Research Reports

Kansantaloustieteen tutkimuksia, No. 122:2010

Dissertationes Oeconomicae

HENRI NYBERG

STUDIES ON BINARY TIME SERIES MODELS WITH

APPLICATIONS TO EMPIRICAL
MACROECONOMICS AND FINANCE

ISBN: 978-952-10-5352-8 (nid.)

ISBN: 978-952-10-5353-5 (pdf)
ISSN: 0357-3257
Contents

Acknowledgements . . . . . . . . . . . . . . . . . . . . . . . . . . . . . v

1 Introduction 1
1.1 Background . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.2 Binary and Qualitative Time Series Models . . . . . . . . . . . . . . 3
1.2.1 Static Univariate Probit Model . . . . . . . . . . . . . . . . 3
1.2.2 Extensions to the Static Probit Model . . . . . . . . . . . . 5
1.2.3 Latent Variable Approach . . . . . . . . . . . . . . . . . . . 9
1.2.4 Bivariate and Multivariate Models . . . . . . . . . . . . . . . 10
1.2.5 Qualitative Models . . . . . . . . . . . . . . . . . . . . . . . 11
1.3 Applications . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
1.3.1 Forecasting Business Cycle Recession Periods . . . . . . . . 13
1.3.2 Predicting Growth Rate Cycles . . . . . . . . . . . . . . . . 19
1.3.3 Directional Predictability in Stock Market Returns . . . . . 22
1.4 Contributions of the Thesis . . . . . . . . . . . . . . . . . . . . . . 24
1.4.1 Recession Forecasts for the U.S. and Germany . . . . . . . . 24
1.4.2 LM Test for the Autoregressive Structure . . . . . . . . . . . 26
1.4.3 Sign Predictability of the U.S. Stock Returns . . . . . . . . . 27
1.4.4 Nowcasting Cycles in the U.S. Economic Activity with Bi-
variate Autoregressive Probit Model . . . . . . . . . . . . . . 28
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30

2 Dynamic Probit Models and Financial Variables in Recession Fore-

casting 39

i
 2.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 40
2.2 Dynamic Probit Models . . . . . . . . . . . . . . . . . . . . . . . . 41
2.2.1 Models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 41
2.2.2 Forecasts for the Recession Indicator . . . . . . . . . . . . . 44
2.3 Empirical Analysis of Recession Periods in the U.S. and Germany . 45
2.3.1 Data and Predictive Variables . . . . . . . . . . . . . . . . . 45
2.3.2 In-Sample Results and Model Selection . . . . . . . . . . . . 46
2.3.3 Out-of-Sample Forecasting Results . . . . . . . . . . . . . . 52
2.3.4 Recession Probabilities in 2006–2008 . . . . . . . . . . . . . 58
2.4 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 60
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 61

3 Testing an Autoregressive Structure in Binary Time Series Mod-

els 65
3.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 66
3.2 Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 66
3.3 LM Tests . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 68
3.4 Simulation Results . . . . . . . . . . . . . . . . . . . . . . . . . . . 70
3.5 Application: U.S. Recession Forecasting Models . . . . . . . . . . . 76
3.6 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 79
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 79

4 Forecasting the Direction of the U.S. Stock Market with Dynamic

Binary Probit Models 81
4.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 82
4.2 Forecasting Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . 84
4.2.1 Dynamic Probit Models in Directional Forecasting . . . . . . 84
4.2.2 An Error Correction Model . . . . . . . . . . . . . . . . . . 86
4.2.3 Recession Forecast as an Explanatory Variable . . . . . . . . 88
4.3 Evaluation of Forecasts . . . . . . . . . . . . . . . . . . . . . . . . . 91
4.3.1 Statistical and Economic Goodness-of-Fit Measures . . . . . 91

ii
 4.3.2 Testing the Statistical Predictability . . . . . . . . . . . . . 93
4.4 Empirical Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . 95
4.4.1 Data and Previous Findings . . . . . . . . . . . . . . . . . . 95
4.4.2 In-Sample Results . . . . . . . . . . . . . . . . . . . . . . . . 97
4.4.3 Out-of-Sample Results . . . . . . . . . . . . . . . . . . . . . 103
4.4.4 Comparison Between Probit and Alternative Predictive Models107
4.4.5 Sign Predictability of Small and Large Size Firms’ Returns . 110
4.5 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 113
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 114

5 A Bivariate Autoregressive Probit Model: Predicting U.S. Busi-

ness Cycle and Growth Rate Cycle Recessions 119
5.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 120
5.2 Bivariate Autoregressive Probit Model . . . . . . . . . . . . . . . . 122
5.3 Parameter Estimation, Testing and Forecasting . . . . . . . . . . . 125
5.3.1 Maximum Likelihood Estimation . . . . . . . . . . . . . . . 125
5.3.2 LM Test for the Correlation Coefficient . . . . . . . . . . . . 128
5.3.3 Forecasting . . . . . . . . . . . . . . . . . . . . . . . . . . . 130
5.4 Empirical Application: Predicting the Current State of the U.S.
Economy . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 132
5.4.1 Binary Indicators for the Business and Growth Rate Cycles . 132
5.4.2 Data Set and Predictive Models . . . . . . . . . . . . . . . . 135
5.4.3 Model Selection and In-Sample Results . . . . . . . . . . . . 137
5.4.4 Out-of-Sample Performance . . . . . . . . . . . . . . . . . . 145
5.4.5 Predictions for 2006–2008 . . . . . . . . . . . . . . . . . . . 147
5.5 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 150
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 150

iii
Acknowledgements

It is a pleasure to thank the many people who made this thesis possible.

It is difficult to overstate my gratitude to my thesis supervisors, Professor

Markku Lanne and Professor Pentti Saikkonen, for excellent supervising and sup-
port throughout my postgraduate studies. They have supervised this thesis pa-
tiently and given important advice during this work. Their encouragement and
guidance to the academic world have provided the best possible basis for this thesis.

I wish to thank the official examiners of my thesis, Professor Denise Osborn

and Professor Markku Rahiala, for many insightful comments and interesting sug-
gestions for future research. I am also grateful to Professor Heikki Kauppi and
Dr. Pekka Pere for examining my Licenciate thesis in 2007–2008. Their comments
have been valuable when writing this thesis. I have also greatly benefited numer-
ous discussions and comments to my presentations in seminars and conferences
organized in Helsinki, elsewhere in Finland and abroad.

I had a great privilege of writing this thesis while working as a full time doctoral
student at the Department of Economics (2007–2009) and at the Department of
Political and Economic Studies and its Economics discipline (from January 2010) in
the University of Helsinki. Most of that time I have worked in the research project
of “Econometrics of Macroeconomics and Finance, and the Interface between the
Macroeconomy and Financial Markets” financed by the Academy of Finland. I
am very grateful to all people who gave me these excellent research facilities. I
also wish to thank the staff and my colleagues at the Economics discipline for a
friendly and inspirational working environment.

The financial support from the Academy of Finland, the Research Foundation

v
of the University of Helsinki, the Okobank Group Research Foundation and the
Finnish Foundation for Advancement of Securities Markets is gratefully acknow-
ledged. Their support have made it possible to concentrate on research and attend
interesting international conferences, such as the Econometric Society meetings in
Milan (2008) and Canberra (2009) as well as conferences in Luxembourg (2008),
Manchester (2009) and Lund (2009), which have been useful when writing this
thesis.
Finally, I would like to thank my family for support and encouragement in my
studies ever since I started at the university.

