Regime Switches in Interest Rates

Andrew Ang

Geert Bekaert

First Version: 25 March 1998

This Version: 16 July 2001
Keywords: regime switching models, interest rates,
term structure, business cycles, forecasting.

The authors would like to thank Rob Engle, Eric Ghysels, Ken Singleton, Jun Liu, Liuren Wu, two
anonymous referees and seminar participants at Stanford University and the 1999 Econometric Society
meetings for many useful comments. Geert Bekaert acknowledges nancial support from an NSF grant.

Columbia Business School and NBER, 3022 Broadway 805 Uris, New York, NY 10027.
(aa610@columbia.edu)

Columbia University and NBER, 3022 Broadway 802 Uris, New York NY 10027.
(gb241@columbia.edu.)
Abstract
We examine the econometric performance of regime switching models for interest rate data
from the US, Germany and the UK. Regime switching models forecast better out-of-sample
than single regime models, including an afne multi-factor model, but do not always match
moments very well. Regime switching models incorporating international short rate and term
spread information forecast better, match sample moments better, and classify regimes better
than univariate regime switching models. Finally, the regimes in interest rates correspond rea-
sonably well with business cycles, at least in the US.
1 Introduction
The stochastic behavior of interest rates varies over time. For example, the behavior of interest
rates in the 1979-1982 period in the US or around the German reunication period seems to
indicate a structural break in the time series. More generally, changes in business cycle condi-
tions and monetary policy may affect real rates and expected ination and cause interest rates
to behave quite differently in different time periods. Regime-switching (RS) models constitute
an attractive class of models to capture these changes in the stochastic behavior of interest rates
within a stationary model. Many authors have built on the seminal work of Hamilton (1989) to
model short rates by a model where the parameters change over time driven by a Markov state
variable (assumed to be unobserved to the econometrician). For example, Hamilton (1988),
Lewis (1991), Evans and Lewis (1994), Sola and Drifll (1994), Garcia and Peron (1996),
Gray (1996) and Bekaert, Hodrick and Marshall (2001) all examine empirical models of regime
switches in interest rates.
Importantly, RS models accomodate regime-dependent mean reversion of interest rates.
Mankiw and Miron (1986), among others, argue that the predictive power of the term spread for
future short rates in the US is very much a function of the monetary policy regime. In particu-
lar, they argue that the interest rate smoothing efforts of the Federal Reserve Bank make the US
short rate behave like a random walk, and this behavior causes rejections of the Expectations
Hypothesis. When a regime switching model is tted to US data however, Bekaert, Hodrick
and Marshall (2001) and Gray (1996) show that such random walk behavior is only true for
low interest rates whereas high interest rates show considerable mean reversion. Several au-
thors (Cecchetti, Lam and Mark 1993 and Garcia 1998) show that single regime models are
econometrically rejected in favor of their RS counterparts.
Despite their economic appeal, RS models are less attractive than one-regime models from
an econometric estimation perspective. Although with the recent work of Gray (1996) and
Hamilton (1994) the likelihood construction has been simplied, estimating RS models is dif-
cult. Often, the data do not allow clear regime-classication, that is, the probability of having
observed a regime ex-post may hover around a half. These problems may explain why there are
few RS term structure models of interest rates (see Naik and Lee 1994, Evans 1998, and Bansal
and Zhou 1999).
In this paper, we provide an analysis of the econometric properties of RS models, both with
constant and state-dependent transition probabilities, for interest rates in the US, Germany and
the UK. Apart from residual diagnostic tests, we use two statistical criteria to compare and rank
alternative one-regime and RS models of short rates. The rst criterion investigates the t of the
1
models with the unconditional moments of the data. One attraction of RS models is that they
may accommodate some of the non-linearities in interest rates which may show up in higher or-
der unconditional moments (see At-Sahalia 1996, Stanton 1997, and Ahn and Gao 2000). The
dependence of mean reversion on the level of the interest rate may also induce an autocorrelo-
gram that is difcult to match by parsimonious ARMA models. The second criterion concerns
the forecasting power of the different models, both for rst and second moments. Finally, we
propose a new metric to compare the performance of different RS models in identifying the
regime over the sample. Our Regime Classication Measure (RCM) uses the simple fact that
the ex-post probability of observing one of the regimes ought to be close to one at all times
when regime classication is perfect.
Given the econometric problems mentioned above, it is not a priori clear that RS models
perform well on these statistical criteria, even when they are the true data generating process
(DGP). Moreover, as Bekaert et al. (2001) stress, the estimation may suffer from a peso prob-
lem, in that the fraction of observations drawn from one particular regime in the sample at hand
may not correspond to the population frequency of that regime. In that case, the estimation is
biased. For example, it is unlikely that we could get a reliable estimate of the mean reversion at
large interest rates in US data, without including the 1979-1982 period. Furthermore, ARMA
models may generally constitute good approximations to any covariance stationary process and
hence may outperform RS models in small samples, if the parameter estimates of the RS models
are severely biased and inefcient.
To help overcome these problems, we extend the effective sample size through two chan-
nels. First, we investigate multi-country systems of interest rates. It is possible that short rates
in the US Granger-cause rates in other countries (or vice versa) and that Granger-causality may
be regime-dependent. Whereas such relations would immediately affect the forecasting perfor-
mance, we may also obtain more efcient estimates if interest rate innovations across countries
are correlated. If some parameters are identical in different countries, further gains in efciency
are to be expected. The model we propose and estimate allows for correlated interest rate in-
novations and Granger-causality between rates in some regimes. We compare the performance
of several variants of the multivariate RS models to their single regime vector-autoregressive
(VAR) counterparts, and to one multi-factor model in the afne term structure class.
Second, we exploit information in the term structure, by adding term spreads to the model.
Under the null of the Expectations Hypothesis, spreads should forecast future short rates, so the
potential for improved performance is obvious. The moments criteria here include the cross-
correlations between short rates and spreads. As Pfann, Schotman and Tscherning (1996) show,
the correlation between short rates and long rates changes with the level of the interest rate,
2
suggesting the correlation may be informative about the regime.
Apart from a number of methodological contributions, this article offers some important
empirical results. First, while RS models do not always outperform single regime models in the
in-sample diagnostics, they forecast very well out-of-sample. Second, multivariate RS models
perform better than univariate models in terms of regime classication and forecasting. The
best forecasting model is invariably a multivariate RS model. Hence, our results greatly expand
on Gray (1996), who examines the out-of-sample forecasting power of a univariate RS model
for second moments of the US short rate. Third, the regime classication implied by RS models
is closely related to economic business cycles and the ex-ante regime probabilities are good
short-horizon predictors of the business cycle in the US.
The paper is organized as follows. Section 2 describes the data and establishes a set of styl-
ized facts. Section 3 outlines the general empirical and econometric framework and discusses
our diagnostic statistics. It presents a general multivariate RS model and considers as special
cases univariate short rate models, multi-country models of the short rate and bivariate short
rate and term spread models for each country. A stark implication of the framework is that
univariate models generally cannot be consistently estimated. Section 4 briey discusses the
empirical estimation results and Section 5 discusses the performance of the various models. To
help interpret the results we perform a Monte Carlo experiment that examines the performance
of single regime and RS models in small samples when the true DGP is a RS model. We con-
sider the quality of regime classication and ask if the regimes are related to the business cycle
in Section 6. Section 7 concludes.
2 Data and Stylized Facts
Our empirical work uses monthly observations on 3 month short rates and 5 year long rates of
zero coupon government bonds from the US, Germany and Great Britain from January 1972 to
August 1996. The data set combines data from Jorion and Mishkin (1991) with a proprietary
data set of zero coupon rates (see Bekaert et al. 2001). We denote the short rates as r
m
t
and
the spreads as z
m
t
for country m. We estimate models based on an in-sample period, with
forecasting done on an out-of-sample period of the last 30 months. Hence, our in-sample period
has 267 observations.
Table 1 reports the rst four central moments of the short rates and spread data on the in-
sample period. The table also shows the autocorrelations for each country, the cross-correlations
of short rates for each pair of countries and correlations of short rates and spreads within each
3
country. We note that the short rates for Germany and Great Britain do not showexcess kurtosis.
Short rates are very persistent, with the UK showing the least persistence. Spreads are also
autocorrelated, but less so than short rates. Turning to international cross-correlations, lagged
short rates of the US are more highly correlated with current German and UK rates than present
levels of US short rates. This suggests that lagged US short rates may Granger-cause movements
in short rates in Germany and the UK. The contemporaneous correlations of short rates across
countries are not very high except for the US and UK rates.
In Table 2 we attempt to determine whether the behavior of the term structure depends
on the business cycle. For the US, we use the NBER dates for business cycle expansions
and contractions, and dates for the Germany and UK are from for the US can be found at
www.nber.org/cycles.html, dates for Germany and the UK are from the Center for International
Business Cycle Research at Columbia University (see Zarnowitz 1997). The table divides the
interest rate observations into periods of expansions and contractions and performs
2
tests
for the equality of various moments assuming independence across the cycles. As Zarnowitz
(1997) notes, only the US has a business cycle history which is ofcial, in the sense of being
accepted by governmental authorities, and the dating of the cycles for other countries is less
reliable. This means we must interpret the results for Germany and the UK with caution.
Focusing on the country with the best cycle dating, the US, Table 2 reveals that recessions
are characterized by signicantly higher interest rates, and somewhat more variable interest
rates. The variability is, somewhat surprisingly, not signicantly different across expansions
and recessions. Interest rates in expansions exhibit higher kurtosis than in recessions and they
are signicantly less mean-reverting. Spreads are lower and more variable in recessions but only
the mean of the spread is signicantly different across cycles. In recessions there is signicantly
more skewness (or a lack of negative skewness) and spreads are more mean-reverting.
These patterns are not perfectly replicated in Germany and the UK. In these countries auto-
correlations of the short rate and spread are not signicantly different across the business cycle.
In Germany, the patterns are similar to the US, except for mean reversion which is insigni-
cantly higher in expansions. In the UK, the volatility of both spreads and interest rates is higher
in expansions, although the p-values are not very low. Although the point estimates of mean
reversion follow the same pattern as the US, the differences across cycles are not statistically
signicant.
Finally, in the US and UK the correlation between the short rate and the spread varies over
the business cycle. The difference in correlations suggests that in expansions the long rate is
4
less sensitive to short rate shocks than in recessions. To see this, note that:
(r
l
t
, r
t
) = w
1
[w
2
(z
t
, r
t
) + 1] (1)
where w
1
= (r
t
)/(r
l
t
), w
2
= (z
t
)/(r
t
) which is less than 1 empirically, r
t
is the short rate,
z
t
is the spread, r
l
t
is the long rate, and (x, y) is the correlation between x and y. In expansions,
(z
t
, r
t
) is more negative and correspondingly the correlation between short and long rates is
lower.
For the US, the picture that emerges is one where in expansions, short rates are more persis-
tent, the long rate is not as sensitive to short rate shocks and the short rate-spread correlation is
more negative. In expansions, the interest rate persistence may arise from the smoothing efforts
of the monetary authorities. In recessions, long rates are more sensitive to short rate shocks de-
spite the lower persistence of short rates. Here, shocks to the short rate are more likely to move
the whole term structure. The difference in the short rate-spread correlation across expansions
and recessions is signicant at the 5% level in the US, but only signicant at the 10% level in
the UK and not signicant in Germany. However, the pattern of the short rate-spread correlation
across expansions and recessions in the UK is quantitatively similar to the pattern in the US.
Overall, Table 2 implies the following points about the behavior of interest rates across
the business cycle. First, the moments of interest rates vary from recessions to expansions;
in particular, the mean is higher in recessions. Second, the spread is informative about the
regime, with the spread increasing during expansions and correlations between the spread and
short rate changing across the business cycle. Third, mean reversion in the US is signicantly
different across economic regimes. These patterns can potentially be accomodated in models
which contain a regime variable.
3 The Empirical and Econometric Framework
3.1 A General Multivariate Regime Switching Model
We describe a general multivariate RS model of short rates r
t
= (r
us
t
, r
ger
t
, r
uk
t
)

and spreads
z
t
= (z
us
t
, z
ger
t
, z
uk
t
)

. Let y
t
= (r

t
, z

t
)

. We assume that the information set I

t
for our econo-
metric model is composed of [y

t
, y

t1
, . . . ]

