002 2003 2004 2005 2006 2007

3.4 Report of the estimation results:

Figure 2: shows the evolution of the spreads for different vertices. There is great
volatility from the beginning of the sample until 1998. After the changes in exchange rate
regime in January of 1999, the spreads are less volatile and with an apparent large memory.

Figure 2: Spreads for different maturities.

S60_30 S90_30
S180_90 S360_180
0.03

0.02

0.01

0.00

-0.01

-0.02

1996 1997 1998 1999 2000 2001 2002 2003 2004 2005 2006 2007

Based on exogenous information about the evolution of the Brazilian economy it

was proposed the existence of four possible regimes:

7
 • regime 1: from 1995:1 to 1997:9 – (macroeconomics stabilization process);
• regime 2: from 1997:9 to 1998:12 – (Asian Crisis);
• regime 3: from 1999:1 to 2006:12 (floating exchange rate era and second Cardoso
administration and first Lula administration);
The following models were estimated:
• Model 01: Unrestricted with three complete different regimes.

1 1 0 0
− 1 0 0 0 

 0 −1 1 0
• Model 02: β ' (1) = β ' ( 2) = β ' (3) =  
0 0 −1 1 
0 0 0 − 1
 
 ρ1 ρ 2 ρ 3 ρ 4 

• Model 03: Restrictions imposed in model 02 plus α (1) = α (2) = α (3) = α .

There is no structural change in first moment of the process.

Table 1: Likelihood ratio test for proposed simplifications.

r m lags p D Number of parameters Log-Likelihood

Model 1 4 3 2 5 1 138 457.79
Model 2 4 3 2 5 1 122 455.74
Model 3 4 1 2 5 1 82 380.09
Model 4 4 2 2 5 1 117 452.81

In Table 2 the simplifications from the general model are tested. The hypothesis the
spreads and the average term premium does not differs significantly in the three regimes is
not rejected from the data (line 1). The hypothesis that there is no structural change in first
moment of the process is strongly rejected from the data (line 3).
Then using this structure it is possible to document the existence of three different
regimes. The first starts in 1995:1 and ends in 1997:9. The second starts in 1997:10 and
ends in 1998:12. The third regime starts in 1999:1 and lasts until the end of the sample.
Finally it was tested whether or not the common trends remains the same across the
regimes. For this to happen it must exist a matrix (a) that multiplied by α(t) in all regimes

8
that results in zero (a’* α(t)=0). If α is constant in all regimes then the common trends are
the same in all regimes. Suppose that this structure is correct α(t)= α∗∗ρ(t) where α∗ is a p
x r matrix and ρ(t) is a r x r matrix, then it’s possible to find a matrix a that satisfies the
restrictions a’* α(t)= a’* α∗∗ρ(t)=0.
The model under this restriction can be estimated using GRRR and a likelihood
ratio test can be formulated to test the validity of this hypothesis. Then the hypothesis that
the common trends remain the same during the period of 1997:9 to 2006:12 is not rejected
by the data (line 3). Despite the fact that there is evidence of structural change under the
whole period, the change does not affected the long run implications of the expectation
hypothesis. The spreads are found to be stationary and the factors that drive the interest rate
remains intact during the whole sample.
The evidence of a changing structure can be rationalized under the McCallum (1994)
framework. In this paper he had shown that the spreads can have memory if the risk
premium evolves as a process that has memory (as an autoregressive process for example).
A change in the risk pattern or the rules followed by central bank can imply in a relation
between spreads and first difference of short rates that change throughout the time.
To sum up the evidence in this paper is favorable of the expectation hypothesis if
it’s assumed a time variant risk premium. However it must be noted that not all the
implications of the expectation hypothesis is tested. For example, if the expectation
hypothesis holds under McCallum (1994) framework the spreads can be approximated by
an autoregressive process but only the past value of the spreads can help to predict the
current values of the spreads. (See for example Serna and Arribas (2006) and Gallmeyer,
Hollifield and Zin (2005))

Table 2: Likelihood ratio test for proposed simplifications.