Helsinki, March 2010

Henri Nyberg

vi
Chapter 1

Introduction

1.1 Background
In recent decades, nonlinear time series models have attracted serious attention
in the econometric literature. Various nonlinear models have been proposed to
describe potential nonlinear characteristics in the underlying data generation pro-
cess. Threshold (TAR) models (see, e.g., Tong, 1990), smooth transition (STAR)
models (see, e.g., Granger and Teräsvirta, 1993), Markov switching models (Hamil-
ton, 1989), as well as ARCH (Engle, 1982) and GARCH models (Bollerslev, 1986)
are examples of nonlinear time series models that have been applied in numerous
economic applications (see also, for example, the book by Franses and van Dijk,
2000).
In all of the above-mentioned models the dependent variable is “continuous”
indicating that its observed value can, in principle, be any real number. However,
the dependent variable can also be qualitative with only a limited number of
possible outcomes. This is the case, for example, if the dependent variable is
binary or a count variable. Models of this type are referred to as qualitative
response models in the econometric literature. Qualitative time series models can
also be seen as generalizations of the generalized linear models designed for cross-
sectional data considered extensively in the previous literature in statistics.
The objective of this thesis is to consider time series models with a binary

1
dependent variable. There are various potential empirical applications of these
models. In this thesis, we are interested in applications to empirical macroeco-
nomics and finance. For example, forecasting the recession periods of the economy
or the signs of stock market returns are of interest for many economic decision
makers. Policymakers, such as central banks and governments are interested in
predicting the probability of a recession in the future or assessing the probability
that the economy is already in recession. Correspondingly, financial investors can
benefit from probability forecasts of the future developments in financial markets
in making their investment decisions. For instance, binary time series models can
be used to evaluate the probability that the stock market return is positive in the
next period.
In this thesis, the main interest is on different dynamic extensions of the tradi-
tional “static” univariate probit model commonly used in the previous literature.
The static model can be extended in various ways. Models with the so-called au-
toregressive model structure (see Kauppi and Saikkonen, 2008) are of particular
interest throughout the thesis.
The rest of this introductory chapter consists of a survey of binary time series
models and their economic applications. In addition, summaries of the contents
of the four studies to be presented in Chapters 2–5 are provided. In Section 1.2,
different binary time series models are discussed. Although the main interest
is on different dynamic extensions to the static univariate probit model, a few
closely related models, such as multinomial models and count data models, are also
briefly reviewed. Section 1.3 provides an introduction to the empirical applications
considered in Chapters 2–5. As in much of the previous literature, forecasting the
recession periods of the economy is the main empirical application examined in
the thesis, especially in Chapter 2. Finally, Section 1.4 gives a short summary of
the main results and contributions of the thesis.

2
1.2 Binary and Qualitative Time Series Models
Many economic applications are concerned with variables whose range is discrete or
limited. Time series of discrete directions of change constructed from observed con-
tinuous variables are also often studied for reasons of forecasting. For instance, in
empirical finance, the directional predictability of financial asset return is suggested
to be more closely related to the profitability than other statistical goodness-of-fit
measures (see, e.g., Leitch and Tanner, 1991).
Qualitative response models, also known as discrete or categorical models, have
been used in various cross-sectional and panel data applications. Amemiya (1980),
Maddala (1983), Gourieroux (2000) and Wooldridge (2002, chapters 15 and 19)
provide comprehensive introductions to these models from the viewpoint of econo-
metric applications. In applied microeconometrics, the qualitative dependent mod-
els for cross-sectional and panel data are predominantly employed, as emphasized
by Cameron and Trivedi (2005, chapter IV). In the context of panel data Hon-
ore and Kyriazidou (2000) and Honore and Lewbel (2002), among others, have
proposed various dynamic models for binary dependent variables.
The models for cross-sectional and panel data mentioned above are closely re-
lated to the binary time series models considered in this thesis. In statistics, similar
models have been studied under the heading of generalized linear models (see, e.g.,
McCullagh and Nelder, 1989). Li (1994), Shephard (1995) and Benjamin, Rigby
and Stasinopoulos (2003), among others, have considered generalized linear mod-
els designed for time series. Overall, the literature on dynamic models for binary
time series is scant, but a number of new models have recently been suggested.
Some of these models are surveyed in more detail in the next section.

1.2.1 Static Univariate Probit Model

The simplest example of a qualitative time series model is a model where the
dependent variable is binary. For simplicity, and without loss of generality, let 0
and 1 denote the values of the dependent variable. Typically, the value 1 indicates
that some event occurs, and 0 that it does not occur. Now, let yt , t = 1, 2, ..., T , be

3
a univariate binary time series. We denote by Ωt−1 the information set containing,
for example, lagged values of yt and explanatory variables. Conditional on the
information set Ωt−1 , yt follows a Bernoulli distribution

yt |Ωt−1 ∼ B(pt ), (1.1)

where, according to the properties of Bernoulli distribution,

pt = Et−1 (yt ) = Pt−1 (yt = 1). (1.2)

In this expression, Et−1 (·) and Pt−1 (·) signify the conditional expectation and the
conditional probability given the information set Ωt−1 , respectively.
A probit model is obtained by specifying the conditional probability pt in (1.2)
as
pt = Φ(πt ), (1.3)

where Φ(·) is the standard normal cumulative distribution function and πt is a

linear function of variables included in the information set Ωt−1 . A logit model,
where the function Φ(·) is replaced by the logistic function, is an alternative speci-
fication.1 In empirical applications, results obtained with logit and probit mod-
els usually yield very similar results (see, e.g., Maddala, 1983, 23; Davidson and
MacKinnon, 1993, 516). In this thesis, we concentrate on probit models.
In a “static” probit model, the variable πt is specified as
′
πt = ω + xt−1 β, (1.4)
′
where ω is a constant term and xt−1 contains the explanatory variables. This
specification has mainly been employed in the previous literature.2 It is “static” in
the sense that explanatory variables have an immediate effect on the conditional
probability pt which does not change unless values of the explanatory variables
change.
1
In this context, a classical linear model designed for continuous dependent variables is
referred to the linear probability model (see, e.g., Maddala 1983, 15–16). However, it has vari-
ous weaknesses (see Gourieroux 2000, 6–8). For example, in this binary case, the conditional
probability pt in (1.2) is not necessarily between zero and one.
2
Note that in Chapter 3 we denote πt (θ), where θ is the vector of parameters.

4
1.2.2 Extensions to the Static Probit Model

A limitation of the static model specification (1.4) is that it does not allow for the
potential autocorrelation in yt . Therefore, various dynamic extensions have been
proposed, and we review some of them in this section.
An obvious dynamic extension of the static specification (1.4) is obtained by
augmenting it with a lagged value of yt . Cox (1981), Zeger and Qaqish (1988),
Shephard (1995), and Dueker (1997), among others, have considered the “dynamic”
model
′
πt = ω + δ1 yt−1 + xt−1 β, (1.5)

where a positive value of δ1 implies that yt tends to take the same value respectively.
This kind of “clustering effect” appears quite plausible in many applications. In
model (1.5), only one lagged value of yt is included, but the model can be made
more general by increasing the number of lags (see, e.g., Cox, 1981; Kaufmann,
1987).
In this thesis, the main interest is on probit models with an autoregressive
structure. Following Kauppi and Saikkonen (2008), consider the model

′
πt = ω + α1 πt−1 + δ1 yt−1 + xt−1 β, (1.6)

where the condition |α1 | < 1 is assumed. The inclusion of the lagged value πt−1
on the right hand side yields a first-order autoregressive structure for πt . There-
fore, Kauppi and Saikkonen (2008) called model (1.6) the “dynamic autoregressive
model”.3 On the other hand, Anatolyev (2009) refers to this model as the “gener-
alized autoregressive model”. This name is motivated by the analogy of the model
to the GARCH model (Bollerslev, 1986) and the ACD model (Engle and Russell,
1998) used for continuous dependent variables.4 For simplicity, only the first lag
of πt is included in (1.6), but it is also possible to have several lags of πt in the
model.
3
Throughout this thesis we use the same terminology as in Kauppi and Saikkonen (2008).
4
A model somewhat similar to (1.6) has been proposed by Russell and Engle (2005) for a
multinomial dependent variable. However, these authors replace πt−1 by pt−1 = Φ(πt−1 ) on the
right hand side of (1.6).

5
 Using recursive substitution and the assumption |α1 | < 1 in model (1.6) we
obtain the equation
∞
X ∞
X ∞
X ′
πt = α1i−1 ω + δ1 α1i−1 yt−i + α1i−1 xt−i β, (1.7)
i=1 i=1 i=1

which shows that πt can be expressed in terms of the infinite history of yt and the
explanatory variables. This suggests that a model with an autoregressive structure
(i.e. a model including πt−1 ) may be a parsimonious alternative in cases where a
large number of lagged values of yt and the explanatory variables appear necessary.
A special case of model (1.6) is obtained by restricting the coefficient δ1 to zero.
This restriction yields the “autoregressive” model,

′
πt = ω + α1 πt−1 + xt−1 β, (1.8)

where |α1 | < 1 and the predictive power comes completely from the infinite history
of the explanatory variables (cf. (1.7)).
In the previous literature, the so-called “generalized linear autoregressive” (GLAR)
models closely related to model (1.6) have also been considered. Shephard (1995)
formulated a class of GLAR models which also contains a moving average term.
These models are referred to as GLARMA models.5 Further, Rydberg and Shep-
hard (2003) proposed the model

′
πt = ω + xt−1 β + gt , (1.9)

where
gt = a1 gt−1 + δ1 yt−1 .

It is also possible to augment the expression of gt by including a moving average

term based on the difference between yt−1 and pt−1 = Φ(πt−1 ) in the model. Note
that Kauppi (2008) suggested the following alternative to (1.9)

πt = ω + δ1 yt−1 + gt , (1.10)
5
Benjamin et al. (2003) derive very closely related “GARMA” (Generalized Autoregressive
Moving Average) model. See also the multinomial models by Russell and Engle (2005) and
Liesenfeld, Nolte and Pohlmeier (2006).