. Our most general model is a RS Vector Autore-

gression (VAR):
y
t
= (s
t
) +A(s
t
)y
t1
+
1
2
t1
(s
t
)
t
, (2)
where s
t
denotes the regime realization at time t, and
t
IID N(0, I). We restrict attention to
rst-order VARs since in our empirical work we usually estimate at most rst-order systems.
5
The process s
t
follows a Markov chain with K regimes, and with transition probabilities
which may be logistic functions of lagged endogenous variables:
p(s
t
= i|s
t1
= j, I
t1
) =
e

i,j
+

i,j
y
t1
1 +e

i,j
+

i,j
y
t1
. (3)
Let y
T
= (y

T
y

T1
. . . y

1
y

0
)

and denote the parameters of the likelihood by . Then fol-

lowing the methodology of Hamilton (1994) we write the conditional likelihood as:
f( y
T
; ) =
T

t=1
_
K

i=1
f(y
t
|I
t1
, s
t
= i; )p(s
t
= i|I
t1
; )
_
. (4)
The ex-ante probability p
it
= p(s
t
= i|I
t1
; ) can be written as:
p
it
=
K

j=1
p(s
t
= i|s
t1
= j, I
t1
; )
_
f(y
t1
|s
t1
= j, I
t2
; )p(s
t1
= j|I
t2
; )

K
m=1
f(y
t1
|s
t1
= m, I
t2
; )p(s
t1
= m|I
t2
; )
_
, (5)
where the rst term in the sum is the transition probability which can be state-dependent, and
the other terms follow from Bayes Rule.
We start the algorithm using (5) with p(s
1
= i|I
0
) equal to the ergodic probabilities of the
system at t = 1 given by:

i
=
X
ii

K
j=1
X
jj
, (6)
where X
ii
is the ii
th
cofactor of the matrix X = I P
1
, and P
1
is the KxK transition matrix
of the system at t = 1 which can depend on our conditional information set I
0
. In the special
case of constant transition probabilities we start at the ergodic probabilities of the transition
matrix P which solve = P

.
3.2 Special Cases
Since the regime-variable is unobserved to the econometrician and must be factored out of the
likelihood function, under what conditions we can obtain inefcient but consistent estimates
when ignoring some variables? Let Z
t
represent variables which do not enter into our estimation
and X
t
represent variables which do, so y
t
= (Z

t
, X

t
)

. Using conditioning arguments we can

6
write:
f( y
T
; ) =
T

t=1
f(y
t
|I
t1
; )
=
T

t=1
_
K

i=1
f(y
t
|s
t
= i, I
t1
; )p(s
t
= i|I
t1
; )
_
=
T

t=1
_
K

i=1
f(Z
t
|X
t
, s
t
= i, I
t1
; )f(X
t
|s
t
= i, I
t1
; )p(s
t
= i|I
t1
; )
_
. (7)
To take f(Z
t
|X
t
, s
t
= i, I
t1
; ) out of the sum, assume that the excluded variables do not
depend on the regime:
f(Z
t
|X
t
, s
t
= i, I
t1
; ) = f(Z
t
|X
t
, I
t1
; ). (8)
We parameterize the model so that = (

X
)

and {
Z
} {
X
} = , where
Z
and
X
affect
the conditional distribution of the excluded variables and the included variables respectively.
We also assume that the ex-ante probability of being in a particular regime depends only on
X
:
p(s
t
= i|I
t1
; ) = p(s
t
= i|I
t1
;
X
). (9)
The likelihood can be written:
L( y
T
; ) =
T

t=1
log f(Z
t
|X
t
, I
t1
;
Z
) +
T

t=1
log
_
K

i=1
f(X
t
|s
t
= i, I
t1
;
X
)p(s
t
= i|I
t1
;
X
)
_
. (10)
Maximizing the second sum in (10) yields consistent but inefcient estimates relative to full
information maximum likelihood.
Estimation of the full systemis infeasible given the dimension of , so we focus on models of
subsets of the variables. Our choice here is partially based on previous literature and partially on
economic reasoning. We believe that regimes in either real rates, expected ination or business
cycles are the source for potential regimes in nominal interest rates (see Garcia and Perron 1996
and Evans and Lewis 1995). To obtain parsimony in modeling, we assume the existence of a two
state Markov regime variable in every country driving the entire term structure. These country
specic regime variables are assumed independent across countries. It is conceivable that there
is a world business cycle driving interest rates in many countries simultaneously and in some
of the models we consider we allow for interdependence of various forms across countries.
Nevertheless, it should be noted that the correlation between spreads and short rates within
7
a country is typically of a higher magnitude than the correlation of short rates and spreads
across countries (see Table 1) providing empirical motivation for this assumption. Although
the two regime specication may seem restrictive, it is the most the data can bear without
extreme computational problems in estimation, and it sufces to capture the main empirical non-
linearities. In particular, Ang and Bekaert (2000) show that two-state RS models can replicate
the non-parametric drift and volatility functions of the short rate estimated by At-Sahalia (1996)
and Stanton (1997). Finally, most of the past RS literature has focused on two-state models,
with the exception of Garcia and Perron (1996) and Bekaert et al., who estimate three state RS
models.
Since most of the RS literature focuses exclusively on univariate interest rate models, we
start by analyzing univariate short rate models for the US, Germany and UK. As (8) shows, to
consistently estimate univariate short rate RS models, the distribution of the term spreads or
short rates from other countries should not depend on the regime of the short rate we consider.
If regimes capture business cycle effects, the different correlations in the US across economic
cycles in Table 2 violate the assumptions needed for consistent estimation.
Incorporating the extra information from international and term structure data allows us to
weaken the implicit assumptions but makes estimation much more complex. In a second set
of models, we add information from the short rates from other countries. In our multi-country
model (below), dening the regime variable s
t
becomes more involved as it embeds all possible
combinations of the country-specic regime variables for the three countries.
Finally, we consider models in which term spreads are added to the short rate and their
dynamics remain driven by one country-specic regime variable. In most term structure mod-
els, the term spread is an exact function of a number of factors that also drive the short rate.
However, the evidence from a growing literature looking at the response of the term structure
to various shocks, suggests that the spread contains additional independent information which
may help in the classication of regimes. For example, Evans and Marshall (2000) show that
monetary policy shocks have large effects on the short rate but leave the long rate unaffected,
hence shrinking the spread. However, shocks from real economic activity affect the whole
term structure and correspond to a level effect increasing the interest rate but leaving the spread
largely unaffected. Estrella and Mishkin (1997) nd that the spread is useful in predicting future
activity, and the spread contains predictive information which is not captured by other monetary
policy variables. A reduced-form model where the spread and short rate have correlated innova-
tions and different feedback rules, in which spreads help predict future regimes, may be a good
representation of such a world. We estimate the short rate-spread model country by country but
also consider one estimation which uses cross-country information.
8
Table 3 presents a summary of the models estimated, their abbreviations used throughout
the paper and the number of parameters in parentheses. We now brieftly outline each of these
models. (Parameter estimates are available in an Appendix which is available from the authors
on request.)
3.2.1 Univariate Models
For each country m, we consider special cases of the following general model considered in
Gray (1996):
r
m
t
= (s
m
t
) +(s
m
t
)r
m
t1
+h
m
t1
(s
m
t
)
t
, (11)
where
t
IID N(0, 1). The conditional volatility h
m
t1
(s
m
t
) is specied as:
(h
m
t1
(s
t
))
2
= a
0
(s
m
t
) +a
1
(s
m
t
)
2
t1
+b
1
(s
m
t
)(h
m
t2
)
2
+b
2
(s
t
)(r
m
t1
), (12)
where (h
m
t1
)
2
= E
t1
[(r
m
t
)
2
] (E
t1
[r
m
t
])
2
and
t
= r
m
t
E
t1
[r
m
t
]. The regime variable s
t
is
either 1 or 2, and has transition probabilities:
p(s
m
t
= j|s
m
t1
= j, r
m
t1
) =
e
a
j
+b
j
r
m
t1
1 +e
a
j
+b
j
r
m
t1
, j = 1, 2. (13)
We denote constant transition probabilities as P and Q for j = 1, 2 respectively. We evaluate
E
t1
[r
m
t
] and E
t1
[(r
m
t
)
2
] as:
E
t1
[r
m
t
] =
2

j=1
p
t,j
(
j
+
j
r
m
t1
)
E
t1
[(r
m
t
)
2
] =
2

j=1
p
t,j
_
(
j
+
j
r
m
t1
)
2
+ (h
m
t1,j
)
2
_
, (14)
where subscripts indicate the state s
m
t
= j.
The special cases we consider involve setting a
1
= b
1
= b
2
= 0 (RS AR(1)), b
2
= 0
(RS GARCH(1,1)), a
0
= a
1
= b
1
= 0 (RS CIR). The last model is the RS equivalent of the
discretized square root model of Cox, Ingersoll and Ross (1985).
In practice, many interest rate RS models yield one unit-root or near unit-root regime, and
one more mean-reverting regime. Ang and Bekaert (1998) and Holst, Lindgren, Holst and
Thuvesholmen (1994) prove that such processes retain covariance stationarity as long as the
unconditional autocorrelation is strictly less than one. This is guaranteed by appropriate mixing
of the two regimes. With constant transition probabilities, a sufcient condition is that the
ergodic probability associated with the stationary regime is non-zero.
9
3.2.2 Multi-Country Models
For r
t
= (r
us
t
r
ger
t
r
uk
t
)

, we consider the following general multi-country RS model:

r
t
=
_
_
_
_

us
(s
us
t
)

ger
(s
ger
t
)

uk
(s
uk
t
)
_
_
_
_
+A(s
us
t
, s
ger
t
, s
uk
t
)r
t1
+
1
2
(s
us
t
, s
ger
t
, s
uk
t
)
t
, (15)
with
t
= (
us
t

ger
t

uk
t
)

IID N(0, I).

We assume that there are two regimes per country with constant probabilities, so for country
mthe transition matrix is
_
P
m
1P
m
1Q
m
Q
m
_
. For computational tractability, and to keep the number
of parameters as small as possible, we do not consider state-dependent transition probabilities
in the multi-country model.
We assume regimes in different countries to be independent of the regimes in another coun-
try. Formally, let S
m
= {s
m
t
, s
m
t1
, . . . } denote the past history of regimes for country m. Then:
p(s
m
t
|S
us
t
, S
ger
t
, S
uk
t
) = p(s
m
t
|S
m
t
) = p(s
m
t
|s
m
t1
). (16)
Intuitively this means that the regime for one country is unaffected by the regime in another
country. We may justify this by interpreting the regimes as arising from country specic fac-
tors. This independence assumption can only be relaxed at considerable computational cost and
proliferation of parameters. With 2 regimes for 3 countries, it is possible to enlarge the state
space to 2
3
= 8 regimes, where the regimes are dened as s
t
= 1, . . . , 8 (see Hamilton 1994):
s
t
US GER UK
1 1 1 1
2 2 1 1
3 1 2 1
4 2 2 1
5 1 1 2
6 2 1 2
7 1 2 2
8 2 2 2
We then calculate an 8x8 transition matrix, where for example, p(s
t
= 1|s
t1
= 1) = P
us
P
ger
P
uk
.
With the regimes now redened as s
t
= 1, . . . , 8, we re-write (15) as:
r
t
= (s
t
) +A(s
t
)r
t1
+u
t
, (17)
where u
t
N(0,
1
2
(s
t
)). From hereon subscript is refer to the values each specic countrys
state comprises in the overall state i. For example, for s
t
= 4:
_

us
4

ger
4

uk
4
_
=
_

us
(s
us
t
=2)

ger
(s
ger
t
=2)

uk
(s
uk
t
=1)
_
.
10
Given the number of parameters, estimation of the full model is infeasible. To gain ef-
ciency, we test whether some parameters are identical in the one-regime VAR. In particular, we
test for Granger-causality on each countrys short rates. These results are presented in Table
4. The table shows that a joint test for no country Granger-causing another just fails to reject
(p-value = 0.0528). Nevertheless, there is some evidence that US rates Granger-cause German
and UK rates (p-value = 0.0029).
The results of Table 4 lead us to consider consider two formulations of A
i
, a triangular
formulation where A
i
=
_

us
i
0 0

ger
i

ger
i
0

uk
i
0
uk
i
_
, which we refer to as a Granger-causality formulation,
and a diagonal formulation where A
i
=
_

us
i
0 0
0
ger
i
0
0 0
uk
i
_
.
To impose further structure on the error terms, we model the errors as:
_
_
_
_
u
us
t,i
u
ger
t,i
u
uk
t,i
_
_
_
_
=
_
_
_
_
h
us
t1,i