Ha: Ho: Likelihood Degree of p-value
Ratio freedom

Line 1 Model 1 Model 2 4.10 16 99.87%

Line 2 Model 1 Model 3 155.41 21 0.00%
Line 3 Model 1 Model 4 9.96 21 97.94%

9
Table 3: Cointegrated vectors estimated from the best model.
i30 1 1 0 0
i60 -1 0 0 0
i90 0 -1 1 0
i180 0 0 -1 1
i360 0 0 0 -1
constant 0.08% 0.18% 0.18% 0.57%

The poor performance of expectation hypothesis for Brazilian data reported in the
literature is related to non-modeled conditional heterocedasticity. In terms of long run
implications of the expectations hypothesis it was not possible to reject that the short
spreads are stationary in all regime: spread (60, 30), spread (90, 30). The spread (180, 90)
and spread (360, 180) are found to be stationary in the whole sample. In the first sample the
hypothesis that these spreads are stationary are rejected. One possible tentative explanation
to this fact is the conjunction of the effects of Mexican crisis and stabilization process in
Brazilian economy started in 1994 on term structure did not encourage the agents to
arbitrage the difference between short and long run rates due to risk considerations
(expected devaluation of Brazilian currency and inflation concern). The estimated
cointegration vectors are reported in Table 4.

4 Comparison with other papers:

Hansen (2003) implemented the same test for American data. The author divides the
sample in 3 periods. There are some differences from this work. He controls for
heteroscedasticity but imposes the same short-run structure to all regimes. It is not possible
to reject the hypothesis of stationary of the spreads. The constant risk premium hypothesis
across the sub-samples is rejected. The hypothesis that the common trend remains the same
is not rejected.
The work of Hansen (2003) implies in an evidence in favour of the long-run
implications of the expectation hypothesis. The spreads are stationary and this avoids
arbitrage opportunities at least in the long-run. The common trends are the same in all
regimes which implies long-run movement of interest rate are caused by the same factors.

10
 The study of the term structure for Brazilian data is at its early stages. The following
works uses cointegration techniques to test the expectation hypothesis: Brito, Guillen and
Duarte (2004), Lima and Isler (2003), Marçal (2004) and Marçal and Valls Pereira (2007).
Just Marçal (2004), Marçal and Valls Pereira (2007) tries to model more than two vertices
simultaneously using a cointegrated VAR. They have found evidence of cointegration but
the evidence of stationary for longer spreads is weak.
Brito, Guillen and Duarte (2004) use cointegration techniques to test the expectation
hypothesis. They have worked with daily data from 01/07/1996 to 31/12/2001. The
cointegration hypothesis is tested using Johansen cointegration test. The authors conclude
that the cointegration vector is equal to the spreads and this validates the expectation
hypothesis. The frequency of the data is daily while the frequency of this work is monthly.
In high frequency data it’s more likely to find some sort of conditional heteroscedasticity. If
some conditions are not satisfied the trace and maximum eigenvalue statistics cannot be
used (Rahbek, Hansen and Dennis (2002)). Under monthly data the conditional
heteroscedasticity can be seen as a minor problem. But the fact is that the result of this
work is different from their work.
Lima and Isler (2003) uses the ADF and Phillips-Perron test to evaluate whether or
not the spreads are stationary. They have used monthly data from January of 1995 to
December of 2001. They have obtained evidence in favour of no unit root in spreads (Lima
and Isler (2003), p. 886). These results are confirmed by bivariate cointegration tests Lima
and Isler (2003), pages. 888 e 889) applied to short and long interest rate. No effort to build
up a VEC was done.
Marçal (2004) and Marçal and Valls Pereira (2007) test the expectation hypothesis
using the methodology employed by Campbell and Shiller (1991) as well as cointegration
techniques. Not just the long run implications of the expectation hypothesis is tested but
also the short-run implications implied by the theory in line with the works of Johansen and
Swensen (1999) and Johansen and Swensen (2003). The results are not favorable to the
expectation hypothesis particularly for longer spreads. It is also tested which vertices
causes the long-run movements of the interest rate in line with the work of Gonzalo and
Granger (1995). It was found that the common trends contain just elements related to long
rate series.
11
5 Conclusion:
This work investigates the evidence of structural change in Data Generator Process
of Brazilian term structure of interest rate data. It is documented the presence of 3 regimes.
The first lasts from 1995:1 to 1997:9. The second lasts from 1997:10 to 1998:12. The third
lasts from 1999:1 to the end of the sample (2006:12).
The cointegration vector equals the spreads in all regimes and the common trends
remains unchanged despite the evidence of structural change. Future studies of term
structure using Brazilian data should be aware of the differences among the three periods
and particularly with the data sooner after the Real Plan (particularly before 1997). As there
is the ‘1979-1982 data problem’ in American term structure of interest rate (Seo (2003),
Hansen (2003) and Cuthbertson and Nitzsche (2005)), it seems reasonable to state that there
is a similar data problem for Brazilian data (post Real Plan and Asian Crisis data - 1995:1
to 1998:12) but the long-run implications of expectation hypothesis is satisfied.