6
where
′
gt = a1 gt−1 + xt−1 β.

Compared with model (1.6), in models (1.9) and (1.10), the autoregression is
related to the expression for gt instead of πt .
Model (1.6) and its special cases discussed above can be estimated by the
method of maximum likelihood (ML).6 Recently, de Jong and Woutersen (in press)
have shown that, under appropriate regularity conditions, the conventional large
sample theory of ML estimation applies to the dynamic probit model (1.5). Ex-
tending this theory to models with autoregressive structure, such as (1.6) and
(1.8), still remains to be done.
An important advantage of model (1.6) and its special cases is that one-period
and multiperiod forecasts can be computed by explicit formulae (see Kauppi and
Saikkonen, 2008). This is in contrast to many other nonlinear time series models
and alternative dynamic extensions of the static probit model (1.4) to be discussed
in Section 1.2.3. This is a very useful property, especially in this thesis, because in
our empirical applications forecasting binary time series is of interest (see Chapters
2, 4 and 5).
To illustrate the properties of the introduced probit models, we simulate real-
izations from the dynamic autoregressive model (1.6) and its special cases. In the
general case, we use the process

πt = 0.1 + 0.70πt−1 + 0.5yt−1 − 0.20xt−1 , (1.11)

where the simulated explanatory variable xt follows the AR(1) process

xt = 0.1 + 0.95xt−1 + ηt , ηt ∼ NID(0, 1). (1.12)

The special cases of model (1.6) (i.e. models (1.4), (1.5) and (1.8)), are obtained
by imposing zero restrictions on the coefficients in (1.11). For example, in the

6
Note that alternative estimation methods have also been suggested for binary time series
models (see, e.g., Manski 1975, 1985; Elliott and Lieli, 2007).

7
 static model (1.4), πt is generated from

πt = 0.1 − 0.20xt−1 .

model (1.4) model (1.5)

0.9 1

0.8 0.9

0.8
0.7

0.7
0.6
0.6
0.5
t

t
0.5
P

P
0.4
0.4
0.3
0.3

0.2
0.2

0.1 0.1

0 0
0 100 200 300 400 500 0 100 200 300 400 500

model (1.6) model (1.8)

1 1

0.9 0.9

0.8 0.8

0.7 0.7

0.6 0.6
t

0.5 0.5
P

0.4 0.4

0.3 0.3

0.2 0.2

0.1 0.1

0 0
0 100 200 300 400 500 0 100 200 300 400 500

Figure 1.1: Simulated realizations of length 500 from different probit models. Con-
ditional probability pt (line) and the realized values of yt (see (1.1)) with the shaded
areas corresponding to yt = 1 are depicted.

Figure 1.1 shows the fact that the conditional probability pt is much more per-
sistent in models with an autoregressive structure (models (1.6) and (1.8)). There
seem to be periods characterized by a higher conditional probability with yt typi-
cally taking the value 1, and similarly periods characterized by a lower conditional
probability where yt = 0. Thus, the realized values of the corresponding series are
clustered in such a way that the outcome (yt = 1 or yt = 0) is typically followed
by the same outcome also in the next period. Periods of economic recession and

8
expansion, for example, follow a similar clustered pattern (see Chapters 2 and 5)
which suggests that these models may be able to capture business cycle variation.

1.2.3 Latent Variable Approach

An alternative to the way the probit model was specified above is provided by latent
variable approach (see, e.g., Davidson and MacKinnon, 1993, 514–515). In this
approach, it is assumed that the outcome yt = 1 is obtained when an unobserved
variable yt∗ takes a positive value. In other words,

 1, if y ∗ > 0,
t
yt = (1.13)
 0, if y ∗ ≤ 0.
t

The static probit model (1.4) can now be based on the equation

′
yt∗ = c + xt−1 b + ut , ut ∼ NID(0, 1), (1.14)

where the normality assumption yields the probit model. As in Section 1.2.1, the
conditional probability of the outcome yt = 1 is (cf. (1.2) and (1.4))

′
Pt−1 (yt = 1) = Pt−1 (yt∗ > 0) = Φ(c + xt−1 b).

According to the classification of Cox (1981) (see also, e.g., Benjamin et al., 2003),
model (1.6) belongs to the class of “observation-driven” models, whereas the model
based on the latent variable approach can be seen as a “parameter-driven” model.
Within the latent variable approach, various dynamic extensions of the static
probit model have been suggested. Poirier and Ruud (1988) consider a model
where the error term in (1.14) is assumed to follow a stationary first-order autore-
gressive process (see also Gourieroux, 2000, 35–36). Chauvet and Potter (2005)
have proposed a dynamic model in which the latent variable yt∗ is generated by

′
yt∗ = c + φyt−1
∗
+ xt−1 b + σt ut , (1.15)

where the latent variable yt∗ follows a first-order autoregression and ut is as in

(1.14) but the variance of yt∗ may be time varying via the variable σt . Compared
to the dynamic autoregressive model (1.6), model (1.15) is computationally more

9
demanding. Unlike in model (1.6), the application of maximum likelihood methods
requires high-dimensional multiple integrations over the unobserved lagged latent
variable yt−1
∗
and, therefore, Bayesian methods, such as the Gibbs sampler, have
been used to evaluate the likelihood function. Note also that forecasts from model
(1.6) can be computed by explicit formulae, which is not the case in model (1.15).
Kauppi (2008) has made some comparisons between the two alternative ap-
proaches to specify the probit model. He argues that although the latent variable
yt∗ can have a meaningful economic interpretation, it is rather difficult to see what
kind of dynamic properties of yt are implied by a model such as (1.15). For in-
stance, as the lagged value of yt is not included in (1.15), it is not very easy to see
how lagged values of yt drive the conditional probability (1.2).
All in all, further research is needed to compare different dynamic probit mod-
els and their properties in these two approaches. As the applications of this thesis
are based on the observation-driven models introduced in Section 1.2.2, such com-
parisons are left for future research.

1.2.4 Bivariate and Multivariate Models

So far, we have considered univariate binary time series models. In practice, we

are often interested in joint dynamic behavior and interrelationships of several
variables. Generalizing univariate binary time series models to the bivariate and
multivariate cases is therefore of interest.
Bivariate and multivariate binary time series models have not much been ex-
amined in the previous literature. Ashford and Sowden (1970) have proposed the
static bivariate probit model, where the joint conditional probabilities of the four
possible outcomes of the random vector (y1t ,y2t ) (cf. the univariate model in

10
(1.1)–(1.2)) are obtained as

P11,t = Pt−1 (y1t = 1, y2t = 1) = Φ2 (π1t , π2t , ρ),

P10,t = Pt−1 (y1t = 1, y2t = 0) = Φ2 (π1t , −π2t , −ρ),

P01,t = Pt−1 (y1t = 0, y2t = 1) = Φ2 (−π1t , π2t , −ρ),

P00,t = Pt−1 (y1t = 0, y2t = 0) = Φ2 (−π1t , −π2t , ρ). (1.16)

Here Pij denotes the conditional probability of the outcome of vector (y1t = i, y2t =
j), i, j = 0, 1, and Φ2 (·) is the bivariate standard normal cumulative distribution
function. As in the univariate case, π1t and π2t are linear functions of variables in
the information set Ωt−1 . Specifically, in the static bivariate model we have
      
′
π1t ω1 x1,t−1 0 β1
 = +
′
 , (1.17)
π2t ω2 0 x2,t−1 β2
′ ′
where x1,t−1 and x2,t−1 are vectors of explanatory variables.
To the best of our knowledge, only two bivariate and multivariate dynamic
probit models have been proposed in the literature. Mosconi and Seri (2006)
have suggested a dynamic extension to the static bivariate model defined in (1.16)
and (1.17). Their model is based on the latent variable approach discussed in
the univariate case in Section 1.2.3. Quite recently, Anatolyev (2009) has consid-
ered dynamic extension of the static multivariate models of Ekholm, Smith and
McDonald (1995, 2000). Anatolyev’s bivariate model is somewhat similar to the
model proposed in Chapter 5, which can also be seen as an extension of the static
bivariate model introduced above.

