Yield Curve As A Predictor of Recessions: Evidence From Panel
Data
Huseyin Ozturk
Luis Felipe V. N. Pereira
Abstract
In this study we test empirically whether the slope of the yield curve - yield spread - is a
good predictor of recessions. We follow Estrella and Trubin (2006) although instead of
time series, we adopt an unbalanced panel data framework for 32 OECD countries from
1990 to 2011. This modification allows one to apply this model for countries with short
time series. Furthermore, we include four quarter lagged GDP in the model to assure that
yield spread is a good predictor of recessions, even when controlling for GDP changes. Our
models are estimated using panel Probit, Logit and MLE. The results show that with a Type
I error of 25%, the models deliver a power of roughly 63% and can be used as an effective
instrument to predict recessions one year ahead.
Keywords: term structure, panel probit, panel logit, recession
Resumo
Este estudo testa a capacidade da inclinação da curva de rendimentos – “spread” de juros –
para prever recessões. A metodologia segue o proposto por Estrella e Trubin (2006), mas
substitui as séries de tempo por dados em painel não balanceado, incluindo 32 países da
OCDE de 1990 a 2011. Essa modificação permite aplicar o modelo proposto a países com
séries de tempo curtas. Ainda, incluímos defasagens do PIB para garantir que o “spread” de
juros é útil na previsão de recessões, mesmo quando se controla para os efeitos de
mudanças no PIB. Os modelos são estimados utilizando painel Probit, Logit e MLE. Os
resultados indicam que com um nível de erros Tipo-1 de 25%, os modelos possuem poder
de aproximadamente 63% e confirmam que esta metodologia pode ser utilizada como um
instrumento efetivo para prever recessões com um ano de antecedência.
Palavras-chave: estrutura a termo, painel probit, painel logit, recessão
JEL Classification: C23, C24, C25
Área 7 – Microeconomia, Métodos Quantitativos e Finanças

Undersecretariat of Treasury of the Republic of Turkey and Hacettepe University. Email:
huseyin.ozturk@hazine.gov.tr

Brazil National Treasury Secretariat and Catholic University of Brasilia. Email:
luis.fpereira@catolica.edu.br.
The views and opinions expressed in this work are those of the authors and do not necessarily reflect those of
the Brazil National Treasury or the Undersecretariat of Treasury of the Republic of Turkey.
1. Introduction
It has been widely discussed in recent literature that expectations about short-term interest
rates, which determine a substantial part of the movement of long-term interest rates,
depend upon macroeconomic variables. Therefore, one would expect macroeconomic
variables and modeling exercises to be quite informative in explaining and forecasting the
yield-curve movements. Likewise, yield curve movements would be informative in
explaining and forecasting future economic performance.
A strong rationality supports the relationship between the slope of the yield curve and real
economic activity. There are at least three main reasons that explain the relationship and
thus explain why the yield curve might contain information about future recessions. In
general, this relationship is positive and, essentially, reflects the expectations of financial
market participants regarding future economic growth. A positive spread between longterm and short-term interest rates, i.e. a steepening yield curve, is associated with an
expectation of an increase in real economy activity, while a negative spread, i.e. a flattening
or inverted yield curve, is associated with an expectation of a decline in real activity.
The first reason stems from the expectations hypothesis of the term structure of interest
rates. This hypothesis states that long-term interest rates reflect the expected path of future
short-term interest rates. Thus, long-term rates can be considered a weighted average of
expected future short-term rates. In particular, it claims that, for any choice of holding
period, investors do not expect to get different returns from holding bonds of different
maturity dates (Anderson et. al. 1996). An anticipation of a recession implies the
expectation of a decline of future interest rates that is reflected in a decrease of long-term
interest rates. These expected reductions in interest rates might be induced from countercyclical monetary policy designed to stimulate the economy. Yet, the monetary policy and
yield curve dynamics is out of the scope of this study.
Another reason that explains the aforementioned relationship is related to the effects of
monetary policy on yield curve. For example, when monetary policy is tightened, shortterm interest rates rise; long-term rates also typically rise but usually by less than the
current short rate, leading to a downward-sloping term structure. The monetary contraction
can eventually reduce spending in sensitive sectors of the economy, causing economic
growth to slow and, thus, the probability of a recession to increase. Estrella and Mishkin
(1998) show that the monetary policy is an important determinant of term structure spread
in this sense.
The third reason is given by Harvey (1988) and Hu (1993) and is based on the
maximization of the inter-temporal consumer choices. The central assumption behind their
model is that consumers prefer a stable level of income rather than very high income during
expansion and very low income during slowdowns. Therefore, if the consumers expect a
reduction of their income, in case of a recession, they prefer to save and buy long-term
bonds in order to get payoffs in the slowdown. By doing that they increase the demand for
long-term bonds and that leads to a decrease of the corresponding yield. Further, to finance
the purchase of the long-term bonds, a consumer may sell short-term bonds whose yields
2
will increase. As a result, when a recession is expected, the yield curve flattens or gets
inverted.
Indeed, yield spread between long – term and short – term interest rates may contain
significant information about future economic activity, in particular, recessions. Based on
this background, this paper focuses on the study of the information content of the yield
curve in order to predict recessions in OECD. Based on the vast literature, our study revisits
the issue and extends the validity of the methodology to these countries that have short time
series data. As per the stylized fact suggests, we test for the information content of the yield
curves. The problem faced is that some countries have short time series. Econometrics
analysis traditionally makes use of long time series. We overcome this problem by
employing panel data. By doing so, we will be able to investigate this relationship for the
countries who have short time series as well. As for the best of our knowledge, this study is
the first to employ panel data to investigate this issue.
First, by using panel data we adopt standard probit, logit and MLE methods to predict
recessions. Second, we include in our a model the GDP lagged four quarters, Our results
show that both models in all three methods deliver roughly the same level of power.
Regarding the the logit and probit methods, both are expected to have same estimation
results1. The MLE estimates are also roughly the same. Lastly, we compare predicted
recessions with actual recessions by defining Type I and Type II errors and evaluating the
power of each model.
The paper is organized as follows: the next section will have a brief literature survey.
Section 3 explains the data and methodology and focus on methodological procedures to
make accurate estimations. Section 4 will present our econometric results for probit, logit
and MLE. The robustness of the results is also tested in this section. This section will
discuss the results for the estimation result for those countries that have short time series.
Section 5 has some concluding remarks.
2. Literature
Several economics and finance papers have explored the macroeconomic determinants of
the unobservable factors of the yield curve identified by empirical finance studies. Wu
(2001) examines the relationship between the Federal Reserve's monetary policy
"surprises" and the movement of the "slope" factor of the yield curve in the U.S. after 1982.