1
t
h
ger
t1,i

2
t
+
ger
i

1
t
h
uk
t1,i

3
t
+
uk
i

1
t
_
_
_
_
, (18)
where
t
= (
1
t
,
2
t
,
3
t
)

are drawn from an IID N(0, I) distribution and the conditional volatility
of country m, h
m
t1,i
, is specied either as a constant, h
m
t1,i
=
m
i
or as a square root process,
h
m
t1,i
=
m
i
_
r
m
t1
. In this specication the errors from the US also shock the interest rates
of Germany and the UK, but not vice versa. Another interpretation along the lines of a world
business cycle is that there are world shocks which drive the dominant US economy while
Germany and the UK are also subject to these shocks as well as country-specic shocks.
The extent to which these countries are exposed to the world shock depends on the state of
the domestic economy. Given the dominance of the US in the world economy such a structure
seems reasonable. The conditional covariance matrix, conditional on state s
t
= i is given by:

t
(s
t
= i) = E[u
t
u

t
|I
t1
, s
t
= i]
=
_
_
_
_
(h
us
t1,i
)
2

ger
i
h
us
t1,i

uk
i
h
us
t1,i

ger
i
h
us
t1,i
(h
ger
t1,i
)
2
+ (
ger
i
)
2

ger
i

uk
i

uk
i
h
us
t1,i

uk
i

ger
i
(h
uk
t1,i
)
2
+ (
uk
i
)
2
_
_
_
_
. (19)
This specication arises because the errors u
m
t,i
inherit a multivariate normal distribution from
the normality of the errors
m
t,i
. Note that German and UK shocks are conditionally correlated
to the extent only that they correlate with the US shock.
It is possible to obtain probability inferences for a particular country by summing together
the relevant joint probabilities. For example if we want the ex-ante probability p(s
us
t
= 1|I
t1
)
we just sum over the probabilities p(s
t
|I
t1
) where s
us
t
= 1. In this case, we would sum over
states s
t
= 1, 3, 5, 7.
11
3.2.3 Term Spread Models
For y
m
t
= (r
m
t
z
m
t
)

, the short rate and spread for country m, the RS term spread model is:
y
m
t
= (s
m
t
) +A(s
m
t
)y
m
t1
+ (s
m
t
)
t
, (20)
where
t
N(0, I). We use 2 regimes, with constant transition probabilities, and also logistic
state-dependent transition probabilities where:
p(s
m
t
= j|s
m
t1
= j, y
m
t1
) =
exp(a
j
+ b
j
r
m
t1
+c
j
z
m
t1
)
1 + exp(a
j
+b
j
r
m
t1
+c
j
z
m
t1
)
j = 1, 2. (21)
We also estimate the term spread model jointly across the US, Germany and UK, following
Bekaert et al. (2001). This estimation views each country as an independent draw of the DGP,
by assuming independence of the regimes across countries, lack of cross-country correlation
and the same parameters across countries.
We consider two classes of one-regime models as potential benchmarks. First, we estimate
unconstrained VARs of the short rate and the term spread, restricting attention to rst and
second-order VARs. Second, we consider the afne class of term structure models (see Dufe
and Kan 1996). In these models, zero coupon yields are afne (constant plus linear term)
functions of the unobservable factors. This implies that we can represent y
m
t
(n), the yield for
maturity n for country m, as an afne function of the state variables X
m
t
for country m:
y
m
t
(n) =

A(n) +

B(n)

X
m
t
, (22)
where the scalar

A(n) and vector

B(n) are functions of the model parameters. We represent the
dynamics of X
m
t
, without loss of generality, by a rst-order VAR:
X
m
t
= + X
m
t1
+
m
t
, (23)
where
t
N(0, 1). The one-month yield (which we do not observe) takes the form:
y
m
t
(1) =
0
+

1
X
m
t
, (24)
where
0
is a scalar and
1
is a vector. The specication of a pricing kernel
m
t+1
, for each
country m, completes the model. The pricing kernel prices all nominal bonds through the
recursive relation:
P
m
t
(n + 1) = E
t
[
m
t+1
P
m
t+1
(n)], (25)
where P
m
t
(n) is the zero coupon bond price of maturity n for country m.
12
Different afne models make different assumptions about the state variable dynamics and
the specication of the pricing kernel, specically the specication of the prices of risk. Stan-
dard models assume either homoskedastic state variable dynamics with constant prices of risk,
for example correlated Vasicek (1977) models, or square-root process with time-varying prices
of risk or a combination of the two. Duffee (2001) demonstrates that standard afne term struc-
ture models forecast very poorly out of sample. Therefore, we consider an alternative afne
model not considered by Duffee. We consider Gaussian, homoskedastic state variables, but
time-varying prices of risk. More specically, we assume that the pricing kernel has the form:

t+1
= exp(
1
2

1
X
m
t

t+1
), (26)
where the risk premia
t
are time-varying:

t
=
0
+
1
X
m
t
, (27)
where
0
is a vector and
1
is a matrix.
For identication purposes, we impose the following parameter restrictions:
=
_
0
0
_
, =
_

11
0

12

22
_
, = I,
0
=
_

01
0
_
, and
1
=
_

11
0
0
22
_
. (28)
We call this bivariate correlated factor model the Gaussian Afne TermStructure Model (ATSM)
with time-varying risk premia.
The model has a structural VAR representation in terms of the observable yields. The short
rate and spread for country m can be written as:
y
m
t

_
r
m
t
z
m
t
_
=
_

A(3)

A(60)

A(3)
_
+
_

B(3)

B(60)

B(3)
_

X
m
t
(29)
or, by appropriately dening

Aand

B, as y
m
t
=

A+

BX
m
t
. The discrete-time recursive relations
determining

A(n) and

B(n) are derived in Ang and Piazzesi (2001). By substituting (23) into
(29), it is straightforward to show that:
y
m
t
= +Ay
m
t1
+
t
, (30)
where
t
N(0, I), = (I

B

B
1
)

A, A =

B

B
1
and = B. This representation makes
both maximum likelihood estimation and forecasting using the observed yields easy. Clearly,
the ATSM is simply a VAR model with cross-equation restrictions. Whereas the estimation of
this model went smoothly for the US, the likelihood surfaces for the UK and Germany proved
13
very at. Models with
1
restricted to 0, that is, standard correlated Vasicek (1977) models, do
not converge at all for all countries. Dai and Singleton (2001) show that a Gaussian model with
afne prices of risk matches the deviations from the Expectations Hypothesis observed for US
data, but they ignore small sample biases (see Bekaert, Hodrick and Marshall 1997).
3.3 Model Diagnostics
We start by reporting a number of standard in-sample residual tests for our various models. Our
second diagnostic more easily leads to comparisons across a large number of non-nested models
of varying complexity. We measure the t of the unconditional moments implied by the models
to the sample estimates of the unconditional moments. Single regime models may perform rea-
sonably well along these dimensions even though they are not the true DGP. However, they are
less likely to perform well over tests that exploit the changing behavior of interest rates across
regimes. To easily rank the performance across all models, we focus on summary statistics for
out-of-sample forecast errors. Finally, we compare different RS models, using a measure of the
quality of the regime classication. We discuss these in turn.
3.3.1 Residual Tests
We report two tests on in-sample scaled residuals e
m
t
of short rates of country m where e
m
t
=
(r
m
t
E
t1
[r
m
t
])/h
m
t1
. The conditional volatility h
m
t1
is given by:
(h
m
t1
)
2
= var
t1
(r
m
t
E
t1
[r
m
t
])
= E
t1
[(r
m
t
)
2
] (E
t1
[r
m
t
])
2
. (31)
For a univariate RS model E
t1
[r
m
t
] and E
t1
[(r
m
t
)
2
] are evaluated using equation (14).
Following Bekaert and Harvey (1997), we use a GMM test of the moment conditions on the
mean of the scaled residuals:
E[e
m
t
e
m
tj
] = 0 for j = 1, 2, . . . , k, (32)
which we refer to as mean residual tests, and a GMM test of the moments of the variance of
the scaled residuals:
E[((e
m
t
)
2
1) ((e
m
tj
)
2
1)] = 0 for j = 1, 2, . . . , k, (33)
which we refer to variance residual tests. In both tests we choose k = 6 and correct for
heteroskedasticity in the residuals following Andrews (1991).
14
3.3.2 Unconditional Moment Comparisons
We compute the unconditional population moments of our various models using analytical ex-
pressions where possible but use a simulation for the RS models for time-varying probabilities.
Analytical formulae for moments are available only for one-regime CIR and GARCH processes
as well as for autoregressive regime switching models with constant probabilities (see Timmer-
man 2000). Because of the high persistence of the series, we use sample sizes of one million.
To enable comparison across several models, we introduce the point statistic:
H = ( g g)

1
g
( g g), (34)
where g are sample estimates of unconditional moments, g are the unconditional moments from
the estimated model, and
g
is the covariance matrix of the sample estimates of the uncondi-
tional moments.
g
is estimated using a GMM estimation of the unconditional moments, and
for the purposes of this paper, we use a Newey-West (1987) estimate with 6 lags. The point
statistic assigns weights to the deviations between the unconditional moments implied by var-
ious models and the sample unconditional moments, which are inversely proportional to the
error by which the sample moments are estimated.
We test for the rst four central moments, the autocorrelogram and cross-correlations. In the
rst case g contains the mean, variance, skewness and kurtosis; for the autocorrelogram the rst
10 autocorrelations; and for cross-correlations lags from -3 to +3. Generally, the high persis-
tence of interest rates may lead to poor estimation of unconditional moments. Therefore, there
are instances where high correlation between the estimated moments leads to somewhat poorly
conditioned weighting matrices. Hence, we also calculate a related statistic H

, which uses
as a weighting matrix the diagonal of
g
. Strong correlations between the estimated moments
sometimes imply that the model minimizing H does not minimize H

.
3.3.3 Forecast Comparisons
Our forecast methodology is to estimate only using the in-sample period and forecast without
updating the parameters on the out-of-sample period. We use two point statistics for compar-
ison of unconditional forecast errors, the root mean squared error RMSE, and mean absolute
deviation MAD. For a time series
t
, these are dened as:
RMSE =
_
1
T

(
t

t
)
2
MAD =
1
T

|
t

t
|, (35)
15
where hatted values denote conditional forecast values. In our application we let
t
= r
t
for
univariate and multi-country models, looking at rst and second moments k = 1, 2. In term-
spread models we also consider
t
= z
t
and the cross-moment
t
= r
t
z
t
.
3.3.4 Regime Classication
Previous specication tests for RS models have focused on properties of residuals (Gray 1996)
or scores (Hamilton 1996), but here we propose a summary point statistic which captures the
quality of regime classication. A RS model assumes that at each point of time the data are
drawn from one of the regimes which is observed by agents in the economy but not by the
econometrician. To conduct inference about the regime, most papers focus on the smoothed
(ex-post) regime probabilities, p(s
t
= 1|I
T
) which we denote as p
t
. Weak regime inference
implies that the RS model cannot successfully distinguish between regimes from the behavior
of the data and may indicate misspecication. An ideal RS model should classify regimes
sharply so that p
t
is close to one or zero; in inferior models p
t
may hover close to a half.
To measure the quality of regime classication, we therefore propose the regime classica-
tion measure (RCM), dened for two states as:
RCM = 400
1
T
T

t=1
p
t
(1 p
t
). (36)
The constant serves to normalize the statistic to be between 0 and 100. Good regime classica-
tion is associated with low RCM statistic values: a value of 0 means perfect regime classica-
tion and a value of 100 implies that no information about the regimes is revealed. Since the true
regime is a Bernoulli random variable, the RCM statistic is essentially a sample estimate of its
variance.
The statistic easily generalizes to multiple regimes. A general denition of the statistic for
K regimes is:
RCM(K) = 100K
2
1
T
T

t=1
_
K

i=1
p
i,t
_
, (37)
where p
i,t
= p(s
t
= i|I
T
).
16
4 Empirical Results
4.1 Estimation Results of the Regime-Switching Models
Estimation of regime switching models in nite samples is plagued with the presence of mul-
tiple local maxima. To ensure that we identify the global maximum for the 31 RS models we
estimate, we use the following procedure. First, we obtain estimates for a large set of starting
values and select a candidate global maximum. Second, to check for stability of the global we
re-estimate using starting values randomly chosen in a 10% interval around the parameters
of the provisional global maximum. When models have trouble converging to a well-behaved
global using this procedure, we either dropped the model or simplied it, rather than continuing
the numerical search towards poorly identied models.
The RS models all produce one regime with a unit root and lower conditional volatility and
a second regime which is stationary with higher conditional volatility. This type of estimation
is found in univariate, multi-country and term spread models. Economically the rst regime
corresponds to normal periods where monetary policy smoothing makes interest rates behave
like a random walk. When extraordinary shocks occur, interest rates are driven up, volatility
becomes higher and interest rates become more mean-reverting.
In general, models with time-varying transition probabilities have many insignicant coef-
cients in the probability terms which suggests over-parameterization. Previous studies with
time-varying probabilities such as Gray (1996) also document this. For some of our models, we
fail to reject the null hypothesis of constant probabilities. Nevertheless, the general pattern that
emerges in the majority of cases is as expected: higher short rates (and spreads) increase the
probability of switching to the high volatility regime.
To highlight the features of specic models we discuss univariate, multi-country and term
spread RS models in turn. Recall that Table 3 presents the nomenclature scheme of the models.
4.1.1 Univariate Models
As Table 3 shows, we consider three different conditional volatility specications. We retain
constant transition probability models for all countries for all the formulations, except for the
UK GARCH model. We do not estimate state-dependent models for the GARCH formulation,
since the constant probability models are already over-parameterized. In estimating models
with state-dependent transition probabilites, we only nd signicant state dependence for the
US CIR model and the German RS AR(1) model. We drop the RS AR(1) model with state-
dependent transition probabilities for the US.
17
4.1.2 Multi-Country Models
RSD1 is a diagonal model with the same parameters (
i
,
i
,
i
) across countries and homoskedas-
tic within-regime errors. The RSG1-model is identical but has square root errors. Constraining