1.2.5 Qualitative Models

Although the main interest in this thesis is on binary time series models, it may
also be useful to review other related qualitative time series models. A natural ex-
tension of binary models is a model, where the dependent variable is multinomial.
In these models, there are at least three possible values that the dependent vari-
able can take. Early work in this area includes Eichengreen, Watson and Grossman

11
(1985), Hausman, Lo and MacKinlay (1992) and Dueker (1999a). Recently, Russell
and Engle (1998, 2005) and Kauppi (2007) have suggested new dynamic multino-
mial models with autoregressive dynamics of the same type as in the univariate
model (1.6).
As discussed by Kauppi (2007), in economics multinomial models have typically
been used to predict interest rate target changes made by central banks, especially
the U.S. Federal funds rate target changes made by the Federal Reserve (see also
the models proposed by Dueker (1999b), Hamilton and Jorda (2002), and Hu and
Phillips (2004)). Predicting the direction of the financial asset returns is another
example of potential application of multinomial models (see Liesenfeld et al., 2006).
Models for count variables is another important class of qualitative time series
models. In these models, the dependent variable is typically assumed to follow a
conditional Poisson distribution leading to the Poisson regression model (see, e.g.,
Maddala, 1983, 51–54). The classification into observation-driven and parameter-
driven models discussed in Section 1.2.3 is also applicable here. Parameter-driven
models based on the latent variable have been considered by Zeger (1988), Chan
and Ledolter (1995) and Davis, Dunsmuir and Wang (2000), whereas Zeger and
Qaqish (1988) provide an early discussion on observation-driven models (see also
Davis, Dunsmuir and Street, 2003). References to the count time series models
can also be found in the literature of GLARMA and GARMA models discussed
in Section 1.2.2 (see, e.g., Shephard, 1995; Benjamin et al. 2003). Furthermore,
related to the observation-driven models, Heinen (2003) has proposed a univariate
autoregressive conditional Poisson (ACP) model with an autoregressive structure
similar to that in the dynamic autoregressive probit model (1.6). The ACP model
can also be extended in various ways, such as to the multivariate case as in Heinen
and Rengifo (2007).
In addition to the introduced multinomial and count data models there are
also various other qualitative dependent models (see, e.g., the Qual VAR model
of Dueker (2005)) examined in the econometric literature (see also, e.g., Maddala,
1983; Gourieroux, 2000).

12
1.3 Applications
In this section, we review three applications of binary time series models. In Section
1.3.1, we consider forecasting “classical” business cycle recession periods. In Section
1.3.2, the predictability of cycles in aggregate economic activity is discussed by
considering growth rate cycles. The sign predictability of asset prices is a major
issue in empirical finance. We concentrate on this application in Section 1.3.3 with
emphasis on sign predictability of stock market returns.

1.3.1 Forecasting Business Cycle Recession Periods

Fluctuations in economic activity are a central topic in theoretical and empirical

macroeconomic research. Predicting cycles in economic activity is one of the most
challenging and also one of the most important applications in macroeconomic
forecasting.7 The importance of predictions of the macroeconomic state of the
economy is based on the fact that business cycles, and especially recession periods,
are costly. For example, central banks and governments may try to mitigate fluc-
tuations by stabilization policy. However, if the current state of the the economy
is unknown or forecasts for the future development are inaccurate, the timing of
policy actions may not be optimal, and the policy actions may even amplify further
business cycle fluctuations.
After the Great Depression in 1930s, the National Bureau of Economic Research
(NBER) developed a methodology for the empirical analysis of cycles in the U.S.
economic activity. NBER applies the definition of the business cycle, which was
originally suggested by Burns and Mitchell (1946). They emphasized that business
cycles are co-movements of several macroeconomic variables which determine the
turning points, peaks and troughs, in aggregate economic activity. Recession starts
just after the economy reaches a peak of a business cycle and ends at the trough,
7
At the time of writing this thesis in 2008–2009, the financial crisis had taken place all
around the world, and most of the industrialized countries have faced one of the most severe re-
cessions since the Great Depression in 1930s. This has once again increased interest in forecasting
recession periods.

13
and vice versa with expansionary period. Recessions and expansions are recurrent,
but not periodic. The duration of a business cycle from peak to peak or trough to
trough is at least one year, but it may be much longer, even more than ten years.
Forecasting the recession periods with binary time series models is based on a
binary-valued recession indicator. Without loss of generality, the recession indi-
cator takes the value 1 when the economy is in a recession, and 0 otherwise. As
above, a recession begins when the aggregate economic activity has reached a peak
and ends at the trough. Thus, to construct recession periods, it is first necessary
to determine the peak and trough dates for business cycle phases.
Using the definition of Burns and Mitchell (1946), the Business Cycle Dat-
ing Committee of the NBER determines the U.S. business cycle turning points,
that is, the peak and trough months. In contrast to the commonly used rule of
two consecutive quarters of decline in real GDP, the NBER defines the recession
as “a significant decline in economic activity spread across the economy, lasting
more than a few months, normally visible in real GDP, real income, employment,
industrial production, and wholesale-retail sales.” See Hall et al. (2003) for details.
The recession dating procedure applied by the NBER is informal in the sense
that it is up to the judgment of the Business Cycle Dating Committee. However,
due to the informational lags and revisions between the initial and final values of
the macroeconomic variables given in the recession definition, the announcements
of peak and trough months are available with a substantial delay. This, in turn,
leads to a delay in the publication of the recession indicator. This delay is referred
to as a “publication lag” in this thesis. The most recent announcements of peak
and trough months presented in Table 1.1 (see also the left panel of Figure 1.2)
show that the publication lag has been between six months and almost two years
in the U.S.
In the literature on recession forecasting, the NBER recession periods remain
the benchmark. At the time of writing (fall 2009), the latest business cycle turning
point in the U.S. is the peak in December 2007, whereas the last trough month is
November 2001 (see Hall et al., 2008).

14
 Table 1.1: NBER’s U.S. business cycle peak and trough chronology from 1970s.
Peak Trough Recession time Publication lag
(months) (months)
peak trough
1973 M11 1975 M03 16 – –
1980 M01 1980 M07 6 5 12
1981 M07 1982 M11 16 6 8
1990 M07 1991 M03 8 10 21
2001 M03 2001 M11 8 8 20
2007 M12 12
Note: Recessions start at the peak of a business cycle and end at the trough. Publication lag is
the time between the business cycle turning point month (peak or trough) and the month when
the Business Cycle Dating Committee of the NBER has made an announcement of the turning
point (see details in http://www.nber.org/cycles/cyclesmain.html [2 July 2009]). Note that
publication lags for the first peak and trough months are not available.

In addition to the informal “NBER approach”, there is also a large alternative

literature on dissecting business cycles and determining recession periods. Bry and
Boschan (1971) and Harding and Pagan (2002), among others, have developed for-
mal mathematical methods to determine recession and expansion periods. These
methods are typically based on the informational content of the same monthly
macroeconomic variables that NBER uses in their approach.
Another way to locate the turning points is based on the use of detrending
methods. The idea is to filter out the cyclical component of the output series (e.g.
the real GDP), and then to apply a dating rule to locate the peaks and troughs.
However, in this approach, as Canova (1998) has noted, the resulting turning
points are very sensitive to the detrending method and dating rule employed.
Therefore, depending on the exact methods, the obtained recession periods can be
very different.
Overall, despite the large number of different methods available for finding the
turning points, we take the NBER recession periods as given in this thesis. One
reason for this is the fact that the Economic Cycle Research Institute (ECRI) uses
the same type of recession definition as the NBER and in Chapter 2 we use the

15
recession periods they provide for Germany along with the U.S. recession periods.
The literature on recession forecasting based on the binary recession indicator
dates back to the beginning of the 1990s. To the best of our knowledge, the
first authors in this area were Estrella and Hardouvelis (1991) who used the static
probit model (1.4) to predict U.S. recession periods. The same model with different
predictive variables was subsequently employed in various other studies (see, e.g.,
Estrella and Mishkin, 1998; Bernard and Gerlach, 1998; Estrella, Rodrigues and
Schich, 2003, among others). In recent years the dynamic models discussed in
Sections 1.2.2 and 1.2.3 have been employed (see, e.g., Chauvet and Potter, 2005;
Dueker, 2005; Kauppi and Saikkonen, 2008; Startz, 2008).
Variables used to predict future economic activity are often referred to as “lead-
ing indicators”. In fact, there is a branch in the literature in empirical macroe-
conomics focused on finding and constructing reliable leading indicators for busi-
ness cycle fluctuations. Marcellino (2006) provides a comprehensive survey of this
literature, especially from the viewpoint of constructing coincident and leading
indices to economic activity (see also, e.g., Stock and Watson, 1989; Mariano and
Murasawa, 2003; Aruoba, Diebold and Scotti, 2009).
Marcellino (2006) discusses properties that a useful leading indicator variable
should have. First, a leading indicator should systematically anticipate recessions
and expansions. Second, the potential predictive power of a leading indicator
should be supported by economic theory. Third, values of a leading indicator
should be regularly available without (major) subsequent revisions. The real time
availability of financial variables supports the use of those variables as predictors
in recession forecasting. Financial variables, such as interest rates and stock mar-
ket returns, are available even at a very high frequency, and they are measured
precisely without revisions.8 This is not typically the case for most macroeconomic
predictors.
Much of the previous research on the use of financial variables as predictors
in recession forecasting lends support to the term spread between the long-term
8
Present value models for the term structure of interest rates and the stock price provide
theoretical basis for the potential predictive power of these variables.