His study identifies monetary policy "surprises" in several ways to make the analysis more
robust; the results indicate a strong correlation between such monetary policy "surprises"
and the movement of the "slope" factor over time.
Ang and Piazzesi (2001) examine the influences of inflation and real economic activity on
the yield curve in an asset-pricing framework. In their model, bond yields are determined
1
Logit and probit models produce comparable results. In both estimation methodologies, error terms are
distributed logistically and normally. These distributions have very much common except for their tails being
one of them is fatter. Our sample is not so large to catch the differentiation in tails. Therefore, logit and probit
model estimation produce similar results.
3
not only by the three unobservable factors—level, slope, and curvature—but also by an
inflation measure and a real activity measure. They find that incorporating inflation and real
activity into the model is useful in forecasting the yield curve's movement. However, such
effects are quite limited. Inflation and real activity help explain the movements of shortterm bond yields and medium-term bond yields (up to a maturity of one year), but most
movements of long-term bond yields are still accounted for by the unobservable factors.
Therefore, they conclude that macroeconomic variables cannot substantially shift the level
of the yield curve.
Evans and Marshall (2001) analyze the same problem using a different approach. They
formulate several models with rich macroeconomic dynamics and look at how the "level",
"slope," and "curvature" factors are affected by the structural shocks identified in those
models. Their conclusion confirms Ang and Piazzesi's (2001) result that a substantial
portion of short- and medium-term bond yields is driven by macroeconomic variables.
However, they also find that in the long run macroeconomic variables do indeed explain
much of the movement of the long-term bond yields, and the "level" factor responds
strongly to macroeconomic variables. For instance, their identification results indicate that
the changes in households' consumption preferences induce large, persistent, and
significant shifts in the level of the yield curve.
Ahrens (2002), analyzes the concept with another question. He studies the informational
content of the term structure as a predictor of recessions in eight OECD countries. The
results of the study suggest that for all countries in the analysis the term spread turns out to
be a successful predictor for recession.
Karunaratne (2002) also studies the relationship between recession and yield curve. After
testing for the unit root, stationary variables reveals out that the yield curve gives the best
forecast on economic activity. Non-nested tests of other financial indicators, that are
nominees to be used for predicting recession, demonstrate that yield curve is the
outperformer in predicting recession. In his analysis four quarter time horizon gives the best
results in predicting recession for Australia.
Afonso and Martins (2010) study fiscal behavior and the sovereign yield curve in the U.S.
and Germany during the period 1981:I-2009:IV. The latent factors, level, slope and
curvature, obtained with the Kalman filter approach, are used in a VAR with macro and
fiscal variables, controlling for financial stress conditions.
Last but not the least, De Pace (2012) describes of the leading properties of the term
structure for output growth relationship using use time-varying-parameter models and realtime data to shed light on the dynamic characteristics of the yield curve. He investigates
five European economies and USA over the last decades and until the third quarter of 2010.
Interestingly, the main argument of the paper is that the predictive content of the term
spread is not a reliable predictor of output growth over time and across the country set. He
also contend that the yield spread significantly contributes to the forecast performance of
simple growth regressions in Europe but not in the USA in recent years and the variance of
the random shocks to the term spreads tends to fall in all countries.
4
The increasing interest to predict recession from the term structure is evident. Since Estrella
and Mishkin (1998), many studies have been conducted that investigate power of long –
term and short – term yield spread as a recession predictor. Estrella and Trubin (2006) have
also argued that the slope of the yield curve – in a bold definition, the spread between long
and short – term interest rates – is a good predictor of future economic activity. The focus
of their study was the slope of the curve could be a good forecasting tool in real time,
although it was instrumental in predicting past recessions. We follow Estrella and Mishkin
(1998) and Estrella and Trubin (2006) in modeling recession. By employing panel data,
information content of yield curve as a recession predictor has been tested for these
countries, even if they have short time series data. We employ binary response models and
MLE in estimation process.
3. Data and Methodology
Empirically, we build a model that converts the steepness of the yield curve at the present
time into a likelihood of a recession four quarter ahead. The reason why we make
predictions four quarter ahead is that Estrella and Mishkin (1998) makes forecasts one, two,
four and six quarters ahead. The basic finding they arrived is that the performance of the
yield curve spread improves considerably as the forecast horizon lengthens to two and four
quarters. By the same token, Estrella and Trubin (2006), predicts recessions four quarter
ahead.
All in all, our analysis is composed of four items: a measure of steepness of the yield curve,
a definition of recession, two models that combine both of them and reliable econometric
techniques for estimation.
3.1. A measure of steepness
For the sake of simplicity we define the steepness of the yield curve as the difference
between long – term and short – term interest rate (Estrella and Trubin 2006, Estrella and
Mishkin 1998). Although the very recent methods to estimate the term structure of the yield
curve give thorough approximation of the slope, we follow the literature on predicting
recession over the yield curve which defines the steepness of the yield curve as the
difference between long – term and short – term interest rate.
Then, long – term and short – term interest rate need to be identified. We utilize OECD
methodology in identifying long – term and short – term interest rate. According to OECD
definition, the short term interest rate are usually either the three month interbank offer rate
attaching to loans given and taken amongst banks for any excess or shortage of liquidity
over several months or the rate associated with Treasury bills, Certificates of Deposit or
comparable instruments, each of three month maturity. For Euro Area countries the 3month "European Interbank Offered Rate" is used from the date the country joined the
euro.
Whereas, long term (in most cases 10 year) government bonds are the instruments whose
field is used as the representative ‘interest rate’ for this area. Generally the yield is
calculated at the pre-tax level and before deductions for brokerage costs and commissions
5
and is derived from the relationship between the present market value of the bond and that
at maturity, taking into account also interest payments paid through to maturity.
3.2. Definition of recession
There have been many attempts to define recession, yet, there is no official or generally
agreed definition of recession. Although NBER’s archive represents recessions in US
thoroughly, such a statistics does not exist for the other countries subject to our analysis 2.
There is a piece of literature dealing with recession definition from business cycles. In this
literature, the beginning and the end of a recession are turning points in the business cycle,
i.e. the beginning represents a peak in the cycle while the end represents a trough. Artis et
al., (1997) proposes an algorithm detecting turning points in industrial production series of
G7 and the European countries. Ross and Ubide (2001) also apply the same methodology to
the Euro area. This methodology utilizes binary response model and estimate recession. We
opt for more a literal definition of recession. There is not a clear cut definition of recession,
yet, we define recession as follows:
Definition: A recession has occurred if a country reports two consecutive quarters of
negative GDP growth.
For estimation purposes, our binary variables will represent two different approaches in
terms of measuring the duration of such recession. In Model 1, we consider that a recession
actually happened and binary variable takes the value “1” starting only on the second
quarter of negative GDP growth. Alternatively, Model 2 considers a recession actually
happened in both first and second quarters of negative GDP growth, as long as both
quarters are consecutive.