i
to be the same across countries imposes the restriction that the conditional volatility for Ger-
many and the UK is higher than the conditional volatility for the US. We relax this formulation
in the RSG2-model and nd that it makes little qualitative difference.
The estimation results show that Granger-causality by US shocks is important only for the
UK in the second mean-reverting high variance regime. Granger-causality of Germany is in-
signicant in both regimes. Looking at the impact of US shocks on the error terms of Germany
and the UK, the Granger-causality model RSG2 has signicant shock terms for Germany and
the UK in the rst random walk regime. The diagonal model, however, shows US shocks affect-
ing only UK shocks in the rst regime. To summarize, in the normal random walk regime US
shocks propagate into Germany and the UK, while in the second regime only the US Granger-
causes UK short rates.
4.1.3 Term Spread Models
In the RS term spread models we nd that Granger-causality is model dependent. For the US
and Germany, one regime produces a signicant A
i
[1, 2] term, so the spread Granger-causes the
short rate in only one regime (the higher variance regime for the US but the lower variance one
for Germany). The evidence for the UK is less clear as the coefcient is just insignicant in
one regime but very insignicant in the other. Similarly, the short rates Granger-cause spreads
only in one regime but these may not be the same regimes where spreads Granger-cause short
rates. In the US these are in opposite regimes, but for Germany these regimes are the same. In
the joint estimation where we assume independence and the same parameters across countries,
short rates and spreads Granger-cause each other in the same regime (the lower conditional
variance regime).
The correlation between short rates and spreads differs markedly across regimes. The high
variance less persistent regime has more negative correlation than the low variance regime.
Wald tests for equality across the regimes reject with zero p-value for all countries. Short rates
and spreads seem less correlated in the rst regime, which corresponds to normal periods.
However, note from Table 2 that the correlation between the short rate and spread is more
negative in expansions, which is the opposite to what the regime switching models imply. Nev-
ertheless, the high mean, high variance second regime does correspond to economic recessions.
We examine this further in Section 6.
18
For our time-varying probability formulations the transition probabilities depend on both
the short rate and spread, except for the US where we use a model with transition probabili-
ties dependent only on the spread. Likelihood ratio tests for constant transition probabilities
versus time-varying probabilities reject for all countries. The results on Granger-causality and
regime-dependent correlations hold for both the constant and time-varying transition probability
models.
5 Performance Measures
Section 5.1 analyzes residual tests and the moment performance and Section 5.2 analyzes fore-
cast performance. Section 5.3 summarizes the evidence and makes use of a Monte Carlo exper-
iment to help interpret the results. The results are reported in Tables 5 through 10.
5.1 In-Sample Tests
5.1.1 Residual Tests
Table 5 lists the results of the mean and variance residual tests. Turning rst to the US results,
the benchmark single-regime models perform well, passing both the mean and variance residual
tests. However, in each of the univariate, multi-country and term spread models, the variance
residual test has a p-value of only slightly larger than 5%. The only models which comfortably
pass both the mean and variance residual tests incorporate term spread information in a RS
model (RSM1 and RSM2). The single-regime or RS (RS3) univariate GARCH models and the
CIR models fail to pass the mean residual tests. The multi-country RS models generally do
poorer than their single-regime counterparts.
The mean residual tests for Germany reject all the models, with the exception of a second
order VAR, despite a rst order VAR being the optimal AIC and BIC choice. Several RS
specications (RS2 and RSD2) do less well than their single-regime counterparts, with the
variance residual test also rejecting them. In comparison, almost all the models pass the residual
tests on UK data, with univariate RS state-dependent transition probability specications (RS2
and RS5) and the ATSM being the exception.
The Gaussian ATSMs reject the mean residual test at a 5% level across all countries. The
implied factors from afne models are severely biased, which leads to the poor in-sample per-
formance, but the ATSMs manage to pass the variance residual tests. This conrms evidence
in Ghysels and Ng (1998) who reject the conditional mean specication of afne models, but
19
also nd less evidence of mis-specication with second moments.
In summary, no single model passes all the residual tests for all countries. For the US and
UK, RS term spread models comfortably pass the residual tests, while almost all models fail to
pass the residual tests on German data.
5.1.2 Matching Sample Moments
We present H-statistics for univariate models in Panel A of Table 6. For the US, the one-regime
models seem to work better in matching unconditional moments than the RS models. The
dismal performance of models RS1-3 for the US is partly caused by the unit root in one of the
regimes, although the models are theoretically stationary. For Germany, RS2 and RS3 do poorly
because they produce large values for kurtosis. The best ts for the moments for Germany are
for the one-regime and RS CIR models. For the UK, the AR(1) RS processes work best with the
square root processes performing more poorly. RS models with state-dependent probabilities
(RS2, RS5) and GARCH errors (RS3) fare less well than the constant probability models, RS1
and RS4.
Panel B of Table 6 reports H-statistics for the multi-country models. Among the one-
regime models, diagonal models match central moments better than the unconstrained VAR(1)
or Granger-causality models, suggesting over-parameterization in these models. With the ex-
ception of the UK, the RS diagonal model performs better than its one-regime diagonal coun-
terpart. This is quite an achievement considering that this model constrains each country to
have the same parameters. The RS Granger-causality models perform more poorly than the RS
diagonal models for the US and UK but not for Germany. There is little difference when we no
longer constrain
i
to be equal across countries in the RS Granger-causality models.
Table 7 reports the H and H* statistics for the bivariate short rate-spread system. The one-
regime models (VAR(1), VAR(2) and ATSM) generally out-perform the RS models (RSM1 and
RSM2) at matching unconditional moments. For one-regime models, the more parsimonious
VAR(1) denitely does better at matching autocorrelations than VAR(2), with comparable re-
sults for the central moments. For the US, the ATSM performs almost as well as VAR(1) and
VAR(2) in matching central moments, but this is not the case for the UK and Germany. In
matching autocorrelations, the ATSM performs best across the board in Germany, performs
best for the short rate aucorrelations in the UK, and also performs best for spread correlations
in the US. However, the ATSMs perform extremely poorly in all countries matching the short
rate-spread cross-correlation. This is because the off-diagonal term in the companion matrix
of the factors (
12
in (28)) is near zero for Germany and the UK. Turning to the RS models,
20
the state-dependent probability models fare better for the US and Germany than their constant
probability counterparts, but for the UK this result is reversed. One-regime models clearly out-
perform RS VARs for central moments and autocorrelograms. Only for cross-correlations does
RSM2 provide good ts.
Does incorporating extra information improve the performance of RS models? By looking
across Panels A and B of Table 6 we compare the univariate RS models with the multi-country
RS models. We see a dramatic improvement when incorporating multi-country information for
the US but not for Germany or the UK. Comparing the univariate RS models in Table 6 with the
bivariate RS term spread models in Table 7, the term spread information leads to a better match
of moments only for the US, and for autocorrelations only for the UK. Overall, using the extra
information from other countries or the term spread unequivocally helps the US obtain a better
t to unconditional moments, but it denitely does not help for Germany. The evidence for the
UK is mixed.
5.2 Out-of-Sample Tests
Tables 8 and 9 list the forecast performance results. Focusing rst on univariate models in Panel
A of Table 8, the state-dependence of the probabilities in RS AR(1) models produces superior
forecasts, even though many of the estimated coefcients are insignicant and the performance
in matching the sample moments is poor. However, this result is not shared by the RS CIR
model, with only the UKs state dependent formulation performing better. Overall, with the
exception of the UK, the GARCH models produce the best results. For the UK, the superior
performance of the RS2 model, using either the RMSE or MAD criterion for both rst and
second moments, is remarkable given that regime classication in the UK is rather poor (see
Figure 1 discussed below). Relative to their one-regime counterparts, RS models generally
perform better. For all countries with the exception of the one-regime GARCH model, the RS
AR(1) models forecast better than a simple AR(1) and the RS CIR models forecast better than
the single regime CIR models.
Panel B of Table 8 presents the forecasting results for the multi-country models. The diag-
onal one-regime models out-perform the unrestricted VAR on mean forecasts and do worse for
second moment forecasts only for the US, again showing over-parameterization of the uncon-
strained VAR. The multi-country RS diagonal model outperforms the one-regime model despite
having the interest rate DGP constrained to be the same across all countries. This is a strong
endorsement of the importance of regime shifts in forecasting. Granger-causality seems to aid
in forecasting both in one-regime and RS frameworks. The RS Granger models do particularly
21
well for the US and the UK.
Table 9 reports forecast performance in the term spread models. In forecasting the rst and
second moments, the more parsimonious VAR(1) outperforms the VAR(2) for all countries,
suggesting that the VAR(2) is over-parameterized. In the US, the ATSM provides better fore-
casts of the short rate than unrestricted VARs, which conrms the results in Ang and Piazzesi
(2001). This nding is repeated for Germany but not for the UK, where the ATSM fails to beat
the VAR specications. For the US, the ATSM out-performs all the other bivariate specica-
tions for forecasting short rates and second moments of short rates. Duffee (2001) comments
that afne models with constant risk premia forecast very poorly, but he does not consider fore-
casts of afne models with time-varying risk premia as in our ATSM specication. In contrast
to the US results, in Germany and the UK RS models out-perform the one-regime models for
forecasting the level and square of short rates. The results of forecasts of spreads and cross-
moments are mixed. While the RS models out-perform the one-regime specications in the US,
the ATSM and VAR specications provide better forecasts in Germany and the UK. The lowest
RMSE-statistics for cross-moment forecasts belong to the RS models for the US and Germany,
the best cross-moment forecast for the UK is VAR(1).
Adding information from other countries or term spreads to the estimation uniformly im-
proves forecasts. Focusing on the RMSE criterion, Table 8 shows that the multi-country ap-
proach generally yields better forecasts than the univariate models. Table 9 shows that adding
term spreads improves forecasts, with the RS spread models beating univariate forecasts with
the exception of the US, where the ATSM dominates.
5.3 Summary and Interpretation
In general, we nd that in matching sample moments RS models do not systematically outper-
form one-regime models. However, in forecasting out of sample, RS models almost invariably
do better. Focusing on short rates, Table 10 reports the best models with the lowest H, RMSE
and MAD statistics. There is no clear-cut best model. However, it appears that while single
regime models may produce lower H-statistics (for example in the case of the US), RS models
forecast better for all countries. We note that for the US, the ATSM comes very close to giv-
ing the best forecast for the short rate. Moreover, the best RS forecasting models incorporate
information from other countries or the spread. Interestingly, RS models with state-dependent
probabilities tend to forecast better than their constant probability counterparts even though they
perform very poorly at matching sample moments.
How do we interpret these results? As indicated before, the RS models considered here need
22
large simulations to pin down their unconditional moments with any precision. This means that
the small sample behavior of RS models may be poor. In other words, it is conceivable that
more parsimonious one-regime models produce better estimates of the sample unconditional
moments than RS models in small samples, even though a RS model is the true DGP. Here we
run a Monte Carlo experiment to specically investigate this conjecture.
Consider the following RS VAR population model of the short rate and spread, y
t
= (r
t
z
t
)