16
and short-term interest rate being the main leading indicator (see, e.g., Estrella
and Hardouvelis, 1991; Estrella and Mishkin, 1998; Bernard and Gerlach, 1998;
Estrella, 2005). The usefulness of the term spread as a predictor in recession fore-
casting is illustrated in Figure 1.2, where the U.S. and German recession periods
and term spreads are depicted. The term spread is constructed as the difference
between the 10-year and three-month interest rates. It is seen that positive values
close to zero or even negative values of the term spread have preceded recessions in
both countries. This was also the case before the recent recession period that be-
gan in 2008. These preliminary findings suggest that the term spread is a reliable
leading indicator of future recession periods.

5 6

4
4

2
2
SPUS

SPGE

1 0

0
−2

−1

−4
−2

−3 −6
1970 1975 1980 1985 1990 1995 2000 2005 2010 1970 1975 1980 1985 1990 1995 2000 2005 2010
Time Time

Figure 1.2: In the left panel, the U.S. recession periods (shaded areas) and the
U.S. term spread (SPtU S ) from January 1971 to January 2009 are depicted. The
German recession periods and the term spread (SPtGE ) are shown in the right
panel.

As an example of the use of univariate probit models in recession forecasting,

we consider six-month-ahead recession forecasts obtained from the static (1.4) and
the autoregressive model (1.8). In both models, the domestic term spread depicted
in Figure 1.2 is employed as a predictor of the U.S. and German recession periods.
Figure 1.3 shows the estimated recession probabilities. Intuitively, the recession
probability should be as close as possible to one in recession and close to zero in
expansion.
In Figure 1.3, we see that the autoregressive model (1.8) leads to more persis-

17
tent recession forecast compared with the static model (1.4). Based on the pseudo-
R2 goodness-of-fit measure of Estrella (1998), which can be seen as a counterpart to
the coefficient of determination in models for continuous variables, the autoregres-
sive model outperforms the static model in both countries. Especially for the U.S.
(German) recession periods, the autoregressive model yields a higher pseudo-R2
value 0.391 (0.571) compared with the static model 0.181 (0.533). Overall, these
results lead to the conclusion that the recession periods seem to be predictable
with the predictive power provided by the term spread.
model (1.4) model (1.8)
0.9 1

0.8 0.9

0.8
0.7

0.7
0.6
0.6
Probability

Probability

0.5
0.5
0.4
0.4
0.3
0.3

0.2
0.2

0.1 0.1

0 0
1970 1975 1980 1985 1990 1995 2000 2005 2010 1970 1975 1980 1985 1990 1995 2000 2005 2010
Time Time

model (1.4) model (1.8)

1 1

0.9 0.9

0.8 0.8

0.7 0.7

0.6 0.6
Probability

Probability

0.5 0.5

0.4 0.4

0.3 0.3

0.2 0.2

0.1 0.1

0 0
1970 1975 1980 1985 1990 1995 2000 2005 2010 1970 1975 1980 1985 1990 1995 2000 2005 2010
Time Time

Figure 1.3: Estimated recession probabilities based on the static model (1.4) and
the autoregressive model (1.8) where the term spread is employed as a predictor.
In the upper panel, the U.S. recession periods are considered whereas in the lower
panel the German recession periods are examined. The sample period is from
January 1971 to January 2009 where the first 12 observations are used as initial
values.

18
 As emphasized by Marcellino (2006), although the term spread seems to a
useful recession predictor, a model with a single predictor may yield unreliable
forecasts for future recession periods. Experience has taught that each recession
period has its own characteristics that a single explanatory variable may not be
able to explain. Thus, in addition to the term spread, it is useful to consider also
other financial variables as predictors in recession forecasting (see, e.g., the survey
of Stock and Watson, 2003). This issue is one of the main objectives of Chapter 2.

1.3.2 Predicting Growth Rate Cycles

In recession forecasting, values of the recession indicator are related to the level
of aggregate economic activity such that in a business cycle recession the level of
economic activity is decreasing. However, at times the economy faces a slowdown
that is not a decline in the level, but in the growth rate of economic activity. Such
periods are called growth rate recessions, and this predictability can be studied by
models similar to those discussed above.
Economic Cycle Research Institure (ECRI), among others, has determined a
chronology of the growth rate cycle peak and trough months. The dating proce-
dure is based on several macroeconomic variables and, in that sense, it is similar
to the approach that the NBER uses to define business cycle turning points, as
discussed in the previous section.9 Layton and Moore (1989) and Banerji and Hiris
(2001) have described in detail how the growth rate cycles can be determined. In
summary, the methodology is based on the six-month smoothed growth rates of
the macroeconomic variables included in the analysis. These smoothed growth
rates together determine the cyclical turning points, i.e., peaks and troughs, in
the growth rate cycles and hence also the growth rate “recession” and “expansion”
periods.
Note that a distinction is made between growth cycles and growth rate cycles
in the literature. According to Layton and Moore (1989) and Banerji and Hiris
(2001) the difference between these two is that while growth cycles are deviations
9
See details in http://www.businesscycle.com/resources/cycles [2 July 2009].

19
from the trend of economic activity, growth rate cycles are based on the suggested
six-month or 12-month smoothed growth rate of economic activity. Thus, in the
case of growth rate cycles, detrending is avoided, whereas it is inevitable in the
case of growth cycles.10
In this thesis, we restrict ourselves to the ECRI growth rate cycles. To the best
of our knowledge, Osborn, Sensier and van Dijk (2004) is the only study where
growth rate cycle periods are predicted by using binary time series models. They
used the static probit model (1.4) to predict ECRI growth rate cycle periods in
a number of European countries. One of their main finding was that the growth
rate cycles are closely related to the estimated output gap, which is the difference
between the actual and potential output.
Although business cycle recessions and expansions have been studied much
more than growth rate cycle periods, the latter are of importance for various rea-
sons. As illustrated in Section 1.3.1, business cycle recessions are quite infrequent
and, therefore, growth rate cycles give us information about the cyclical variation
observed during business cycle expansions. In Chapter 5, we see that there is much
more regime switching between recessions and expansions defined in terms of the
growth rate cycle than those defined in terms of business cycle. This can be seen
in Figure 1.4 where the values of both cycle indicators are illustrated with the
business conditions index of Aruoba, Diebold and Scotti (2009) (hereafter ADS
index).11 This ADS index is an example of a coincident economic index discussed
in Section 1.3.1 which reflects overall economic activity and is based on various
macroeconomic variables measured at different data frequencies. The average value
of the ADS index is set to zero. Thus, positive values indicate progressively better-
than-average economic conditions, whereas negative values indicate progressively
worse-than-average conditions.
In the left panel of Figure 1.4, we see that the ADS index is substantially
10
Zarnowitz and Ozyildirim (2006) and Marcellino (2006), among others, have surveyed meth-
ods used to determine and characterize business cycles, growth rate cycles and growth cycles.
11
See details at http://www.phil.frb.org/research-and-data/real-time-center/business-
conditions-index [15 March 2010].

20
negative at business cycle recession periods whereas growth rate cycle recessions
are closely related to the upswing and downswing periods in the growth rate of the
index. A growth rate cycle recession is typically occurred at the same time with
the decreasing values of the ADS index, and vice versa with expansion periods.
NBER business cycles ECRI growth rate cycles
3 3

2 2

1 1

0 0
ADS

ADS
−1 −1

−2 −2

−3 −3

−4 −4

−5 −5
1970 1975 1980 1985 1990 1995 2000 2005 2010 1970 1975 1980 1985 1990 1995 2000 2005 2010
Time Time

Figure 1.4: U.S. business cycle (left panel) and growth rate cycle (right panel)
periods (shaded areas are the recession periods) with the real business conditions
index of Aruoba, Diebold and Scotti (2009). The sample period is from January
1971 until January 2009.