3.3. Data
For the econometric modeling we use data from 32 OECD. To obtain evidences on the
usefulness of the slope of the yield curve for predicting recessions in the countries of
interest, we should make use of long time series. The problem here is that some of the
countries in our analysis have very short time series. Moreover to the best of our
knowledge, no attempt has been made to estimate recession in these countries due to this
lack of long time series. To deal with this problem we construct a database composed of the
long – term and short term rates of countries from OECD official website and adopt a panel
data framework. We retrieved quarterly data starting from 1990Q1 to 2011Q13.
The data is not homogenous in terms of their starting data. One of the main foci of this
study is to predict recession in countries that have short time series. In our country set, there
are countries having relatively shorter time series and are subject to repercussions of Euro
zone debt crisis. Recession fears and yield curve dynamics in Euro-zone countries that have
short time series will be interesting to analyse. Therefore, we pay special attention to the
data for Portugal, Iceland, Ireland, Greece and Spain. The data for these countries are also
2
3
Please visit: http://www.nber.org/cycles.html
Details on time series for each country are available in Appendix Table 1.
6
shorter. Investigating the relationship between slope of yield curve and recession with panel
data will enable us to have an insight also for these countries.
3.4. Methodology
We estimate an unbalanced panel model of 32 countries from 1990Q1 to 2011Q1 in order
to ensure robustness of the results, we compare the three most common methodologies for
estimation, namely: Probit Model, Logit Model and MLE. Equation (1) below is the basic
model we estimate. We also estimate a modified model, i.e. equation (2), which includes
the autoregressive series of the state of the economy (the indicator of recession or
expansion currently). A measure of accuracy of the forecast was calculated and compared
among different specifications of the models.
Re cessioni ,t , m   0,i  1,i slopei ,t  4   i ,t
(1)
Re cessioni ,t , m   0,i  1,i slopei ,t  4   2,i GDPi ,t  4   i ,t
(2)
where m  1,2 are the models, that differ in terms of the definition of recession.
i  1,...,32 are the cross section units, represented by the OECD countries and
t  1990Q1,...,2011Q1 represents the time dimension of quarterly data. In our analysis we
define recession as follows:
1, recession occur
Re cessioni ,t ,m  
0, recession does not occur
Regarding the probit and logit model we employed, the main rationale behind including the
lagged GDP into the basic model is as follows: one of the main assumptions of the binary
response models is that the random shocks are i.i.d. normal random variables with zero
mean. In this kind of model the errors are generally auto-correlated. In traditional time
series approach this kind of problem is removed by using an autoregressive moving average
(ARMA) filter. Here, since the shocks to 𝑢 are unobservable this technique is not available.
Therefore, we adopt the solution proposed by Dueker (1997) to remove the serial
correlation in 𝑢 by adding a lag of GDP.
Adding a lag of GDP will also help to deal with auto–correlation in MLE, although ARMA
type of filters is preferred. Yet, we follow a single procedure to remove auto–correlation,
since ARMA filters are not available for binary response models.
3.4. Econometric Procedures
The econometrics of binary outcome models concern is to give the correct treatment to the
discreteness of the dependent variable, and in our case to constrain predicted probabilities
between zero and one. Traditional MLE estimator ignores these concerns. The likelihood
function is the joint density, which given independent observations are the product
Π𝑖 𝑓(𝑦𝑖 |𝑥𝑖 , 𝛽) of the individual densities. The log likelihood function is then the log of a
product, which equals the sum of logs. The estimator 𝛽̂ maximizes:
7
𝑁
1
∑ ln 𝑓(𝑦𝑖 |𝑥𝑖 , 𝛽)
𝑁
𝑖=1
To properly deal with binary variables, literature applies both logit a probit models.
The commonly used models take the form of conditional probability given by:
𝑄𝑁 (𝛽) =


pi  Pr yi 1 x  F ( xi'  )
where yi stands for the dependent variable and xi' stands for a matrix of independent
variables. In the Logit model, function F (.) takes the form of:
ex 
 
'
p F x  
'
i
1 e x 
And for the Probit model the analogous probability is:
'
p   xi'     ( z )dz
xi' 

where  (.) is the standard normal cumulative distribution function with derivative
  z2 
 1 
 , which is the standard normal density function.
 exp 
 2 
 2 
The choice between the econometric models has theoretical and empirical considerations.
Theoretically speaking, the appropriate model depends on the data generation process,
which is unknown. Also theoretically, the MLE model would not be chosen, since it
violates the probability boundaries. However, empirically, there is little or no difference
between predicted probabilities.
 ( z)  
4. Results
The goodness of fit for binary outcome models is usually evaluated by a comparison of
fitted and actual values, although there are alternative diagnostics detailed in Amemiya
(1981) and Maddala (1983).
Consider yi as the dependent variable, and ŷi a prediction, the criterion   yi  yˆi 
2
gives
i
the number of wrong predictions. The most intuitive prediction rule is to set yˆ 1
pˆ  F ( x'  )  0.5 . However, if most of the sample has yˆ 1 , then often
when
  y  yˆ  n(1  y ) ,
2
i
i
i
since it is likely that pˆ  0.5 , and hence yˆ 1 for all observation.
To overcome this problem, typically a cutoff value is considered, letting yˆ 1 if pˆ  c . We
follow the traditional Type I and Type II error definition, commonly used to forecast
business failure. Typically, a Type I error refers to failed firms that are classified by the
models as non-failed ones, while type II error refers to non-failed firms that are classified as
failed (see Zopounidis and Doumpos (1999) and Boritz and Kennedy (1995)). Our
approach for defining Type I and Type II error, follows the same rationale, although it is
adapted to better fit our framework.
8
Boritz and Kennedy (1995) argue that there is no generally accepted basis for trading off
Type I and Type II errors. There may have different perceptions of the relative costs of
Type I and Type II errors. In our analysis we prefer to minimize Type II error and tolerate
Type I error. Therefore, we define the power of estimation as (1 – Type II error). The
following table describes the resulting Type I and Type II errors in our study.
Actually Recession
Actually Non-Recession
Predicted Recession
Correct
Type II error
Predicted Non-Recession
Type I error
Correct
Our results are based on an unbalanced panel data set. This issue is important especially
when one is not able to assure that data was randomly missing (see Baltagi (2005)). To
avoid any critics regarding this issue, we compare the results of our unbalanced panel with
a balanced panel, where we excluded countries that have missing data. As results are almost
the same in terms of statistical significance, fitting and power, we will present only results
for unbalanced data4.
Table I shows the estimation results for Model 1. The model is estimated through probit,
logit and MLE. The parameter estimates for slope and GDP data are both significant at 5%.