:
y
t
= (s
t
) +A(s
t
)y
t1
+
1
2
(s
t
)
t
where
t
N(0, I), s
t
= 1, 2 with Markov state-dependent
logistic transition probabilities depending on lagged y
t
. We use the parameters from the joint
estimation as the population model and nd the true population moments of this model using
a very long simulation. Then, we simulate a small sample of size T + N and compute uncon-
ditional moment estimates over the in-sample of size T and RMSE forecast statistics over the
out-sample of size N for several aproximations to the true model. We set T and N to be the size
of our in-sample and out-sample data sets in this paper, 267 and 30 respectively. The models we
consider are an AR(1) and a RS AR(1) on the short rates with constant probabilities, a VAR(1)
and a RS VAR(1) on the bivariate short rate and spread with constant transition probabilities.
We denote these as AR, RS AR, VAR, and RS VAR respectively.
Unfortunately we cannot include the true model because of the problems we encounter in
nding satisfactory estimates of the RS VAR with time-varying probabilities in small samples.
The many convergence failures that occur even when starting from the true parameters are in
itself proof of the small sample problems RS models face.
To compare the unconditional moment estimators, we calculate H-statistics with the mean,
standard deviation, skewness and kurtosis, and then record which of the four models yields the
best (lowest) statistic value for each simulated sample. To compare out-of-sample forecasts, we
record which model produces the lowest RMSE statistic. We use 1000 Monte Carlo replications.
Table 11 reports the percentage times each model best t the population moments or produced
the best forecasts. For example, for the simulations performed, in 15.9% of cases the AR(1)
model gave the best t to the population moments as measured by the H-statistic even though
the true model was a RS VAR(1) with state-dependent probabilities.
Table 11 shows that the one-regime models are good approximations to the true RS models
in small samples, and that despite the true DGP being a RS model, parsimonious one-regime
models may perform better at matching moments and forecasting. It is notable that RS models
perform quite poorly in matching unconditional moments, but perform better in forecasting.
These results parallel our ndings for the actual RS models estimated on real data.
We also examine the empirical distribution of the moments produced by the models in small
samples. Table 12 reports the population values of the unconditional moments for the short
23
rates and spreads. The table also lists the mean values and standard deviations of the small
sample distribution of the moments produced by the various models. RS models tend to over-
estimate the mean and under-estimate the variance of the short rate, but the population val-
ues lie within 95% condence intervals of the small sample model moments. However, the
AR and VAR single-regime models produce close to unbiased estimates of the mean and vari-
ance. This result may help justify the popularity of VAR-type models to test unconditional term
structure hypotheses, such as the Expectations Hypothesis, even in the presence of signicant
non-linearities in the data.
6 Regime Classication and Regime Interpretation
Figure 1 displays the regime probabilities for the RS VAR state-dependent transition probability
model for the US, Germany and UK. The solid line in the top plots are smoothed probabilities
p(s
t
= 1|I
T
) using information over the full sample of size T and the broken line represents
ex-ante probabilities p(s
t
= 1|I
t1
). Plots of ex-ante and smoothed probabilities for the other
models look similar. For the UK, there is a high frequency of switching between regimes
because the transition probabilities P and Q are very close to a half. In a regime switching
model, if P + Q = 1 the model reduces to a simple switching model. For the UK models, we
often cannot reject this hypothesis and the regime classication also appears poor because the
smoothed regime probability often is far away from 1 or 0.
For a more quantitative examination of regime classication, we present RCM statistics in
Table 13. In univariate models the RS AR(1) model produces the sharpest regime classica-
tion for the US, while RS CIR models produce the sharpest regime classication for Germany
and the UK. For univariate models, moving from constant to state-dependent transition prob-
abilities produces very little improvement. Our multi-country model produces sharper regime
classication for the UK and Germany at the expense of the US. In particular, there is a large
improvement in regime classication for the UK by adding US information. Including term
spread information leads to lower RCM statistics for all countries.
Are the regimes correlated with the business cycle? Table 14 attempts to answer this ques-
tion. The table rst presents correlations between various lags j of the ex-ante probabilities
p
tj+1
and a recession indicator for the business cycles of each country. The ex-ante proba-
bilities are generated from the term spread RS model with time-varying probabilities (RSM2).
We use this model because it is the model with the lowest RCM statistic for the US in Table
13. We report the correlations between the second regime with mean-reverting higher volatility
24
and the economic downturns. The table shows that this regime is associated with economic
recessions, while the normal unit root regime with lower volatility represents economic ex-
pansions. The US and Germany have signicant correlations, while the correlations of the UK
are insignicant.
The business cycle association of the regimes is not surprising for the US. Figure 1 shows
that the ex-ante probabilities during the 1979-1982 period of monetary targeting are near zero,
placing this period in the second regime. During this period high variable interest rates were
accompanied by a large recession. Germany also experienced a similar episode around the
same time (1980:03 to 1983:07), and also went through an earlier recession accompanied by
high interest rates in the early 1970s (1973:09 to 1975:05). The recession brought on by the re-
unication, beginning in mid-1991, also saw rising interest rates but the regimes do not capture
this period as successfully. The poor results for the UK are not surprising given the poor regime
classication of the UK model.
The last four columns of Table 14 report coefcients from a Probit regression with the
recession indicator being the dependent variable, and current and lagged ex-ante probabilities
being the independent variables. The Probit regressions yield signicant coefcients for the US
and Germany. We also list the percentage of correctly forecasted recessions in-sample from the
Probit regressions. For the US, the ex-ante probabilities successfully predict 84% of recessions
one-month ahead, with the success ratio slightly increasing as we try to predict further into the
future. The success ratio is around 60% for Germany and, not surprisingly, only 50% for the
UK.
Harvey (1988) and Estrella and Mishkin (1997) nd that term spreads successfully predict
real economic activity. Table 14 conrms their ndings showing that the magnitude of corre-
lations between recessions and the spread increases with the lag, and that the accuracy of the
Probit forecasts increases with the forecast horizon. This happens in all three countries. Look-
ing specically at the US, the ex-ante regime probabilities have better forecast ratios for one and
two month ahead predictions than the spread. While the forecast ratios increase with horizon
for the spread, the forecast ratios of the the ex-ante probabilities remain essentially at. This
evidence indicates that for the US the ex-ante regime probabilities are better contemporaneous
indicators of the business cycle than the spread, and the spread is a forward looking indicator
with greater forecasting ability at longer horizons. For the other countries, the spread better
predicts recessions than our regime probabilities at all horizons. Given that both the regime
classication and the dating of the actual business cycles is less precise for these countries, this
is not surprising.
25
7 Conclusions
We compare the econometric performance of regime-switching (RS) models relative to their
one-regime counterparts in several ways. First, residual tests show that RS models often per-
form worse than single regime models. However, for the US only RS models with term spread
information comfortably pass the residual tests. Second, the moments implied by RS models
do not always t the sample moments as well as simpler models do because of the difculties in
estimating RS models in small samples. A Monte Carlo experiment conrms that this happens
even when the RS model is the true data generating process. Finally, RS models invariably fore-
cast better than one-regime models, although a parsimonious multi-factor afne term structure
model with time-varying prices of risk performs almost as well for US short rates.
To improve the econometric performance of RS models it is important to incorporate addi-
tional information. In fact, univariate RS models yield inconsistent estimates when the omitted
variables contain information on the regime. We compare the performance of univariate with
multi-country short rate models and models incorporating term spreads. In particular, US short
rates improve both the regime classication and the statistical performance for German and UK
short rates (but not vice versa). Furthermore, inclusion of term spread information leads to gen-
eral improvements over univariate models in forecasting and to dramatic superior performance
in regime inference. The inclusion of additional cross-sectional country short rates or term
spreads does not always improve the t of the unconditional moments. However, the regimes
correspond well with business cycle expansions and contractions.
26
References
Ahn, D. H., and Gao, B. (2000), A Parametric Nonlinear Model of Term Structure Dynamics, Review of
Financial Studies, 12, 4, 721-762.
At-Sahalia, Y. (1996), Testing Continuous-Time Models of the Spot Interest Rate, Review of Financial Studies,
9, 2, 385-426.
Ang, A., and Bekaert, G. (1998), Regime Switches in Interest Rates, Working Paper 6508, National Bureau of
Economic Research.
Ang, A., and Bekaert, G. (2000), Short Rate Nonlinearities and Regime Switches, Journal of Economic
Dynamics and Control, in press.
Ang, A., and Piazzesi, M. (2001), A No-Arbitrage Vector Autoregression of Term Structure Dynamics with
Macroeconomic and Latent Factors, Working Paper 8363, National Bureau of Economic Research.
Andrews, D. (1991), Heteroskedasticity and Autocorrelation Consistent Matrix Estimation, Econometrica, 59,
817-858.
Bansal, R., and Zhou, H. (1999), Term Structure of Interest Rates with Regime Shifts, mimeo, Duke University.
Bekaert, G., and Harvey, C. R. (1997), Emerging Equity Market Volatility, Journal of Financial Economics, 43,
1, 29-77.
Bekaert, G., Hodrick, R. J., and Marshall, D. A. (1997), On Biases in Tests of the Expectations Hypothesis of the
Term Structure of Interest Rates, Journal of Financial Economics, 44, 3, 309-348.
Bekaert, G., Hodrick, R. J., and Marshall, D. A. (2001), Peso Problem Explanations for Term Structure
Anomalies, forthcoming Journal of Monetary Economics.
Cecchetti, S. J., Lam, P., and Mark, N. C. (1993), The Equity Premium and the Risk-Free Rate: Matching the
Moments, Journal of Monetary Economics, 31, 21-46.
Cox, J., Ingersoll, J., and Ross, S. (1985), A Theory of the Term Structure of Interest Rates, Econometrica, 53,
385-408.
Dai, Q., and Singleton, K. (2001), Expectation Puzzles, Time-varying Risk Premia, and Afne Models of the
Term Structure, Journal of Financial Economics, in press.
Duffee, G. R. (2001), Forecasting Future Interest Rates: Are Afne Models Failures? Journal of Finance, in
press.
Dufe, D., and Kan, R. (1996), A Yield-Factor Model of Interest Rates, Mathematical Finance, 6, 4, 379-406.
Estrella, A., and Mishkin, F. S. (1997), The Predictive Power of the Term Structure of Interest Rates in Europe
and the United States: Implications for the European Central Bank, European Economic Review, 41, 7,
1375-1401.
Evans, C., and Marshall, D. (2000), Economic Determinants of the Term Structure of Nominal Interest Rates,
mimeo, Chicago Federal Reserve.
Evans, M. D. (1998), Regime Shifts, Risk and the Term Structure, mimeo, Georgetown University.
Evans, M, D., and Lewis, K. K. (1995), Do Expected Shifts in Ination Affect Estimates of the Long-Run Fisher
Relation, Journal of Finance, 50, 1, 225-253.
27
Garcia, R. (1998), Aymptotic Null Distribution of the Likelihood Ratio Test in Markov Switching Models,
International Economic Review, 39, 3, 763-788.
Garcia, R. and Perron, P. (1996), An Analysis of the Real Interest Rate under Regime Shifts, The Review of
Economics and Statistics, 78, 1, 111-125.
Ghysels, E, and Ng, S. (1998), A Semiparametric Factor Model of Interest Rates and Tests of the Afne Term
Structure, Review of Economics and Statistics, 80, 4, 535-48.
Gray, S. F. (1996), Modeling the Conditional Distribution of Interest Rates as a Regime-Switching Process,
Journal of Financial Economics, 42, 27-62.
Hamilton, J. D. (1988), Rational-Expectations Econometric Analysis of Changes in Regime: An Investigation of
the Term Structure of Interest Rates, Journal of Economic Dynamics and Control, 12, 385-423.
Hamilton, J. D. (1989), A New Approach to the Economic Analysis of Nonstationary Time Series and the
Business Cycle, Econometrica, 57, 357-384.
Hamilton J. D. (1994), Time Series Analysis, Princeton University Press, New Jersey.
Hamilton, J. D. (1996), Specication Testing in Markov-Switching Time Series Models, Journal of
Econometrics, 70, 127-157.
Harvey, C. R. (1988), The Real Term Structure and Consumption Growth, Journal of Financial Economics, 22,
2, 305-33.
Holst, U., Lindgren, G., Holst, J., and Thuvesholmen, M. (1994), Recursive Estimation in Switching
Autoregressions with a Markov Regime, Journal of Time Series Analysis, 15, 5, 489- 506.
Jorion, P., and Mishkin, F. (1991), A Multicountry Comparison of Term-Structure Forecasts at Long Horizons,
Journal of Financial Economics, 29, 59-80.
Lewis, K. K. (1991), Was There a Peso Problem in the US Term Structure of Interest Rates: 1979-1982?
International Economic Review, 32, 159-173.
Mankiw, N. G., and Miron, J. A. (1986), The Changing Behavior of the Term Structure of Interest Rates,
Quarterly Journal of Economics, 101, 211-228.
Naik, V., and Lee, M. H. (1994), The Yield Curve and Bond Option Prices with Discrete Shifts in Economic
Regimes, mimeo, University of British Columbia.
Newey, W. K., and West, K. D. (1987), A Simple Positive Semi-Denite, Heteroskedasticity and Autocorrelation
Consistent Covariance Matrix, Econometrica, 55, 703-8.
Pfann, G. A., Schotman, P. C., and Tschernig, R. (1996), Nonlinear Interest Rate Dynamics and Implications for
the Term Structure, Journal of Econometrics, 74, 149-176.
Sola, M., and Drifll, J. (1994), Testing the Term Structure of Interest Rates Using a Vector Autoregression with
Regime Switching, Journal of Economic Dynamics and Control, 18, 601-628.
Stanton, R. (1997), A Nonparametric Model of Term Structure Dynamics and the Market Price of Interest Rate
Risk, Journal of Finance, 52, 5, 1973-2002.
Vasicek, O. (1977), An Equilibrium Characterization of the Term Structure, Journal of Financial Economics, 5,
177-188.
Timmerman, A. (2000), Moments of Markov Switching Models, Journal of Econometrics, 96, 75-111.
Zarnowitz, V. (1997), Some New Lessons from International Cyclical Indicators, in Oppenl ander, K., ed.,
Business Cycle Indicators, Ashgate Publishing, 175-204.
28
Table 1: Sample Moments
Panel A: Sample Central Moments
Parameter US GER UK
short rate spread short rate spread short rate spread
mean 7.3381 1.2198 6.9045 0.4984 10.5605 0.0643
(0.4449) (0.2028) (0.4197) (0.2719) (0.4268) (0.2491)
variance 8.3103 2.0366 7.1111 3.1241 8.2388 2.7458
(1.9390) (0.3833) (1.3380) (0.6714) (1.4354) (0.5292)
skewness 0.8172 -0.7281 0.6806 -0.5410 -0.1521 -0.2596
(0.2167) (0.2782) (0.2515) (0.3227) (0.1797) (0.2404)
kurtosis 3.6102 3.5921 2.6987 3.3732 2.5406 2.8086
(0.6718) (0.7179) (0.4405) (0.5768) (0.3264) (0.4071)
Panel B: Sample Autocorrelations
Lag US GER UK
short rate spread short rate spread short rate spread
1 0.9706 0.8669 0.9845 0.9657 0.9565 0.9322
(0.0181) (0.0292) (0.0216) (0.0265) (0.0237) (0.0238)
2 0.9295 0.7663 0.9583 0.9207 0.8948 0.8776
(0.0347) (0.0497) (0.0436) (0.0507) (0.0450) (0.0425)
3 0.8931 0.6958 0.9253 0.8715 0.8271 0.8234
(0.0513) (0.0689) (0.0638) (0.0711) (0.0637) (0.0596)
4 0.8551 0.6221 0.8858 0.8127 0.7627 0.7692
(0.0653) (0.0820) (0.0812) (0.0868) (0.0784) (0.0753)
5 0.8256 0.5873 0.8428 0.7502 0.7006 0.7200
(0.0778) (0.0836) (0.0957) (0.0999) (0.0895) (0.0895)
Panel C: Sample Cross Correlations
Short rates of countries Short rates/Spreads
Lag US/DEM US/UK DEM/UK US GER UK
-3 0.4197 0.6470 0.3279 -0.3655 -0.7929 -0.6524
(0.1334) (0.0777) (0.1007) (0.1130) (0.0563) (0.0727)
-2 0.4205 0.6549 0.3523 -0.4213 -0.8326 -0.7016
(0.1322) (0.0725) (0.0964) (0.1091) (0.0435) (0.0607)
-1 0.4120 0.6521 0.3696 -0.4907 -0.8656 -0.7375
(0.1315) (0.0686) (0.0939) (0.1038) (0.0317) (0.0521)
0 0.3953 0.6454 0.3808 -0.5920 -0.8804 -0.7637
(0.1310) (0.0678) (0.0933) (0.0976) (0.0284) (0.0459)
1 0.3756 0.6139 0.3782 -0.5952 -0.8634 -0.7057
(0.1325) (0.0698) (0.0945) (0.0982) (0.0335) (0.0539)
2 0.3542 0.5758 0.3717 -0.5715 -0.8389 -0.6608
(0.1335) (0.0754) (0.0974) (0.1013) (0.0406) (0.0629)
3 0.3294 0.5485 0.3650 -0.5522 -0.8097 -0.6210
(0.1328) (0.0828) (0.1008) (0.1080) (0.0477) (0.0718)
NOTE: Sample period 1972:01 to 1993:02 (in-sample period). Standard errors are in paren-
theses and are estimated using GMM with 6 Newey-West (1987) lags. In Panel C, the
cross-correlations are the estimates of
cov(r
m
1
t+j
,r
m
2
t
)