As argued by Osborn et al. (2004), the relationship between growth rate reces-
sions and the estimated output gap makes growth rate cycles of interest because
the output gap is an important variable to policymakers, especially those concerned
with monetary policy. Evidence of this kind can be found in Figure 1.5, where we
depict the estimated U.S. output gap and growth rate cycle recession and expan-
sion periods. The monthly output gap is approximated by the difference between
the U.S. industrial production, which is used to measure economic activity in a
monthly basis, and its trend component extracted by using the Hodrick-Prescott
filter. It appears that the growth rate cycle periods have typically been related to
the turning points of the estimated output gap. A growth rate recession is often
began after the output gap has reached its local maximum and vice versa with
expansion periods.
It should be pointed out that despite the importance of the output gap, it
is not available in real time because of the delays and revisions in real GDP or

21
industrial production typically employed to construct the output gap. Therefore,
if the growth rate cycle phases, i.e. the turning points in the estimated output
gap, are predictable, the predictions may yield important information for monetary
policy decisions.

0
Output gap

−2

−4

−6

−8
1970 1975 1980 1985 1990 1995 2000 2005 2010
Time

Figure 1.5: Estimated monthly U.S. output gap and the U.S. growth rate cycle
recession (shaded areas) and expansion periods.

1.3.3 Directional Predictability in Stock Market Returns

According to recent empirical evidence, stock market returns or, more precisely
excess stock market returns over the risk-free rate, are predictable, although the
predictability is rather weak out of sample. Macroeconomic variables included
in traditional linear regression models have been capable of predicting U.S. stock
returns (see, e.g., Chen, Roll and Ross, 1986; Fama and French, 1989; Chen, 1991;
Pesaran and Timmermann, 1995), and the recent work of Rapach, Wohar and
Rangvid (2005), Ang and Bekaert (2007), among others, suggests that interest
rates are the best predictors.
Christoffersen and Diebold (2006) and Christoffersen et al. (2007) argue that
the sign of stock market returns may be predictable even if the returns themselves

22
are not predictable. The arguments used to support this view have been based
on models with continuous dependent variables. For instance, Christoffersen et al.
(2006, 2007) have shown that the predictability of stock return volatility implies
the sign predictability of the returns.
Despite the potential sign predictability of stock returns, surprisingly few stud-
ies have considered binary time series models in this context. However, they have
been incorporated into models where the stock return is decomposed into different
components. The simplest decomposition is

rt = sign(rt )|rt |, (1.18)

where a return (rt ) consists of a sign component (sign(rt )) and an absolute value
component (|rt |). This decomposition has been employed by Anatolyev and Gospodi-
nov (2010) who specified a probit model for the sign component.
Rydberg and Shephard (2003) suggest a different decomposition where the re-
turn is decomposed into three distinct processes when analyzing the dynamics of
(high frequency) trade-by-trade price movements in asset prices. This decomposi-
tion can be presented as follows,

rt = At Dt St , (1.19)

where a binary variable At takes value one when the price moves and zero other-
wise. Conditional on At = 1, the second binary variable Dt indicates the direction-
of-change (Dt = 1 if the price moves upwards). Finally, conditional on the price
movement and the direction-of-change, the variable St gives the size of the move-
ment. The dynamics of the two binary variables can be modeled by binary time
series models. Similarly, Liesenfeld et al. (2006) incorporate two binary models
into their trinomial autoregressive conditional model (see also the model of Russell
and Engle, 1998, 2005).
Instead of using binary models, it is also possible to extract sign forecasts from
models built for the level of a stock return. Sign forecasts are obtained by the
decision rule that, say, a positive forecast is a signal for a positive stock return.
To the best of our knowledge, Leung, Daouk and Chen (2000) is the only study to

23
compare the predictive performance of continuous dependent models and different
discrete models, including probit models. They find that discrete models and
various classification-based methods outperform traditional continuous models in
terms of the number of correct sign forecasts and subsequent investment returns.
In Chapter 4, we compare sign forecasts from binary and continuous predictive
models for U.S. stock returns.
Finally, binary time series models may also be used to predict “bull” and “bear”
market phases in stock markets which can be determined in the same fashion as
recession and expansion periods of the macroeconomy (see Section 1.3.1). Maheu
and McCurdy (2000), Pagan and Sossounov (2003), and Candelon, Piplack, and
Straetmans (2008) have suggested methods to determine the bull and bear markets.
Recently, Chen (2009) considered a Markov switching model and the static probit
model (1.4) to forecast bear market phases. He concluded that the probit model
is capable of predicting the stock market phases and that the term spread and
inflation rate are the most useful predictors.

1.4 Contributions of the Thesis

In this thesis, we concentrate on different probit models based on the dynamic
autoregressive probit model (1.6) and its special cases. We also suggest new model
variants, whose most important feature is the inclusion of an autoregressive struc-
ture to the model. In this section, we introduce and summarize the main contri-
butions to be presented in Chapters 2–5.

1.4.1 Recession Forecasts for the U.S. and Germany

In Chapter 2, we explore various probit models and their ability to predict the
U.S. and German recession periods. The monthly data from January 1972 to
March 2007 consists of observations of dependent recession indicators and financial
predictive variables, such as interest rates and stock market returns.
The empirical findings show that models with an autoregressive structure out-

24
perform their competitors, especially the static model (1.4) used in many previous
studies. We also propose a new “autoregressive interaction” model

′ ′
πt = ω + α1 πt−1 + xt−1 β + yt−1 z t−1 γ, (1.20)

′
where the autoregressive model (1.8) is augmented by the interaction term yt−1 z t−1
′
in which z t−1 includes the predictors. Compared with model (1.6), yt−1 is now
′
excluded from the model, but the predictive variables included in the vector z t−1
are allowed to have an asymmetric effect depending on the lagged state of the
economy (yt−1 ).
As suggested in the previous literature, the term spread between the long-term
and short-term interest rates turns out to be the main predictor, but various other
financial variables are also found to have useful predictive power. As already sug-
gested by Estrella and Mishkin (1998) in the case of the static probit model (1.4),
stock market returns have predictive power for both countries. In addition to the
term spread and stock market returns, the foreign term spread is a statistically
significant predictor for both countries. We also find that the short-term interest
rate differential between the U.S. and Germany is a statistically significant pre-
dictor for recession periods for Germany. Our findings show that these additional
predictive variables have also useful out-of-sample predictive power over and above
the term spread for both countries.
In an out-of-sample forecasting exercise covering the period from January 1995
to March 2007, the best probit models give good forecasts for the state of the
business cycle at least six months ahead. The best out-of-sample forecasts are
obtained from the interaction model (1.20), where the domestic or the foreign
′
term spread is included in the interaction term yt−1 z t−1 .
In Chapter 2, we also report out-of-sample forecasts for the years 2006–2008.
This period is of interest given that the latest recession period in the U.S. began
in December 2007 and in April 2008 in Germany. Our results show that the best
dynamic probit models, such as model (1.20), were able to predict the beginning
of the recent recession for both countries.

25
1.4.2 LM Test for the Autoregressive Structure

The findings of this thesis lend support to the usefulness of models with an au-
toregressive structure. Therefore, it is of interest to test the statistical significance
of the autoregressive structure when specifying the model. From the practical
point of view, this is useful because parameter estimation in model (1.6) is more
complicated than, say, in model (1.5) where estimation can be carried out by the
procedures available in standard econometric software packages.
In Chapter 3, we propose a LM test for testing the adequacy of the restricted
model (1.5) with no autoregressive structure. Thus, the null hypothesis imposed
by model (1.5) is
H0 : α1 = 0. (1.21)

The rejection of the null hypothesis would provide evidence in favor of the unre-
stricted model (1.6). As a matter of fact we propose two LM test statistics for
the null hypothesis (1.21). Asymptotically these test statistics are equivalent, but
their small-sample properties seem to be different. Our simulation results show
that both LM tests can be severely oversized in small samples. As a remedy,
we propose a parametric bootstrap method which makes the finite sample size of
tests acceptable. Both tests have reasonable size-adjusted power, even in sample
as small as 150 observations.
As an illustration, we test for significance of the autoregressive model structure
in various binary models for U.S. recession periods. The data set consists of the
monthly observations of the U.S. recession indicator and explanatory variables from
January 1953 until December 2006. Thus, the sample size of 634 observations is
somewhat larger than in Chapter 2. The main findings are, however, the same as
in Chapter 2. The autoregressive structure is found to be a statistically significant
addition to the traditional static recession prediction model, possibly augmented
with the forecast horizon specific predictor yt−15 , used in many previous studies.
However, when the first lag of the recession indicator (yt−1 ) is employed in the
dynamic autoregressive model (1.6), it is the main predictor and the autoregressive
structure is in fact statistically insignificant indicating that model (1.6) reduces to

26
the dynamic model (1.5). It is worth noting that, in both cases, the proposed LM
tests yield the same conclusion as the Wald and likelihood ratio tests.