Table II presents the cut-off value, Type I, Type II and the power of each estimation. We
fix exogenously level of Type I errors as close as possible to 25%. This means that the
power we present in Table II is compatible with a model that gives one “false alarm” for
each four recession predictions. The results are interesting in a sense that the powers are
roughly the same at 63%. This means that a recession is correctly predicted 63% of the
time, even for countries with short time series. Furthermore, one should note that Log
Likelihood, AIC, BIC and Pseudo R2 are very close in Logit and Probit estimates, and they
slightly improve with the inclusion of lagged GDP. Results for MLE should be used only
for comparison purposes, since they do not represent properly our binary framework. By
combining Tables I and II, one can confirm the goodness of fit and robustness of our panel
data models and its reliability to predict future recessions.
4
Results for balanced panel data are available upon request.
9
Table I – Estimation Results for Model 1
Slope (t-4)
GDP (t-4)
Constant
Observations
No. of countries
Log Likelihood
AIC
BIC
Pseudo R2
Probit
-0.107**
-0.113**
(0.017)
(0.018)**
-0.086
(0.037)
-1.397**
-1.317**
(0.061)
(0.065)
2086
1959
32
32
-562,339
-541,564
1130,678
1091,128
1147,607
1113,449
0,36389
0,38739
Logit
-0.230**
-0.262**
(0.042)
(0.046)
-0.152**
(0.069)
-2.433**
-2.291**
(0.124)
(0.135)
2086
1959
32
32
-563,378
-541,753
1132,756
1091,507
1146,685
1113,827
0,362848
0,387305
MLE
-0.020** -0.022**
(0.003)
(0.003)
-0.013**
(0.006)
0.095**
0.110**
(0.008)
(0.010)
2086
1959
32
32
-219,218 -243,773
446,4365 497,5462
469,0085 525,4472
-0,236
-0,37444
Note: * and ** indicate 90% and 95% of statistical significance level, respectively. In parenthesis are the
standard deviations. AIC and BIC stand for Akaike and Bayesian information criteria, respectively.
Table II – Power of Model 2
Cut-off
Type I Error
Type II Error
Power
Probit
Without
GDP
With GDP
0,07
0,07
25,0%
25,3%
36,5%
36,3%
63,5%
63,7%
Logit
Without
GDP
With GDP
0,07
0,07
25,0%
25,3%
36,6%
36,2%
63,4%
63,8%
MLE
Without
With
GDP
GDP
0,08
0,08
25,0%
25,3%
36,0%
36,2%
64,0%
63,8%
Table III shows estimation results for Model II. All coefficients have proper signs and are
statistically significant. Pseudo-R2 values in Table II also increase with the inclusion of
lagged GDP, which can be understood as better fit of the model. According to Pseudo-R2,
AIC and BIC one would prefer to rely on the model with lagged GDP.
Results in Table IV show that the Power of the models decrease with the second definition
of recession, to the range 56.9-58.4%. Therefore, the models are sensitive to the definition
of recession. However, robustness of the models are again reaffirmed, since the Power is
roughly the same for Logit, Probit and MLE.
10
Table III - Estimation results for Model 2
Slope (t-4)
GDP (t-4)
Constant
Observations
No. of countries
Log Likelihood
AIC
BIC
Pseudo R2
Probit
-0.105**
-0.113**
(0.017)
(0.017)
-0.061*
(0.034)
-1.211**
-1.139**
(0.068)
(0.072)
2086
1959
32
32
-712,16
-685,327
1430,32
1378,655
1447,249
1400,975
0,370526
0,394243
Logit
-0.218** -0.253**
(0.038)
(0.052)
-0.101*
(0.061)
-2.069** -1.938**
(0.130)
(0.139)
2086
1959
32
32
-713,012 -685,041
1432,042 1378,082
1448,971 1400,403
0,369962 0,394678
MLE
-0.023** 0.026**
(0.004)
(0.004)
-0.012*
(0.006)
0.130** 0.147**
(0.012)
(0.131)
2086
1959
32
32
-540,678 -542,588
1089,355 1095,177
1111,927 1123,078
0,279072 0,276525
Note: * and ** indicate 90% and 95% of statistical significance level, respectively. In parenthesis are the
standard deviations. AIC and BIC stand for Akaike and Bayesian information criteria, respectively.
Table IV– Power of Model 2
Probit
Cut-off
Type I Error
Type II Error
Power
Without GDP
0,10
25,1%
43,1%
56,9%
With GDP
0,10
25,0%
41,6%
58,4%
Logit
Without
GDP
With GDP
0,10
0,10
25,1%
25,0%
43,1%
43,1%
56,9%
56,9%
MLE
Without
With
GDP
GDP
0,11
0,12
25,1%
25,0%
43,1%
40,3%
56,9%
59,7%
Figure 1 below plots the probability of a recession in the blue line, against vertical bars
representing actual recessions. The red dashed line represents our cutoff value. It is
interesting to note that the model we employed is successful in predicting recession one
year ahead in countries where there exist long time series. Especially, it is striking that the
recessions both in US and UK have been predicted successfully. In both country cases,
there are some quarters when there was not an actual recession, and a recession is falsely
estimated, i.e. false alarms, though. Mexico is another very good case where 8 recession
period has been predicted very successfully. The latest recession period has been predicted
on the edge, where the estimated value is roughly the cut-off value. The model is also very
successful in the remainder of country cases.