var(r
m
1
t
)

var(r
m
2
t
)
for j = 3, 2, . . . , +2, +3
and country m
1
and country m
2
.
29
T
a
b
l
e
2
:
I
n
t
e
r
e
s
t
R
a
t
e
B
e
h
a
v
i
o
r
o
v
e
r
t
h
e
B
u
s
i
n
e
s
s
C
y
c
l
e
U
S
G
e
r
m
a
n
y
U
K
r
e
c
e
s
s
i
o
n
e
x
p
a
n
s
i
o
n

2
p
-
v
a
l
r
e
c
e
s
s
i
o
n
e
x
p
a
n
s
i
o
n

2
p
-
v
a
l
r
e
c
e
s
s
i
o
n
e
x
p
a
n
s
i
o
n

2
p
-
v
a
l
N
u
m
b
e
r
o
f
o
b
s
e
r
v
a
t
i
o
n
s
5
0
2
4
7
1
4
9
1
4
8
1
2
8
1
6
9
m
e
a
n
9
.
6
4
6
6
6
.
5
9
7
0
0
.
0
0
0
1
7
.
4
3
1
9
5
.
8
3
2
7
0
.
0
0
4
3
1
1
.
9
6
9
5
8
.
6
8
5
6
0
.
0
0
0
0
(
0
.
7
0
6
4
)
(
0
.
3
1
1
8
)
(
0
.
4
7
0
5
)
(
0
.
3
0
2
8
)
(
0
.
3
4
7
8
)
(
0
.
4
1
7
8
)
v
a
r
i
a
n
c
e
8
.
4
5
1
8
6
.
3
1
7
3
0
.
3
5
8
1
8
.
9
2
3
7
4
.
0
2
5
3
0
.
0
0
5
5
4
.
7
0
4
9
8
.
1
8
3
5
0
.
0
1
7
7
(
1
.
8
7
7
5
)
(
1
.
3
6
7
7
)
(
1
.
3
8
9
8
)
(
1
.
0
8
9
4
)
(
0
.
8
3
9
6
)
(
1
.
2
0
2
6
)
s
h
o
r
t
s
k
e
w
n
e
s
s
0
.
6
8
4
1
1
.
0
3
6
0
0
.
5
1
9
9
0
.
2
7
8
2
1
.
3
3
1
8
0
.
0
2
9
8
0
.
3
1
5
1
0
.
4
1
8
6
0
.
7
5
5
1
r
a
t
e
(
0
.
4
7
4
5
)
(
0
.
2
7
2
1
)
(
0
.
2
4
4
8
)
(
0
.
4
1
8
5
)
(
0
.
2
5
5
9
)
(
0
.
2
1
1
3
)
r
k
u
r
t
o
s
i
s
2
.
0
0
7
7
4
.
5
4
9
9
0
.
0
1
5
9
2
.
0
5
9
0
5
.
1
6
4
1
0
.
0
4
0
8
2
.
5
6
2
7
2
.
1
2
4
8
0
.
3
9
0
6
(
0
.
7
1
9
8
)
(
0
.
7
7
0
8
)
(
0
.
2
1
1
1
)
(
1
.
5
0
3
6
)
(
0
.
4
2
2
7
)
(
0
.
2
8
5
4
)

1
0
.
7
8
5
8
0
.
9
5
0
3
0
.
0
7
5
0
0
.
9
4
3
6
0
.
8
8
9
4
0
.
2
7
7
1
0
.
8
6
5
0
0
.
9
2
4
3
0
.
2
9
8
1
(
0
.
0
9
0
2
)
(
0
.
0
2
0
1
)
(
0
.
0
3
8
3
)
(
0
.
0
3
1
9
)
(
0
.
0
5
0
4
)
(
0
.
0
2
6
6
)

2
0
.
5
6
5
7
0
.
9
0
6
1
0
.
0
2
7
6
0
.
8
8
7
8
0
.
7
6
4
1
0
.
1
7
5
4
0
.
7
6
7
4
0
.
8
3
5
3
0
.
4
5
0
4
(
0
.
1
5
0
1
)
(
0
.
0
3
6
8
)
(
0
.
0
6
7
7
)
(
0
.
0
6
1
3
)
(
0
.
0
7
4
6
)
(
0
.
0
5
0
3
)
m
e
a
n
0
.
5
5
6
8
1
.
3
8
3
5
0
.
0
4
7
4
0
.
0
2
4
7
1
.
2
4
4
3
0
.
0
0
0
4
-
0
.
4
8
9
6
0
.
8
0
2
5
0
.
0
0
0
1
(
0
.
3
9
0
3
)
(
0
.
1
4
6
9
)
(
0
.
2
6
8
8
)
(
0
.
2
1
7
9
)
(
0
.
2
1
2
8
)
(
0
.
2
5
1
7
)
v
a
r
i
a
n
c
e
3
.
1
7
2
4
1
.
5
3
6
2
0
.
1
2
0
6
3
.
0
6
5
7
2
.
1
7
1
9
0
.
2
8
6
6
1
.
6
8
9
1
2
.
9
7
3
2
0
.
0
8
3
2
(
1
.
0
1
7
0
)
(
0
.
2
7
7
6
)
(
0
.
5
9
5
6
)
(
0
.
5
9
0
6
)
(
0
.
2
8
4
3
)
(
0
.
6
8
4
5
)
s
k
e
w
n
e
s
s
0
.
2
9
2
8
-
1
.
0
5
0
7
0
.
0
0
3
8
-
0
.
5
5
8
4
-
0
.
8
0
9
2
0
.
6
4
3
6
-
0
.
3
9
0
3
-
0
.
8
3
5
0
0
.
1
6
3
1
(
0
.
4
0
9
0
)
(
0
.
2
1
8
8
)
(
0
.
2
9
9
6
)
(
0
.
4
5
1
7
)
(
0
.
2
2
0
7
)
(
0
.
2
3
0
0
)
s
p
r
e
a
d
k
u
r
t
o
s
i
s
3
.
4
9
0
0
3
.
9
6
9
9
0
.
6
4
8
7
2
.
8
9
9
5
4
.
9
1
1
1
0
.
1
2
1
0
2
.
2
5
8
3
3
.
5
9
2
2
0
.
1
1
9
2
z
(
0
.
7
7
4
2
)
(
0
.
7
1
4
4
)
(
0
.
4
9
6
3
)
(
1
.
1
9
8
6
)
(
0
.
3
8
0
2
)
(
0
.
7
6
7
2
)

1
0
.
6
4
6
1
0
.
8
6
3
0
0
.
0
3
6
2
0
.
9
0
1
0
0
.
8
5
9
0
0
.
5
7
5
1
0
.
8
4
8
7
0
.
9
1
2
3
0
.
2
8
8
9
(
0
.
0
9
8
7
)
(
0
.
0
3
1
3
)
(
0
.
0
5
8
8
)
(
0
.
0
4
6
5
)
(
0
.
0
5
2
9
)
(
0
.
0
2
8
4
)

2
0
.
3
0
0
0
0
.
7
7
0
0
0
.
0
0
0
7
0
.
8
0
8
9
0
.
6
9
1
8
0
.
3
6
9
8
0
.
7
6
4
6
0
.
8
2
0
0
0
.
5
2
7
8
(
0
.
1
3
0
8
)
(
0
.
0
4
3
5
)
(
0
.
0
9
5
2
)
(
0
.
0
8
9
5
)
(
0
.
0
7
0
7
)
(
0
.
0
5
1
7
)

(
r
,
z
)
-
0
.
2
5
8
0
-
0
.
6
9
4
7
0
.
0
4
1
3
-
0
.
8
8
1
3
-
0
.
8
5
9
1
0
.
6
6
3
7
-
0
.
6
2
7
2
-
0
.
7
9
4
8
0
.
0
8
6
3
(
0
.
1
7
0
3
)
(
0
.
0
8
8
6
)
(
0
.
0
3
0
0
)
(
0
.
0
4
1
4
)
(
0
.
0
8
7
6
)
(
0
.
0
4
3
3
)
N
O
T
E
:
S
a
m
p
l
e
p
e
r
i
o
d
:
1
9
7
2
:
0
1
t
o
1
9
9
6
:
0
8
(
f
u
l
l
s
a
m
p
l
e
)
.
R
e
c
e
s
s
i
o
n
s
a
r
e
d
e

n
e
d
t
o
b
e
f
r
o
m
t
h
e
p
e
a
k
t
o
t
h
e
t
r
o
u
g
h
o
f
t
h
e
b
u
s
i
n
e
s
s
c
y
c
l
e
.
S
t
a
n
d
a
r
d
e
r
r
o
r
s
a
r
e
i
n
p
a
r
n
e
t
h
e
s
i
s
a
n
d
a
r
e
c
o
m
p
u
t
e
d
u
s
i
n
g
G
M
M
w
i
t
h
3
N
e
w
e
y
-
W
e
s
t
l
a
g
s
.