1.4.3 Sign Predictability of the U.S. Stock Returns

As discussed in Section 1.3.3, some predictability in excess stock returns over the
risk-free interest rate has been found in the previous literature. The out-of-sample
predictability has mostly been rather weak leading to the suggestion that it may
only be the sign of an excess return that is predictable.
In Chapter 4, we study the in-sample and out-of-sample sign predictability
of U.S. stock returns using different probit specifications. The data set consists
of monthly U.S. data from January 1968 to December 2006. The monthly excess
stock return is constructed as the difference between the U.S. nominal return on the
S&P500 index and the risk-free interest rate. Several financial variables included in
the data set are considered as predictors. The models are estimated with the data
up to December 1988 and the remaining sample period is left for out-of-sample
forecasts. Thus, the first out-of-sample forecasts are made for January 1989 and
last ones for December 2006.
We propose a new framework where the recession forecast (see Chapter 2) is
used as a predictor in a probit model to predict the sign of the return. Second,
alternative dynamic probit specifications are considered in sign forecasting. To the
best of our knowledge, this type of model has not been considered previously. This
model can also be interpreted as a “structural” model where a recursive structure
is imposed on two binary variables. First, we construct forecasts for the recession
periods, and thereafter compare the forecasts for the sign of the return using the
generated recession forecast as a predictor.
It turns out that the six-month recession forecast is the main predictor of the
sign of the return. This indicates that forecasts of the future state of the economy
are informative for the sign of the next month’s stock return. In addition to
recession forecasts, the differenced short-term and long-term interest rates have
statistically significant predictive power out of sample.

27
 Among the alternative probit models considered, our new “error correction”
model yields the best out-of-sample sign forecasts. This is obtained by imposing
the restriction δ1 = 1 − α1 on model (1.6),
′
πt = ω + α1 πt−1 + (1 − α1 )yt−1 + xt−1 β. (1.22)

We refer this model as the error correction model because it can be written as
error correction form
′
∆πt = ω + xt−1 β + (1 − α1 )(yt−1 − πt−1 ),

where the difference (yt−1 − πt−1 ) is interpreted as a “disequilibrium error”.

In addition to probit models, in terms of out-of-sample sign forecasting, model
(1.22) also outperforms a number of other models, such as ARMAX models and
predictive models based on volatility forecasts (cf. Christoffersen et al., 2006,
2007). These findings are in line with those of Leung et al. (2000).

1.4.4 Nowcasting Cycles in the U.S. Economic Activity with

Bivariate Autoregressive Probit Model

Chapter 5 extends the univariate analysis of binary time series considered in Chap-
ters 2–4 to the bivariate case by introducing a new bivariate autoregressive probit
model. This model is an extension of the static bivariate model of Ashford and
Sowden (1970) described briefly in Section 1.2.4. Our model is given by
         
′
π ω α α12 π x 0 β
 1t  =  1  +  11   1,t−1  +  1,t−k  1 ,
′
π2t ω2 α21 α22 π2,t−1 0 x2,t−k β2
(1.23)
where π1t and π2t , with the correlation coefficient ρ that appears in bivariate cumu-
lative normal distribution function define the joint conditional probabilities given
in (1.16) for the four possible outcomes of the vector (y1t , y2t ). The main difference
between the static bivariate probit model (1.17) and the bivariate autoregressive
model (1.23) is the absence of an autoregressive structure in the former. Model
(1.23) can also be interpreted as a bivariate extension to the univariate autore-
gressive model (1.8). Like its univariate counterpart (1.7), model (1.23) has an

28
infinite-order representation implying that π1t and π2t depend on the infinite his-
tory of the predictive variables included in x1,t−k and x2,t−k .
Before applying the bivariate autoregressive probit model it may be of interest
to test whether the correlation coefficient ρ in the bivariate cumulative normal
distribution (1.16) defining the model is zero (ρ = 0). If this is the case, the bi-
variate probit model reduces to two univariate models, which simplifies parameter
estimation. To this end, we propose a Lagrange Multiplier test for testing the
hypothesis of ρ = 0.
We apply the bivariate model to nowcasting the current state of the U.S. econ-
omy defined in terms of business cycle and growth rate cycle recession periods.
By nowcasting we mean that the forecast horizon is one month (h = 1) indicating
that we are interest in forecasts of the state of the economy at the current time.
The data set covers the U.S. data from January 1972 to December 2005. This type
of bivariate analysis of the two binary cycle indicators has not been considered in
the previous literature. It extends the univariate recession forecasting studies of
the previous literature and Chapter 2.
Empirical results show that additional predictive power can be gained by now-
casting business cycle and growth rate cycle periods jointly within a bivariate
model. The bivariate autoregressive probit model (1.23) outperforms the univari-
ate models yielding the best in-sample and out-of-sample predictions for the state
of the U.S. economy.
As in the previous literature and also in Chapter 2, the U.S. term spread is
the main predictive variable for business cycle recession periods. For growth rate
cycles, however, the lagged differenced Federal funds rate and stock market returns
are the best predictors. The U.S. term spread also has some predictive ability for
growth rate cycles, but it is less powerful than the lagged differenced Federal funds
rate and stock market returns. Furthermore, the best out-of-sample nowcasts for
growth rate cycle recessions are obtained from the bivariate autoregressive probit
model in which the effect of “Great Moderation” time period is also taken into
account.

29
References
Amemiya T. 1980. Qualitative response models: A survey. Journal of Economic
Literature 19: 1483–1536.

Anatolyev S. 2009. Multi-market direction-of-change modeling using dependence

ratios. Studies in Nonlinear Dynamics and Econometrics 13, article 5.

Anatolyev S, Gospodinov N. 2010. Modeling financial return dynamics via decom-

position. Journal of Business and Economic Statistics 28: 232–245.

Ang A, Bekaert G. 2007. Stock return predictability: Is it there? Review of Fi-

nancial Studies 20: 651–707.

Aruoba SB, Diebold FX, Scotti R. 2009. Real-time measurement of business con-
ditions. Journal of Business and Economic Statistics 27: 417–427.

Ashford JR, Sowden RR. 1970. Multivariate probit analysis. Biometrics 26: 535–
546.

Banerji A, Hiris L. 2001. A framework for measuring business cycles. International

Journal of Forecasting 17: 333–348.

Benjamin MA, Rigby RA, Stasinopoulos DM. 2003. Generalized autoregressive

moving average models. Journal of American Statistical Association 98: 214–223.

Bernard H, Gerlach S. 1998. Does the term structure predict recessions? The in-
ternational evidence. International Journal of Finance and Economics 3: 195–215.

Bollerslev T. 1986. Generalized autoregressive conditional heteroskedasticity. Jour-

nal of Econometrics 52: 5–59.

Bry G, Boschan C. 1971. Cyclical analysis of time series: Selected procedures and
computer programs. National Bureau of Economic Research, Columbia University
Press.

Burns AF, Mitchell WC. 1946. Measuring Business Cycles. National Bureau of
Economic Research, New York.

30
Cameron CA, Trivedi PK. 2005. Microeconometrics: Methods and Applications.
Cambridge University Press, New York.

Candelon B, Piplack J, Straetmans S. 2008. On measuring synchronization of bulls

and bears: The case of East Asia. Journal of Banking and Finance 32: 1022–1035.

Canova F. 1998. Does detrending matter for the determination of the reference
cycle and the selection of turning points? Economic Journal 109: 126–150.

Chan KS, Ledolter J. 1995. Monte Carlo EM estimation for time series models
involving counts. Journal of American Statistical Association 90: 242–251.

Chauvet M, Potter S. 2005. Forecasting recession using the yield curve. Journal
of Forecasting 24: 77–103.

Chen N. 1991. Financial investment opportunities and the macroeconomy. Jour-

nal of Finance 46: 529–554.

Chen N, Roll R, Ross SA. 1986. Economic forces and the stock market. Journal
of Business 56: 383–403.

Chen SS. 2009. Predicting the bear stock market: Macroeconomic variables as
leading indicators. Journal of Banking and Finance 33: 211–223.

Christoffersen PF, Diebold FX. 2006. Financial asset returns, direction-of-change

forecasting, and volatility dynamics. Management Science 52: 1273–1287.

Christoffersen PF, Diebold FX, Mariano RS, Tay AS, Tse YK. 2007. Direction-of-
change forecasts based on conditional variance, skewness and kurtosis dynamics:
International evidence. Journal of Financial Forecasting 1: 3–24.

Cox DR. 1981. Statistical analysis of time series: Some recent developments.
Scandinavian Journal of Statistics 8: 93–115.

Davidson R, MacKinnon JG. 1993. Estimation and Inference in Econometrics.

Oxford University Press, New York.