11
0
Recession
Recession
Recession
Slope
Slope
Slope
Cutoff
1
Cutoff
0
0,02
Australia
Austria
1
Belgium
Canada
0
0
0,18
0,16
0,14
0,12
0,1
0,08
0,06
0,04
0,02
0
1
0,14
1
Cutoff
0
1
Cutoff
0,02
0
0
Finland
0,25
0
0
Q1-1992
Q1-1993
Q1-1994
Q1-1995
Q1-1996
Q1-1997
Q1-1998
Q1-1999
Q1-2000
Q1-2001
Q1-2002
Q1-2003
Q1-2004
Q1-2005
Q1-2006
Q1-2007
Q1-2008
Q1-2009
Q1-2010
Q1-2011
0,16
Q1-1991
Slope
Q1-1990
1
Q1-1990
Q1-1991
Q1-1992
Q1-1993
Q1-1994
Q1-1995
Q1-1996
Q1-1997
Q1-1998
Q1-1999
Q1-2000
Q1-2001
Q1-2002
Q1-2003
Q1-2004
Q1-2005
Q1-2006
Q1-2007
Q1-2008
Q1-2009
Q1-2010
Q1-2011
Q1-1990
Q1-1991
Q1-1992
Q1-1993
Q1-1994
Q1-1995
Q1-1996
Q1-1997
Q1-1998
Q1-1999
Q1-2000
Q1-2001
Q1-2002
Q1-2003
Q1-2004
Q1-2005
Q1-2006
Q1-2007
Q1-2008
Q1-2009
Q1-2010
Q1-2011
Q1-1990
Q1-1991
Q1-1992
Q1-1993
Q1-1994
Q1-1995
Q1-1996
Q1-1997
Q1-1998
Q1-1999
Q1-2000
Q1-2001
Q1-2002
Q1-2003
Q1-2004
Q1-2005
Q1-2006
Q1-2007
Q1-2008
Q1-2009
Q1-2010
Q1-2011
0
Q1-1992
Q1-1993
Q1-1994
Q1-1995
Q1-1996
Q1-1997
Q1-1998
Q1-1999
Q1-2000
Q1-2001
Q1-2002
Q1-2003
Q1-2004
Q1-2005
Q1-2006
Q1-2007
Q1-2008
Q1-2009
Q1-2010
Q1-2011
Recession
Cutoff
Q1-1990
Q1-1991
Q1-1992
Q1-1993
Q1-1994
Q1-1995
Q1-1996
Q1-1997
Q1-1998
Q1-1999
Q1-2000
Q1-2001
Q1-2002
Q1-2003
Q1-2004
Q1-2005
Q1-2006
Q1-2007
Q1-2008
Q1-2009
Q1-2010
Q1-2011
Q1-1990
Q1-1991
Q1-1992
Q1-1993
Q1-1994
Q1-1995
Q1-1996
Q1-1997
Q1-1998
Q1-1999
Q1-2000
Q1-2001
Q1-2002
Q1-2003
Q1-2004
Q1-2005
Q1-2006
Q1-2007
Q1-2008
Q1-2009
Q1-2010
Q1-2011
Slope
0,2
0,18
0,16
0,14
0,12
0,1
0,08
0,06
0,04
0,02
0
Q1-1991
Q1-1990
Q1-1991
Q1-1992
Q1-1993
Q1-1994
Q1-1995
Q1-1996
Q1-1997
Q1-1998
Q1-1999
Q1-2000
Q1-2001
Q1-2002
Q1-2003
Q1-2004
Q1-2005
Q1-2006
Q1-2007
Q1-2008
Q1-2009
Q1-2010
Q1-2011
1
Q1-1990
Q1-1990
Q1-1991
Q1-1992
Q1-1993
Q1-1994
Q1-1995
Q1-1996
Q1-1997
Q1-1998
Q1-1999
Q1-2000
Q1-2001
Q1-2002
Q1-2003
Q1-2004
Q1-2005
Q1-2006
Q1-2007
Q1-2008
Q1-2009
Q1-2010
Q1-2011
Recession
Q1-2000
Q1-2001
Q1-2002
Q1-2003
Q1-2004
Q1-2005
Q1-2006
Q1-2007
Q1-2008
Q1-2009
Q1-2010
Q1-2011
0
Q1-1990
Q1-1991
Q1-1992
Q1-1993
Q1-1994
Q1-1995
Q1-1996
Q1-1997
Q1-1998
Q1-1999
Figure 1. Recessions for Countries with Long Time Series (in-sample)
UK
1
US
0,14
0,12
0,1
0,08
0,06
Recession
Slope
Germany
Recession
Slope
Recession
Recession
Recession
Slope
Slope
Slope
Cutoff
0,04
0
1
0,14
0,12
0,1
0,06
0,04
Cutoff
0
0,12
0,1
0,08
0,06
0,04
Cutoff
1
0,2
0,15
0,05
Cutoff
0,02
0
France
0,25
0,2
0,08
0,15
0,1
Cutoff
0,05
0
0,14
0,12
0,1
0,08
0,06
0,04
0,02
0
0,18
0,16
0,14
0,12
0,1
0,08
0,06
0,04
0,02
0
Denmark
0,3
0,25
0,1
0,2
0,15
0,1
0,05
0
12
Q1-1990
Q1-1991
Q1-1992
Q1-1993
Q1-1994
Q1-1995
Q1-1996
Q1-1997
Q1-1998
Q1-1999
Q1-2000
Q1-2001
Q1-2002
Q1-2003
Q1-2004
Q1-2005
Q1-2006
Q1-2007
Q1-2008
Q1-2009
Q1-2010
Q1-2011
0
Recession
Recession
0
Slope
Slope
Q1-2011
Q1-2010
Q1-2009
Q1-2008
Q1-2007
Q1-2006
Q1-2005
0
1
New Zealand
Netherlands
Cutoff
1
Norway
Portugal
Cutoff
1
0
0
0,18
0,16
0,14
0,12
0,1
0,08
0,06
0,04
0,02
0
1
0,25
1
0
0,12
0
Recession
Slope
0,05
Spain
Cutoff
0
Q1-1999
Q1-2000
Q1-2001
Q1-2002
Q1-2003
Q1-2004
Q1-2005
Q1-2006
Q1-2007
Q1-2008
Q1-2009
Q1-2010
Q1-2011
0,2
Q1-1998
Q1-1996
Q1-1997
Q1-1994
Q1-1995
Q1-1990
Q1-1991
Q1-1992
Q1-1993
Q1-1994
Q1-1995
Q1-1996
Q1-1997
Q1-1998
Q1-1999
Q1-2000
Q1-2001
Q1-2002
Q1-2003
Q1-2004
Q1-2005
Q1-2006
Q1-2007
Q1-2008
Q1-2009
Q1-2010
Q1-2011
0
Q1-1992
Q1-1993
Cutoff
Q1-1996
Q1-1997
Q1-1998
Q1-1999
Q1-2000
Q1-2001
Q1-2002
Q1-2003
Q1-2004
Q1-2005
Q1-2006
Q1-2007
Q1-2008
Q1-2009
Q1-2010
Q1-2011
Slope
Q1-2004
Q1-2003
Q1-2002
Q1-2001
Slope
Q1-1990
Q1-1991
Recession
Q1-2000
Q1-1999
Q1-1998
Q1-1997
Q1-1996
Q1-1995
Q1-1994
Q1-1993
Q1-1992
Q1-1991
Q1-1990
Recession
Q1-1990
Q1-1991
Q1-1992
Q1-1993
Q1-1994
Q1-1995
Q1-1990
Q1-1991
Q1-1992
Q1-1993
Q1-1994
Q1-1995
Q1-1996
Q1-1997
Q1-1998
Q1-1999
Q1-2000
Q1-2001
Q1-2002
Q1-2003
Q1-2004
Q1-2005
Q1-2006
Q1-2007
Q1-2008
Q1-2009
Q1-2010
Q1-2011
0
Q1-1994
Q1-1995
Q1-1996
Q1-1997
Q1-1998
Q1-1999
Q1-2000
Q1-2001
Q1-2002
Q1-2003
Q1-2004
Q1-2005
Q1-2006
Q1-2007
Q1-2008
Q1-2009
Q1-2010
Q1-2011
Q1-1990
Q1-1991
Q1-1992
Q1-1993
1,2
Q1-1990
Q1-1991
Q1-1992
Q1-1993
Q1-1994
Q1-1995
Q1-1996
Q1-1997
Q1-1998
Q1-1999
Q1-2000
Q1-2001
Q1-2002
Q1-2003
Q1-2004
Q1-2005
Q1-2006
Q1-2007
Q1-2008
Q1-2009
Q1-2010
Q1-2011
Q1-1990
Q1-1991
Q1-1992
Q1-1993
Q1-1994
Q1-1995
Q1-1996
Q1-1997
Q1-1998
Q1-1999
Q1-2000
Q1-2001
Q1-2002
Q1-2003
Q1-2004
Q1-2005
Q1-2006
Q1-2007