i
d
e
n
o
t
e
s
t
h
e
i
t
h
a
u
t
o
c
o
r
r
e
l
a
t
i
o
n
,

(
r
,
z
)
d
e
n
o
t
e
s
t
h
e
c
o
r
r
e
l
a
t
i
o
n
b
e
t
w
e
e
n
s
h
o
r
t
r
a
t
e
s
a
n
d
s
p
r
e
a
d
s
,
a
n
d

2
p
-
v
a
l
d
e
n
o
t
e
s
t
h
e
p
-
v
a
l
u
e
f
r
o
m
a

2
t
e
s
t
o
f
e
q
u
a
l
i
t
y
a
s
s
u
m
i
n
g
i
n
d
e
p
e
n
d
e
n
c
e
a
c
r
o
s
s
c
y
c
l
e
p
e
r
i
o
d
s
.
30
Table 3: Summary of Models Estimated
Univariate Models of short rates
One-regime Two-regime equivalents
const probs time-dep probs
AR(1) RS1 RS2
(3) (8) (10)
GARCH(1,1) RS3
(5) (12)
CIR RS4 RS5
(3) (8) (10)
Multi-Country Models of short rates
Model Description
VAR1u unconstrained VAR(1)
(18)
G1 one-regime Granger-causality model, homoskedastic errors
(13)
RSG1 RS Granger-causality with the same
i
,
i
,
i
, P, Q across countries,
(16) square root errors
RSG2 RS Granger-causality with the same
i
,
i
, P, Q across countries,
(20) but different
i
, square root errors
D1 one-regime diagonal model, homoskedastic errors
(11)
RSD1 RS diagonal model with the same
i
,
i
,
i
, P, Q across countries,
(12) homoskedastic errors
Multivariate Models of the Term Spread
One-regime Two-regime equivalents
const probs time-dep probs
VAR(1) RSM1 RSM2
(9) (20) (24)
VAR(2)
(13)
ATSM (Afne Term Structure Model)
(9)
31
Table 4: Granger Tests in the Multi-Country VAR Model
Granger-causality A[i, j] = 0 p-value
no country Granger-causes another all off-diagonal elements = 0 0.0528
US Granger-causes Germany and UK A[2, 1] = A[3, 1] = 0 0.0029
Germany and UK Granger-cause US A[1, 2] = A[1, 3] = 0 0.7332
Germany and UK Granger-cause each other A[2, 3] = A[3, 2] = 0 0.6742
NOTE: Wald tests are performed using GMMwith 6 Newey-West lags. The notation A[i, j]
refers to the element in row i, column j.
Table 5: Residual Tests on Short Rates
US GER UK
mean var mean var mean var
Univariate Models
AR1 0.3170 0.0523 0.0212* 0.4826 0.4855 0.8401
RS1 0.0101* 0.4980 0.0003** 0.2782 0.4868 0.6243
RS2 - 0.0001** 0.0094** 0.0000** 0.0248*
GARCH 0.0085** 0.1529 0.0000** 0.5841 0.4543 0.8894
RS3 0.0153* 0.5452 0.0011** 0.4612 -
CIR 0.0071** 0.0000** 0.0088** 0.1321 0.4410 0.7602
RS4 0.0039** 0.5368 0.0000** 0.1489 0.4519 0.8656
RS5 0.0030** 0.6816 0.0000** 0.0592 0.0000** 0.2700
Multi-Country Models
VAR1u 0.3728 0.0508 0.0137* 0.2463 0.5197 0.6841
G1 0.8767 0.0000** 0.0246* 0.1189 0.7495 0.6441
RSG1 0.0483* 0.1081 0.0002** 0.5369 0.7702 0.5769
RSG2 0.0311* 0.1478 0.0002** 0.5476 0.7753 0.6668
D1 0.2497 0.0513 0.0211* 0.4820 0.4885 0.8407
RSD2 0.0016** 0.0000** 0.0000** 0.0000** 0.6210 0.4704
Term Spread Models
VAR1 0.5087 0.0547 0.0197* 0.4432 0.4598 0.4517
VAR2 0.8715 0.0185* 0.3146 0.1572 0.4568 0.4146
ATSM 0.0120* 0.0911 0.0233* 0.4902 0.0329* 0.4132
RSM1 0.3831 0.1753 0.0011** 0.3581 0.5929 0.8920
RSM2 0.2988 0.1434 0.0013** 0.3682 0.4338 0.9694
32
Table 6: Moments of Univariate and Multi-Country Models
Panel A: Univariate Models
RS1 RS2 RS3 RS4 RS5 AR(1) GARCH CIR
US
Central moments H 330.99 - 327.56 113.72 63.03 30.32 - 3.11*
H* 112.15 - 72.92 36.78 46.24 15.26 - 1.71*
Autocorrelogram H 10.01 - 6.67 8.82 5.23 3.88* - 3.91
H* 20.84 - 7.81 16.31 6.19 1.30* - 5.34
GER
Central moments H 67.03 4563.76 5211.55 100.78 17.03* 165.53 - 34.11
H* 17.80 153.80 4088.45 27.78 9.61 7.98 - 6.54*
Autocorrelogram H 6.82 9.21 5.08* 6.14 7.67 6.91 - 5.96
H* 13.01 22.30 3.07* 12.46 20.59 13.74 - 9.59
UK
Central moments H 3.49* 4.00 - 29.02 36.90 5.81 6.82 25.11
H* 4.38 4.19 - 18.34 34.27 2.83* 4.09 19.13
Autocorrelogram H 7.43 7.75 - 7.98 7.36* 9.51 8.84 8.43
H* 11.06* 13.07 - 14.01 14.61 20.24 17.55 15.58
Panel B: Multi-Country Models
VAR1u D1 RSD1 G1 RSG1 RSG2
US
Central moments H 30.83 21.73 13.38* 30.31 28.66 32.66
H* 15.25 15.76 11.10* 15.26 17.26 22.64
Autocorrelogram H 3.43 3.34* 8.06 3.87 9.77 11.70
H* 0.97 0.25* 13.44 1.29 19.84 27.34
GER
Central moments H 174.98 166.62 54.91 207.78 26.09* 26.47
H* 7.90* 7.98 15.50 8.20 10.09 11.09
Autocorrelogram H 6.12* 6.91 7.19 6.99 7.96 9.12
H* 12.43* 13.82 13.21 15.10 16.55 21.84
UK
Central moments H 6.40 6.06* 64.62 7.94 146.54 287.66
H* 2.76 2.80 55.29 2.71* 81.77 123.56
Autocorrelogram H 10.08 9.63* 26.13 11.67 31.97 34.04
H* 21.90 20.69* 81.71 26.34 106.81 113.68
33
Table 7: Unconditional Moments of Term Spread Models
VAR1 VAR2 ATSM RSM1 RSM2
US
Central Moments r
t
H 31.51 29.99 29.62* 193.20 141.95
H* 15.26* 15.27 15.35 97.05 54.83
z
t
H 10.09 10.07* 10.88 119.70 30.96
H* 7.62* 7.62 7.87 86.30 21.77
Autocorrelations r
t
H 2.46* 890.48 52.25 4.32 5.13
H* 0.84* 17.35 249.45 1.33 12.11
z
t
H 21.70 5724.77 5.30* 16.82 10.58
H* 43.26 68.67 5.67* 69.27 14.52
Crosscorrelation r
t
z
t
H 86.99 444.51 73.71 12.73 2.05*
H* 16.24 8.82 260.11 31.29 0.19*
GER
Central Moments r
t
H 232.01 157.55* 250.97 374.27 268.22
H* 8.24 8.08* 10.42 20.11 11.40
z
t
H 6.43 6.03* 17.11 38.98 18.69
H* 3.29* 3.69 7.49 10.04 5.97
Autocorrelations r
t
H 7.41 2941.41 4.85* 6.65 6.04
H* 15.28 23.09 3.24* 13.90 10.63
z
t
H 8.50 316.92 5.70* 15.57 14.66
H* 17.39 34.19 8.68* 51.67 47.65
Crosscorrelation r
t
z
t
H 6.96* 142.92 1718.14 17.30 10.80
H* 8.89 7.71 4228.46 10.91 4.21*
UK
Central Moments r
t
H 4.84 4.93 4.33* 23.51 32.49
H* 3.03 3.00* 3.10 4.33 5.60
z
t
H 2.25 2.17* 77.39 9.80 11.09
H* 1.42 1.40* 46.42 7.63 9.15
Autocorrelations r
t
H 8.04 50.26 7.24* 8.69 8.98
H* 16.26 60.28 13.31* 19.00 21.69
z
t
H 2.82* 119.41 8.61 2.99 3.09
H* 0.38* 21.27 16.61 2.00 2.42
Crosscorrelation r
t
z
t
H 7.87* 199.73 397.72 17.00 11.36
H* 11.44 10.93 1233.65 9.34 2.09*
34
Table 8: Forecasts of Univariate and Multi-Country Models
Panel A: Univariate Models
RS1 RS2 RS3 RS4 RS5 AR(1) GARCH CIR
US
r
t
MAD 0.1488 - 0.1458 0.1458 0.1487 0.1483 0.1387* 0.1664
RMSE 0.1956 - 0.1943 0.1945 0.1968 0.1888* 0.1999 0.1999
r
2
t
MAD 1.5161 - 1.4696 1.4771 1.5048 1.6540 1.3410* 1.9874
RMSE 1.9421 - 1.9167* 1.9277 1.9525 2.0335 1.9207 2.3042
GER
r
t
MAD 0.1307 0.1299 0.1329 0.1285 0.1327 0.1501 0.1207* 0.1615
RMSE 0.1732 0.1732 0.1716 0.1694 0.1732 0.1900 0.1627* 0.2006
r
2
t
MAD 1.2097 1.1979 1.1822 1.1407 1.1824 1.4568 1.0895* 1.4985
RMSE 1.5936 1.5943 1.5351 1.5114 1.5423 1.8174 1.4736* 1.8637
UK
r
t
MAD 0.2509 0.2137* - 0.2449 0.2288 0.2419 0.2555 0.2539
RMSE 0.2890 0.2668* - 0.2819 0.2771 0.2772 0.2910 0.2893
r
2
t
MAD 3.5109 2.9807* - 3.2783 3.0626 3.3666 3.4550 3.3090
RMSE 4.0030 3.5617* - 3.7319 3.6206 3.8180 3.9192 3.7569
Panel B: Multi-Country Models
VAR1u D1 RSD1 G1 RSG1 RSG2
US
r
t
MAD 0.1619 0.1499 0.1378 0.1483 0.1160* 0.1174
RMSE 0.2002 0.1891 0.1841 0.1888 0.1625* 0.1626
r
2
t
MAD 1.5550 1.7159 1.2139 1.3992 0.9949* 1.0388
RMSE 1.8065 2.0771 1.4980 1.6453 1.1930* 1.2146
GER
r
t
MAD 0.1580 0.1500 0.1429 0.1327* 0.1451 0.1466
RMSE 0.1959 0.1899 0.1868 0.1704* 0.2035 0.2062
r
2
t
MAD 1.6591 1.4557 1.2822 1.4632 1.1957 1.2436*
RMSE 1.8537 1.8164 1.5706 1.6303 1.5206* 1.5899
UK
r
t
MAD 0.2747 0.2410 0.1429 0.2668 0.1124* 0.1142
RMSE 0.3116 0.2762 0.2017 0.3055 0.1766* 0.1833
r
2
t
MAD 2.2897 3.3541 1.6389 2.1274 1.2960 1.2872*
RMSE 2.0020 3.8040 2.1859 1.8801 1.8465* 1.8554
35
T
a
b
l
e
9
:
F
o
r
e
c
a
s
t
s
o
f
T
e
r
m
S
p
r
e
a
d
M
o
d
e
l
s
V
A
R
1
V
A
R
2
A
T
S
M
R
S
M
1
R
S
M
2
U
S
r
t
z
t
r
t
z
t
r
t
z
t
r
t
z
t
r
t
z
t
M
A
D
0
.
1
8
8
5
0
.
2
1
1
7
0
.
1
9
1
8
0
.
2
1
8
6
0
.
1
2
2
4
*
0
.
2
1
3
4
0
.
1
5
3
1
0
.
2
0
7
0
*
0
.
1
5
8
8
0
.
2
0
7
2
R
M
S
E
0
.
2
3
1
2
0
.
2
7
6
0
0
.
2
4
9
0
0
.
2
7
3
5
0
.
1
6
4
1