6 References:
Boswijk, H. Peter, and Y. Zu, 2005, Testing for Cointegration with Nonstationary
Volatility, (Tinbergen Institute & Amsterdam School of Economics).
Brito, R. D., O. T. C. Guillen, and A. J. M. Duarte, 2004, Overreaction of yield spreads and
movements of Brazilian interest rate, Revista de Econometria 24, 1-55.
Campbell, J, and R. J. Shiller, 1991, Yield Spreads and Interest Rate Movements: A Bird's
Eye View, Review of Economic Studies 58, 419-514.
Cuthbertson, Keith, and Dirk Nitzsche, 2005. Quantitative Financial Economics (John
Wiley & Sons Ltd., West Sussex).
Gallmeyer, Michael F., Burton Hollifield, and Stanley E. Zin, 2005, Taylor Rules,
McCallum Rules and the Term Structure of Interest Rates, NBER Working Paper
(NBER, Cambridge).
Gonzalo, J., and C. W. J. Granger, 1995, Estimation of Common Long-Memory
Components in Cointegrated Systems, Journal of business and Economics Statistics
13.
Hansen, Peter Reinhard, 2003, Structural changes in the cointegrated vector autoregressive
model, Journal of Econometrics 114, 261-295.
Hicks, John, 1946. Value and Capital (Oxford University Press, Oxford).
Johansen, Soren, 1988, Statistical Analysis of cointegration vectors, Journal of Economic
Dynamics and Control 12, 231-254.
Johansen, Soren, and Anders Rygh Swensen, 1999, Testing exact rational expectations in
cointegrated vector autoregressive models, Journal of Econometrics 93, 73-91.
Johansen, Soren, and Anders Rygh Swensen, 2003, More on Testing Exact Rational
12
 Expectations in Cointegrated Vector Autoregressive Models:Restricted Drift Terms,
Discussion Papers 348 (University of Oslo, Oslo).
Keynes, J. M., 1930. Treatise on Money.
Lima, A. M., and J. V. Isler, 2003, A hipótese das expectativas na estrutura a termo da taxa
de juros no Brasil: Uma aplicação de modelos a valor presente, Revista Brasileira
de Economia 57.
Marçal, E. F., 2004, Ensaios sobre eficiência, cointegração, componentes comuns, não
linearidades na variância nos mercados financeiros: um estudo da estrutura a termo
das taxas de juros e da volatilidade de títulos da dívida soberana., Departamento de
Economia (Univesidade de São Paulo, São Paulo).
Marçal, E. F., and P. L. Valls Pereira, 2007, A Estrutura a Termo das Taxas de Juros no
Brasil: testando a hipóteses de Expectativas, Pesquisa e Planejamento Econômico
37, 113-147.
McCallum, B. T., 1994, Monetary Policy and the Term Structure of Interest Rate,
Economic Quartely 91, 1-21.
Rahbek, Anders, Ernst Hansen, and Jonathan G. Dennis, 2002, ARCH Innovation and their
impact on Cointegration Rank Testing, (University of Copenhagen, Copenhagen).
Seo, Byeongseon, 2003, Non linear mean reversion in the term structure of interest rate,
Journal of Economic Dynamics and Control 27, 2243-2265.
Serna, Maria Isabel Martinez, and Eliseo Navarro Arribas, 2006, The expectations theory of
the term structure of interest rates and monetary policy, Social Science Reasearch
Network (SSRN).
Shiller, R. J., 1990, The term structure of interest rate, in Benjamin M. Friedman, and Frank
H. Hahn, eds.: Handbook of Monetary Economics (North-Holland).
Tse, Y. and A. Tsui, 2002, A Multivariate GARCH Model with Time-Varying
Correlations, Journal of Business and Economic Statistics,20, 351-362.

13