Q1-2008
Q1-2009
Q1-2010
Q1-2011
1
Mexico
1
Italy
1
0,8
0,6
0,4
Recession
1
Cutoff
Slope
Recession
Slope
Recession
Slope
Recession
Slope
Cutoff
Cutoff
0
0,2
0,15
0,1
0
Cutoff
1
0,1
0,08
0,06
0,04
0,02
Cutoff
0,18
0,16
0,14
0,12
0,1
0,08
0,06
0,04
0,02
0
0,16
0,14
0,12
0,1
0,08
0,06
0,04
0,02
0
0,12
0,1
0,08
0,06
0,04
0,02
0
Sweden
0,18
0,16
0,14
0,12
0,1
0,08
0,06
0,04
0,02
0
Switzerland
0,18
0,16
0,14
0,12
0,1
0,08
0,06
0,04
0,02
0
Note: The figures above plot the probability of a recession against vertical lines representing actual recessions. Grey vertical lines are
actual recessions where blue lines are estimated values. Red dash line is the cut – off value. If the blue line exceeds the cut – off value a
year ahead of actual recession, then the model predicts recession successfully. Otherwise, if the blue line exceeds the cut – off value but
there has not occurred an actual recession a year after, or vice versa, the model is unsuccessful.
In the same way, Figure 2 below plots the probability of a recession in the blue line, against
vertical bars representing actual recessions. The red dashed line represents our cutoff value.
13
Countries in Figure 2 are those with relatively short time series. Recessions in those
countries are also predicted quite well. The model predictions for Portugal, Ireland, Iceland,
Greece and Spain are quite satisfactory. For instance, the very recessions after 2009Q1 are
very well predicted. Yet the remainder has not been so. The reason for this may come from
the market sentiment toward the afore-mentioned countries, especially for Greece. As
rumors during and after 2010, lead to heightened default risk. Short after the increased
anxiety, the top level policy makers commented on the rumors about a possible Greek
default and calmed down the markets. Furthermore, all Euro-zone officials gave implicit
guarantees not to let Greece to default. Although the financial markets were calmed down
thanks to officials’ announcement, the economic activity in Greece continued to slow down
and recession occurred. We deem the slope of the yield curve in Greece has not been
successful for this reason. Yet, recessions in Portugal, Ireland, Iceland and Spain is well
predicted.
Another interesting case is Chile. Chile has a quarterly data starting from 2005. Such a
short time series would not allow predicting recession in the country if the analysis were
carried out with times series. Yet, thanks to the panel data framework, recessions during
2008 – 2009 have been predicted quite well. There recession period have been successfully
predicted.
This holds also for Japan, Hungary, Ireland and the others whose data starts after 2000.
Yield curve have been very successful in predicting recession. For instance, Hungary has
five quarters of recession before and after 2009. Five quarters of recession have been
successfully predicted. For Japan, the estimated values are below but one year ahead
estimated values rises very near to cut – off values although not successful enough to
predict recessions. The remaining country cases are also interesting but the comments are
left to the readers due to lack of space.
14
0
Recession
Recession
Slope
Recession
Slope
Slope
Slope
Cutoff
1
0
Cutoff
1
Cutoff
1
0
Q1-1990
Q1-1991
Q1-1992
Q1-1993
Q1-1994
Q1-1995
Q1-1996
Q1-1997
Q1-1998
Q1-1999
Q1-2000
Q1-2001
Q1-2002
Q1-2003
Q1-2004
Q1-2005
Q1-2006
Q1-2007
Q1-2008
Q1-2009
Q1-2010
Q1-2011
0,12
0
0,12
0
Q1-1990
Q1-1991
Q1-1992
Q1-1993
Q1-1994
Q1-1995
Q1-1996
Q1-1997
Q1-1998
Q1-1999
Q1-2000
Q1-2001
Q1-2002
Q1-2003
Q1-2004
Q1-2005
Q1-2006
Q1-2007
Q1-2008
Q1-2009
Q1-2010
Q1-2011
Q1-1990
Q1-1991
Q1-1992
Q1-1993
Q1-1994
Q1-1995
Q1-1996
Q1-1997
Q1-1998
Q1-1999
Q1-2000
Q1-2001
Q1-2002
Q1-2003
Q1-2004
Q1-2005
Q1-2006
Q1-2007
Q1-2008
Q1-2009
Q1-2010
Q1-2011
Recession
Q1-1990
Q1-1991
Q1-1992
Q1-1993
Q1-1994
Q1-1995
Q1-1996
Q1-1997
Q1-1998
Q1-1999
Q1-2000
Q1-2001
Q1-2002
Q1-2003
Q1-2004
Q1-2005
Q1-2006
Q1-2007
Q1-2008
Q1-2009
Q1-2010
Q1-2011
Q1-1990
Q1-1991
Q1-1992
Q1-1993
Q1-1994
Q1-1995
Q1-1996
Q1-1997
Q1-1998
Q1-1999
Q1-2000
Q1-2001
Q1-2002
Q1-2003
Q1-2004
Q1-2005
Q1-2006
Q1-2007
Q1-2008
Q1-2009
Q1-2010
Q1-2011
1
Japan
Hungary
Chile
Czech Republic
0
0
0,18
0,16
0,14
0,12
0,1
0,08
0,06
0,04
0,02
0
1
0,25
1
0,05
Cutoff
Q1-1990
Q1-1991
Q1-1992
Q1-1993
Q1-1994
Q1-1995
Q1-1996
Q1-1997
Q1-1998
Q1-1999
Q1-2000