*
0
.
2
8
2
5
0
.
1
9
4
8
0
.
2
6
5
3
*
0
.
2
0
2
5
0
.
2
6
5
0
r
2 t
z
2 t
r
t
z
t
r
2 t
z
2 t
r
t
z
t
r
2 t
z
2 t
r
t
z
t
r
2 t
z
2 t
r
t
z
t
r
2 t
z
2 t
r
t
z
t
M
A
D
2
.
1
1
8
3
0
.
7
1
4
1
1
.
2
6
7
2
2
.
1
2
6
7
0
.
7
4
2
2
1
.
2
1
2
2
1
.
2
0
6
5
*
0
.
8
5
0
1
1
.
0
9
7
0
*
1
.
5
9
0
7
0
.
6
2
6
7
*
1
.
1
0
4
4
1
.
6
5
2
9
0
.
6
2
8
0
1
.
0
9
6
4
R
M
S
E
2
.
5
2
2
6
0
.
9
1
8
1
1
.
5
3
0
9
2
.
6
5
2
1
0
.
9
5
3
4
1
.
4
8
3
5
1
.
5
0
0
5
*
1
.
0
2
9
2
1
.
4
8
1
6
1
.
9
9
4
1
0
.
8
2
9
2
*
1
.
1
3
9
1
*
2
.
0
7
5
5
0
.
8
2
8
7
1
.
3
8
7
0
G
E
R
r
t
z
t
r
t
z
t
r
t
z
t
r
t
z
t
r
t
z
t
M
A
D
0
.
1
3
5
9
0
.
2
1
9
1
0
.
1
7
9
6
0
.
2
3
9
5
0
.
1
2
1
2
0
.
2
1
7
6
*
0
.
1
1
9
7
0
.
2
2
1
4
0
.
1
0
7
4
*
0
.
2
1
8
6
R
M
S
E
0
.
1
7
6
5
0
.
2
7
9
6
0
.
2
2
1
9
0
.
2
8
2
6
0
.
1
6
3
2
0
.
2
7
3
4
*
0
.
1
6
1
7
0
.
2
8
2
1
0
.
1
4
7
1
*
0
.
2
7
5
5
r
2 t
z
2 t
r
t
z
t
r
2 t
z
2 t
r
t
z
t
r
2 t
z
2 t
r
t
z
t
r
2 t
z
2 t
r
t
z
t
r
2 t
z
2 t
r
t
z
t
M
A
D
1
.
3
4
6
4
0
.
7
0
8
4
*
0
.
9
8
8
2
1
.
6
5
6
8
0
.
8
0
1
6
1
.
0
0
4
9
1
.
2
1
2
1
0
.
7
5
0
7
0
.
8
9
9
9
*
1
.
1
3
3
3
0
.
7
1
6
8
1
.
0
2
4
3
0
.
9
7
7
2
*
0
.
7
2
2
5
1
.
0
0
9
2
R
M
S
E
1
.
7
1
0
9
0
.
8
7
5
7
*
1
.
3
2
7
3
2
.
1
2
7
8
0
.
9
2
4
1
1
.
2
9
5
7
1
.
5
8
0
5
0
.
8
9
9
9
1
.
2
9
9
2
1
.
5
4
9
5
0
.
8
9
0
9
1
.
3
3
4
6
1
.
3
5
3
1
*
0
.
8
9
6
2
1
.
2
9
3
7
*
U
K
r
t
z
t
r
t
z
t
r
t
z
t
r
t
z
t
r
t
z
t
M
A
D
0
.
2
1
7
2
0
.
2
4
5
5
*
0
.
2
6
0
7
0
.
2
6
0
3
0
.
4
5
1
0
0
.
3
2
7
4
0
.
2
2
1
1
0
.
2
4
9
1
0
.
1
7
6
8
*
0
.
2
5
9
0
R
M
S
E
0
.
2
5
2
9
0
.
3
0
3
1
0
.
3
1
1
2
0
.
3
3
8
4
0
.
4
8
6
7
0
.
4
3
0
8
0
.
2
1
8
0
*
0
.
3
0
7
8
0
.
2
4
1
4
0
.
3
1
2
1
r
2 t
z
2 t
r
t
z
t
r
2 t
z
2 t
r
t
z
t
r
2 t
z
2 t
r
t
z
t
r
2 t
z
2 t
r
t
z
t
r
2 t
z
2 t
r
t
z
t
M
A
D
3
.
0
7
2
9
1
.
0
9
8
9
*
1
.
2
9
9
5
*
3
.
5
8
7
0
1
.
1
9
2
1
1
.
3
2
9
7
6
.
2
2
2
2
1
.
2
1
2
8
1
.
5
1
9
4
3
.
1
0
1
6
1
.
1
1
6
8
1
.
3
1
9
7
2
.
4
4
7
1
*
1
.
1
3
7
0
1
.
3
8
6
4
R
M
S
E
3
.
4
9
8
1
1
.
4
9
9
1
*
1
.
5
2
9
2
*
4
.
1
0
7
4
1
.
6
2
9
0
1
.
6
0
9
7
6
.
6
1
6
6
1
.
9
0
7
8
1
.
9
8
0
9
3
.
5
8
1
4
1
.
5
3
4
3
1
.
5
4
1
6
3
.
1
7
8
6
*
1
.
5
4
2
3
1
.
5
7
4
6
36
Table 10: Overall Moments and Forecast Comparisons for short rates
Best H-statistics
US GER UK
Central moments CIR RS5 RS1
Autocorrelogram VAR1 ATSM ATSM
Best RMSE-statistics Best MAD-statistics
US GER UK US GER UK
r
t
RSG1 RSM2 RSG1 RSG1 RSM2 RSG1
r
2
t
RSG1 RSM2 RSG1 RSG1 RSM2 RSG2
Table 11: Small Sample Experiment: % Time Models do Best
Unconditional Moments Forecasts
AR RS AR VAR RS VAR AR RS AR VAR RS VAR
r
t
central 15.9% 59.9% 14.8% 9.4% r
t
30.6% 16.3% 24.5% 28.6%
(r
t
) 43.4% 3.3% 43.7% 9.6% r
2
t
29.4% 18.0% 20.5% 32.1%
z
t
central 90.1% 9.9% z
t
45.8% 54.2%
(z
t
) 36.3% 63.7% z
2
t
46.1% 53.9%
(r
t
, z
t
) 88.9% 11.1% cross 44.2% 55.8%
NOTE: We simulate data of length 297 from the joint estimation across the US-Germany-UK of a bivariate
system of the short rate r
t
and spread z
t
with time-varying probabilities (Model RSM2). We then estimate
an AR(1), a regime-switching AR(1), a VAR, and a regime-switching VAR, denoted AR, RS AR, VAR and
RS VAR respectively and record which model gives the lowest H and RMSE statistics. The table lists the
percentage times of which model performed the best in each small sample. We conduct 1000 simulations.
37
Table 12: Small Sample Distribution of Moments
Short Rates
Population AR RSAR VAR RSVAR
Mean 7.3289 7.3905 8.5011 7.4066 8.8526
(1.3454) (1.4462) (1.3802) (1.7742)
Variance 11.2885 10.9206 7.8944 11.0027 8.9975
(3.8646) (2.2026) (4.3127) (2.6317)
Skewness 0.5750 0.2032 0.1185
(0.1700) (0.3087)
Kurtosis 3.0639 3.1360 3.2287
(0.3263) (3.3094)
Spreads
Population AR RSAR VAR RSVAR
Mean 0.8642 0.8509 0.3410
(0.3903) (0.4304)
Variance 1.5460 1.4306 1.0500
(0.5161) (0.2705)
Skewness -0.1815 -0.0790
(0.2812)
Kurtosis 3.0084 3.2709
(1.8155)
NOTE: These are the means, with standard errors in parentheses, of the moments of the estimated models in
a small sample of 267, in the experiment of Table (11). Skewness and kurtosis for the AR and VAR models
are theoretically 0 and 3 respectively.
Table 13: RCM Statistics
US GER UK
RS1 10.44 22.57 43.14
RS2 - 23.69 41.54
RS3 19.04 52.53 -
RS4 11.53 21.02 42.29
RS5 12.88 20.45 40.64
RSD1 18.11 13.44 27.23
RSG1 21.16 19.72 28.17
RSG2 21.94 22.25 24.48
RSM1 7.67 14.60 38.70
RSM2 6.68 16.12 34.90
38
Table 14: Markov Regimes and Business Cycles
US
Correlations Probit Forecasting
mths ahead j (1 p
tj+1
, rec
t
) (z
tj
, rec
t
) (1 p
tj+1
) %forecast (z
tj
) %forecast
1 0.4264 -0.3047 1.6203 83.8 -0.2811 80.8
(0.1153) (0.1104) (0.2569) (0.0605)
2 0.4618 -0.3989 1.7537 84.2 -0.3847 82.3
(0.1149) (0.1028) (0.2603) (0.0645)
4 0.4840 -0.5096 1.8428 84.4 -0.5611 86.7
(0.1123) (0.0851) (0.2640) (0.0760)
6 0.4122 -0.5296 1.5569 85.1 -0.5750 87.0
(0.1126) (0.0820) (0.2584) (0.0745)
Germany
Correlations Probit Forecasting
mths ahead j (1 p
tj+1
, rec
t
) (z
tj
, rec
t
) (1 p
tj+1
) %forecast (z
tj
) %forecast
1 0.1892 -0.5276 0.5789 60.2 -0.4903 75.2
(0.1109) (0.0719) (0.1879) (0.0601)
2 0.2162 -0.5830 0.6632 61.5 -0.6073 75.8
(0.1107) (0.0615) (0.1890) (0.0696)
4 0.2472 -0.6590 0.7615 63.9 -0.8474 77.9
(0.1101) (0.0508) (0.1908) (0.0927)
6 0.2392 -0.6811 0.7366 63.6 -0.9400 81.6
(0.1106) (0.0483) (0.1915) (0.1024)
UK
Correlations Probit Forecasting
mths ahead j (1 p
tj+1
, rec
t
) (z
tj
, rec
t
) (1 p
tj+1
) %forecast (z
tj
) %forecast
1 0.0911 -0.3439 0.6856 54.1 -0.2821 67.3
(0.1066) (0.0999) (0.4590) (0.0506)
2 0.0779 -0.3828 0.5864 53.6 -0.3218 69.4
(0.1067) (0.0962) (0.4601) (0.0522)
4 0.0098 -0.4508 0.0740 51.3 -0.4018 74.1
(0.1077) (0.0899) (0.4646) (0.0564)
6 -0.0230 -0.4680 -0.1756 49.0 -0.4274 72.4
(0.1063) (0.0837) (0.4710) (0.0584)
NOTE: Recessions are coded as a 1, expansions as 0. The symbol p
t
represents the ex-ante probabilities
p(s
t
= 1|I
t1
) of the rst regime from the term spread RS model with time-varying transition probabilities
(RSM2). The rst two columns give the correlation of the recession indicator (rec) with the ex-ante proba-
bility of the second regime and the spread z
t
. Standard errors are calculated using GMM with 3 Newey-West
lags. The last four columns show results from tting the Probit model p(rec
t
= 1) = F(+(.)a
tj
), where
F(.) is the normal cumulative distribution function, is the coefcient corresponding to the variable a
tj
,
and we let a
tj
be current and lagged values of 1 p
tj
and z
tj1
. Lags are in months. The %forecast
column is the percentage of correctly forecasted (in-sample) values from the Probit regression.
39
75 80 85 90 95
0
0.2
0.4
0.6
0.8
1
1st Regime Probability
P
r
o
b
a
b
i
l
i
t
y
Year
Short Rate
Spread
US
75 80 85 90 95
5
0
5
10
15
20
I
n
t
e
r
e
s
t

r
a
t
e
Year
75 80 85 90 95
0
0.2
0.4
0.6
0.8
1
1st Regime Probability
P
r
o
b
a
b
i
l
i
t
y
Year
Short Rate
Spread
GER
75 80 85 90 95
5
0
5
10
15
20
I
n
t
e
r
e
s
t

r
a
t
e
Year
75 80 85 90 95
0
0.2
0.4
0.6
0.8
1
1st Regime Probability
P
r
o
b
a
b
i
l
i
t
y
Year
UK
75 80 85 90 95
5
0
5
10
15
20
I
n
t
e
r
e
s
t

r
a
t
e
Year
Short Rate
Spread
Figure 1: Regime Probabilities
40
Figure Legend
Figure 1 shows the ex-ante probabilities p(s
t
= 1|I
t1
) (dotted line) and smoothed probabilities p(s
t
= 1|I
T
)
(solid line) in the top subplots for each country, and the short rate and spread for each country in the bottom
subplots.
41