Q1-2001
Q1-2002
Q1-2003
Q1-2004
Q1-2005
Q1-2006
Q1-2007
Q1-2008
Q1-2009
Q1-2010
Q1-2011
Q1-1990
Q1-1991
Q1-1992
Q1-1993
Q1-1994
Q1-1995
Q1-1996
Q1-1997
Q1-1998
Q1-1999
Q1-2000
Q1-2001
Q1-2002
Q1-2003
Q1-2004
Q1-2005
Q1-2006
Q1-2007
Q1-2008
Q1-2009
Q1-2010
Q1-2011
0
Q1-1990
Q1-1991
Q1-1992
Q1-1993
Q1-1994
Q1-1995
Q1-1996
Q1-1997
Q1-1998
Q1-1999
Q1-2000
Q1-2001
Q1-2002
Q1-2003
Q1-2004
Q1-2005
Q1-2006
Q1-2007
Q1-2008
Q1-2009
Q1-2010
Q1-2011
Figure 2. Recessions for Countries with Short Time Series (in-sample)
Greece
1
Iceland
0,1
0,08
0,06
0,04
0,02
0
Recession
Spain
Recession
Slope
Recession
Slope
Recession
Slope
Slope
Cutoff
1
0,02
Cutoff
0
0,2
0,15
0,1
0
Cutoff
0,5
0,45
0,4
0,35
0,3
0,25
0,2
0,15
0,1
0,05
0
Portugal
0,1
0,12
0,1
0,08
0,08
0,06
0,06
0,04
0,04
0,02
0
0,25
0,2
0,15
0,1
Cutoff
0,05
0
0,12
0,1
0,08
0,06
0,04
0,02
0
15
0
Recession
Slope
Q1-2010
Q1-2009
Q1-2008
1
Cutoff
1
Cutoff
0
1
Cutoff
Q1-1990
Q1-1991
Q1-1992
Q1-1993
Q1-1994
Q1-1995
Q1-1996
Q1-1997
Q1-1998
Q1-1999
Q1-2000
Q1-2001
Q1-2002
Q1-2003
Q1-2004
Q1-2005
Q1-2006
Q1-2007
Q1-2008
Q1-2009
Q1-2010
Q1-2011
Q1-1990
Q1-1991
Q1-1992
Q1-1993
Q1-1994
Q1-1995
Q1-1996
Q1-1997
Q1-1998
Q1-1999
Q1-2000
Q1-2001
Q1-2002
Q1-2003
Q1-2004
Q1-2005
Q1-2006
Q1-2007
Q1-2008
Q1-2009
Q1-2010
Q1-2011
Cutoff
Slovak Republic
Slovenia
0
0
0,18
0,16
0,14
0,12
0,1
0,08
0,06
0,04
0,02
0
1
Q1-1990
Q1-1991
Q1-1992
Q1-1993
Q1-1994
Q1-1995
Q1-1996
Q1-1997
Q1-1998
Q1-1999
Q1-2000
Q1-2001
Q1-2002
Q1-2003
Q1-2004
Q1-2005
Q1-2006
Q1-2007
Q1-2008
Q1-2009
Q1-2010
Q1-2011
Slope
Q1-2007
Slope
Q1-2006
Slope
Q1-2005
Q1-2004
Recession
Q1-2003
Recession
Q1-2002
Q1-1990
Q1-1991
Q1-1992
Q1-1993
Q1-1994
Q1-1995
Q1-1996
Q1-1997
Q1-1998
Q1-1999
Q1-2000
Q1-2001
Q1-2002
Q1-2003
Q1-2004
Q1-2005
Q1-2006
Q1-2007
Q1-2008
Q1-2009
Q1-2010
Q1-2011
Recession
Q1-2001
Q1-2000
Q1-1999
Q1-1998
Q1-1997
0
Q1-1996
Q1-1990
Q1-1991
Q1-1992
Q1-1993
Q1-1994
Q1-1995
Q1-1996
Q1-1997
Q1-1998
Q1-1999
Q1-2000
Q1-2001
Q1-2002
Q1-2003
Q1-2004
Q1-2005
Q1-2006
Q1-2007
Q1-2008
Q1-2009
Q1-2010
Q1-2011
0
Q1-1995
Q1-1994
Q1-1993
Q1-1992
Q1-1991
Q1-1990
0,16
0,02
0
0,35
0,05
0
0,35
0,05
Q1-1990
Q1-1991
Q1-1992
Q1-1993
Q1-1994
Q1-1995
Q1-1996
Q1-1997
Q1-1998
Q1-1999
Q1-2000
Q1-2001
Q1-2002
Q1-2003
Q1-2004
Q1-2005
Q1-2006
Q1-2007
Q1-2008
Q1-2009
Q1-2010
Q1-2011
Q1-1990
Q1-1991
Q1-1992
Q1-1993
Q1-1994
Q1-1995
Q1-1996
Q1-1997
Q1-1998
Q1-1999
Q1-2000
Q1-2001
Q1-2002
Q1-2003
Q1-2004
Q1-2005
Q1-2006
Q1-2007
Q1-2008
Q1-2009
Q1-2010
Q1-2011
1
Ireland
1
Korea
0,14
0,12
0,1
0,08
0,06
0,04
0
Recession
Slope
Luksembourg
Recession
Slope
Poland
Recession
Recession
Cutoff
Slope
Slope
0,2
0,18
0,16
0,14
0,12
0,1
0,08
0,06
0,04
0,02
0
1
Israel
0,3
0,25
0,2
0,15
0,1
0
Cutoff
0,45
0,4
0,35
0,3
0,25
0,2
0,15
0,1
0,05
0
Korea
0,3
1
0,25
0,2
0,15
0,1
Cutoff
Cutoff
0
0,2
0,18
0,16
0,14
0,12
0,1
0,08
0,06
0,04
0,02
0
0,16
0,14
0,12
0,1
0,08
0,06
0,04
0,02
0
Note: The figures above plot the probability of a recession against vertical lines representing actual recessions. Grey vertical lines are
actual recessions where blue lines are estimated values. Red dash line is the cut – off value. If the blue line exceeds the cut – off value a
year ahead of actual recession, then the model predicts recession successfully. Otherwise, if the blue line exceeds the cut – off value but
there has not occurred an actual recession a year after, or vice versa, the model is unsuccessful.
16
5. Concluding Remarks
In this paper the importance of the use of the term spread as predictor of recessions is
confirmed for OECD countries. The results of this paper show that yield spread appears to
contain useful information for predicting recessions. To arrive at this conclusion we define
recession as two consecutive quarters of negative GDP growth. Models differentiate in
terms of the usage of the definition of recession. Both two consecutive quarters adopted as
recession in Model II and only the second quarter as recession in Model I. Variables are
lagged by four quarters, to allow one to predict recessions one year ahead. In order to
control for current dynamics of the economy we also controlled for the GDP growth, i.e.
four quarter lagged GDP growth is added, in a separate estimation. We employed probit,
logit and MLE with panel setting in estimation process. Model specification used our
analysis is proposed by Estrella and Mishkin (1998).
We found that the use of a lagged GDP growth variable helps to forecast historically
recessions in the countries. We have taken into consideration of the accuracy of forecast
specifically. We carried out an exercise of out-of-sample forecasting to investigate the outof-sample performance of the models especially for those countries who have shorter time
series. The results present that the model proposed by Estrella and Mishkin (1998) have
robust predictive power not only for the countries who have long time series but also the
ones who have shorter time series. Panel data estimation also suggests great goodness of fit.
The model estimations deliver power close to 65% with type I error of 25% with good
statistical properties.
As a suggestion for future research, we believe that this model can be applied to several
other countries, even those with short time series. Furthermore, literature still lacks some
results about how other yield curve factors can be applied in these binary models to predict
recession.
References
Artis, M. J., Z.G. Kontolemis, and D. R. Osborn, 1997. "Business Cycles for G7 and
European Countries" The Journal of Business, University of Chicago Press, Vol. 70,
No: 2, 249-79.
Afonso, A. and M. M. F. Martins, 2010.
“Level, Slope, Curvature of the
Sovereign Yield Curve, and Fiscal Behaviour” . ECB Working Paper No. 1276.
Ahrens, R., 2002. “Predicting Recessions with Interest Rate Spreads: A Multicountry
Regime-Switch Analysis.” Journal of International Money and Finance, Vol. 21,
No: 4, 519-37.
Amemiya, T., 1981. “Qualitative Response Models: A Survey”. Journal of Economic
Literature, Vol. 19, 1483-1536.
Anderson, N., F. Breedon, M. Deacon, A. Derry and G. Murphy, 1996. “Estimating And
Interpreting The Yield Curve” Wiley, Chichester.
Ang, A., and M. Piazzesi. 2001. "A No-Arbitrage Vector Autoregression of Term Structure
Dynamics with Macroeconomic and Latent Variables." Working Paper. Columbia
University. Forthcoming, Journal of Monetary Economics.
17
Artis, M. J., Z. G. Kontolemis, and D. R. Osborn, 1997. “Business Cycles for G7 and
European Countries" Journal of Business, vol. 70, no. 2
Boritz, J. Efrim and Duane B. Kennedy, 1995. “Effectiveness of Neural Network Types for
Prediction of Business Failure”. Expert Systems with Applications, Vol. 9, No: 4,
503-512.
Arturo Estrella & Frederic S. Mishkin, 1998. "Predicting U.S. Recessions: Financial
Variables As Leading Indicators," The Review of Economics and Statistics, MIT
Press, vol. 80(1), 45-61.
Baltagi, B. H., 2005. “Econometric Analysis of Panel Data”. John Wiley & Sons,
Chichester, England.
Dueker, M. J., 1997. “Strengthening the case for the yield curve as predictor of U.S.
recessions", Federal Reserve Bank St. Louis Review, Vol. 79. pp. 41-51.
Constantin Zopounidis and Michael Doumpos, 1999. “Business failure prediction using the
UTADIS multicriteria analysis method”. The Journal of the Operational Research
Society, vol. 50(11), 1138-1148.
Estrella, A., and M. Trubin, 2006. “The Yield Curve as a Leading Indicator: Some Practical
Issues” Federal Reserve Bank Of New York, Vol. 12, No:5.
Evans, C., and D. Marshall. 2001. "Economic Determinants of the Nominal Treasury Yield
Curve." FRB ChicagoWorking Paper 01-16.
Harvey, C.R., 1988. "The Real term Structure and Consumption Growth " Journal of
Financial Economics , Vol. 22, No. 2. 305-335.
Hu, Z., 1993. “The yield curve and real activity," IMF Staff Paper, No.40, pp. 781-806,
December.
Karunaratne, N. D., 2002. “Predicting Australian Growth and Recession via the Yield
curve” Economic Analysis & Policy. Vol. 32, No:2, 233-250.
Maddala, G. S. “Limited-Dependent and Qualitative Variables in Economics”, Cambridge,
UK, Cambridge University Press, 1983.
Pace, P. D., 2012. “Gross Domestic Product Growth Predictions Through The Yield
Spread: Time-Variation And Structural Breaks” International Journal of Finance
and Economics (article first published online: 29 Feb 2012)
Ross, K., and A. Ubide, 2001. “Mind the gap: What is the best measure of slack in the euro
area? ", IMF Working Paper, December.
Wu, T., 2001. "Monetary Policy and the Slope Factor in Empirical Term Structure
Estimations." FRB San Francisco Working Paper 2002-07.
18
Appendix Table 1. List of Countries and Initial Data Dates
GDP
Country
Australia
Austria
Belgium
Canada
Chile
Czech Republic
Denmark
Finland
France
Germany
Greece
Hungary
Iceland
Ireland
Israel
Italy
Japan
Korea
Luxembourg
Mexico
Netherlands
New Zealand
Norway
Poland
Portugal
Slovak Republic
Slovenia
Spain
Sweden
Switzerland
United Kingdom
United States
Start
1990Q1
1990Q1
1995Q2
1990Q1
1996Q1
1996Q2
1991Q2
1990Q1
1990Q1
1991Q2
2000Q2
1995Q2
1997Q2
2000Q2
1995Q2
1990Q1
1990Q1
1990Q1
1995Q2
1993Q2
1990Q1
1990Q1
1990Q1
1995Q2
1995Q2
1997Q2
1996Q1
1995Q2
1993Q2
1990Q1
1990Q1
1990Q1
End
2011Q1
2011Q1
2011Q1
2011Q1
2011Q1
2011Q1
2011Q1
2011Q1
2011Q1
2011Q1
2011Q1
2011Q1
2011Q1
2011Q1
2011Q1
2011Q1
2011Q1
2011Q1
2011Q1
2011Q1
2011Q1
2011Q1
2011Q1
2011Q1
2011Q1
2011Q1
2011Q1
2011Q1
2011Q1
2011Q1
2011Q1
2011Q1
Slope
Start
1990Q1
1990Q1
1990Q1
1990Q1
2004Q3
2000Q2
1990Q1
1990Q1
1990Q1
1990Q1
2001Q1
1999Q2
1994Q1
1990Q1
1997Q1
1991Q2
2002Q2
2000Q4
1999Q1
1997Q1
1990Q1
1990Q1
1990Q1
2001Q1
1993Q3
2000Q4
2002Q2
1990Q1
1990Q1
1990Q1
1990Q1
1990Q1
End
2011Q1
2011Q1
2011Q1
2011Q1
2011Q1
2011Q1
2011Q1
2011Q1
2011Q1
2011Q1
2011Q1
2011Q1
2011Q1
2011Q1
2011Q1
2011Q1
2011Q1
2011Q1
2011Q1
2011Q1
2011Q1
2011Q1
2011Q1
2011Q1
2011Q1
2011Q1
2011Q1
2011Q1
2011Q1
2011Q1
2011Q1
2011Q1
19