31
Davis RA, Dunsmuir WTM, Wang Y. 2000. On autocorrelation in a Poisson re-
gression model. Biometrika 87: 491–505.

Davis RA, Dunsmuir WTM, Streett SB. 2003. Observation-driven models for Pois-
son counts. Biometrika 90: 777–790.

de Jong RM, Woutersen TM. In press. Dynamic time series binary choice. Econo-
metric Theory, forthcoming.

Dueker MJ. 1997. Strengthening the case for the yield curve as a predictor of U.S.
recessions. Federal Reserve Bank of St.Louis Review 79: 41–51.

Dueker M. 1999a. Conditional heteroscedasticity in qualitative response models

of time series: A Gibbs-Sampling approach to the Bank Prime Rate. Journal of
Business and Economic Statistics 17: 466–472.

Dueker M. 1999b. Measuring monetary policy inertia in target fed funds rate
changes. Federal Reserve Bank of St. Louis Review 81: 3–10.

Dueker MJ. 2005. Dynamic forecasts of qualitative variables: A Qual VAR model
of U.S. recessions. Journal of Business and Economic Statistics 23: 96–104.

Eichengreen B, Watson MW, Grossmann RS. 1985. Bank rate policy under the
interwar gold standard. Economic Journal 95: 725–745.

Ekholm A, Smith PWJ, McDonald JW. 1995. Marginal regression analysis of a

multivariate binary response. Biometrika 82: 847–854.

Ekholm A, Smith PWJ, McDonald JW. 2000. Association models for a multivari-
ate binary response. Biometrics 56: 712–718.

Elliott G, Lieli RP. 2007. Predicting binary outcomes. Unpublished manuscript,

University of Texas.

Engle RF. 1982. Autoregressive conditional heteroskedasticity with estimates of

the variance of U.K. inflation. Econometrica 50: 987–1008.

32
Engle RF, Russell J. 1998. Autoregressive conditional duration: A new model for
irregular spaced transaction data. Econometrica 66: 1127–1162.

Estrella A. 1998. A new measure of fit for equations with dichotomous dependent
variables. Journal of Business and Economic Statistics 16: 198–205.

Estrella A. 2005. The yield curve as a leading indicator: Frequently asked ques-
tions. Federal Reserve Bank of New York. Available at http:/www.newyorkfed.org/
research/capital_markets/ycfaq.pdf. [1 October 2009].

Estrella A, Hardouvelis GA. 1991. The term structure as a predictor of real eco-
nomic activity. Journal of Finance 46: 555–576.

Estrella A, Mishkin FS. 1998. Predicting U.S. recessions: Financial variables as

leading indicators. Review of Economics and Statistics 80: 45–61.

Estrella A, Rodrigues AP, Schich S. 2003. How stable is the predictive power of
the yield curve? Evidence from Germany and the United States. Review of Eco-
nomics and Statistics 85: 629–644.

Fama EF, French KR. 1989. Business conditions and expected returns on stocks
and bonds. Journal of Financial Economics 25: 23–49.

Franses PH, van Dijk D. 2000. Non-linear Time Series Models in Empirical Fi-
nance. Cambridge University Press, Cambridge.

Gourieroux C. 2000. Econometrics of Qualitative Dependent Variables. Cambridge

University Press, Cambridge.

Granger CWJ, Teräsvirta T. 1993. Modelling Nonlinear Economic Relationships.

Oxford University Press, New York.

Hall RE, Feldstein M, Frankel J, Gordon R, Romer C, Romer D, Zarnowitz V.

2003. The NBER’s recession dating procedure. Business Cycle Dating Commit-
tee, National Bureau of Economic Research. Available at
http://www.nber.org/cycles/recessions.pdf. [2 July 2009].

33
Hall RE, Feldstein M, Frankel J, Gordon R, Poterba J, Romer D, Zarnowitz V.
2008. Determination of the December 2007 peak in economic activity. Business
Cycle Dating Committee, National Bureau of Economic Research. Available at
http://www.nber.org/dec2008.pdf. [2 July 2009].

Hamilton JD. 1989. A new approach to the economic analysis of non-stationary

time series and the business cycle. Econometrica 57: 357–384.

Hamilton JD, Jorda O. 2002. A Model of the Federal funds rate target. Journal
of Political Economy 110 : 1135-1167.

Harding D, Pagan A. 2002. Dissecting the cycle: a methodological investigation.

Journal of Monetary Economics 49: 365–381.

Hausman J, Lo A, MacKinlay C. 1992. An ordered probit analysis of transaction

stock prices. Journal of Financial Economics 31: 319–379.

Heinen A. 2003. Modelling time series count data: An autoregressive conditional

Poisson model. Core discussion paper 62.

Heinen A, Rengifo E. 2007. Multivariate autoregressive modeling of time series

count data using copulas. Journal of Empirical Finance 14: 564–583.

Honore BE, Kyriazidou E. 2000. Panel data discrete choice models with lagged
dependent variables. Eonometrica 68: 839–874.

Honore BE, Lewbel A. 2002. Semiparametric binary choice panel data models
without strictly exogenous regressors. Eonometrica 70: 2053–2063.

Hu L, Phillips PCB. 2004. Dynamics of the Federal Funds target rate: A nonsta-
tionary discrete choice approach. Journal of Applied Econometrics 19: 851–867.

Kaufmann H. 1987. Regression models for nonstationary categorical time series:

Asymptotic estimation theory. Annals of Statistics 18: 79–98.

Kauppi H. (2007): Predicting the Fed’s target rate decisions, HECER Discussion
Paper, 182. Helsinki Center of Economic Research.

34
Kauppi H. (2008): Yield-curve based probit models for forecasting U.S. recessions:
Stability and dynamics. HECER Discussion Paper, 221. Helsinki Center of Eco-
nomic Research.

Kauppi H, Saikkonen P. 2008. Predicting U.S. recessions with dynamic binary

response models. Review of Economics and Statistics 90: 777–791.

Layton AP, Moore GH. 1989. Leading indicators for the service sector. Journal
of Business and Economic Statistics 7: 379–386.

Leitch G, Tanner JE. 1991. Economic forecast evaluation: Profit versus the con-
ventional error measures. American Economic Review 81: 580–590.

Leung MT, Daouk H, Chen AS. 2000. Forecasting stock indices: A comparison
of classification and level estimation models. International Journal of Forecasting
16: 173–190.

Li WK. 1994. Time series models based on generalized linear models: Some fur-
ther results. Biometrics 50: 506–511.

Liesenfeld R, Nolte I, Pohlmeier W. 2006. Modelling financial transaction price

movements: A dynamic integer count data model. Empirical Economics 30: 795–
825.

Maddala GS. 1983. Limited-Dependent and Qualitative Variables in Econometrics.

Cambridge University Press, New York.

Maheu JM, McCurdy TH. 2000. Identifying bull and bear markets in stock re-
turns. Journal of Business and Economic Statistics 18: 100–112.

Manski CF. 1975. Maximum score estimation of the stochastic utility model of
choice. Journal of Econometrics 3: 205–228.

Manski CF. 1985. Semiparametric analysis of discrete response: Asymptotic prop-

erties of the maximum score estimator. Journal of Econometrics 27: 313–333.

Marcellino M. 2006. Leading Indicators in Handbook of Economic Forecasting eds.

Elliott G, Granger CWJ, and Timmermann A. Volume 1, Elsevier.

35
Mariano, RS, Murasawa, Y. 2003. A new coincident index of business cycles based
on monthly and quarterly series. Journal of Applied Econometrics 18: 427–443.

McCullagh P, Nelder JA. 1989. Generalized Linear Models. Second Edition, Chap-
man and Hall. London.

Mosconi R, Seri R. 2006. Non-causality in bivariate binary time series. Journal of

Econometrics 132: 379–407.

Osborn DR, Sensier M, van Dijk D. 2004: Predicting growth regimes for Euro-
pean countries, in Lucrezia Reichlin (eds.), The Euro Area Business Cycle: Stylized
Facts and Measurement Issues. Centre for Economic Policy Research.

Pagan AR, Sossounov KA. 2003. A simple framework for analyzing bull and bear
markets. Journal of Applied Econometrics 18: 23–46.

Pesaran HM, Timmermann A. 1995. Predictability of stock returns: Robustness

and economic significance. Journal of Finance 50: 1201–1228.

Poirier DJ, Ruud PA. 1988. Probit with dependent observations. Review of Eco-
nomic Studies 55: 593–614.

Rapach DE, Wohar ME, Rangvid J. 2005. Macro variables and international stock
return predictability. International Journal of Forecasting 21: 137–166.

Russell JR., Engle RF. 1998. Econometric analysis of discrete-valued irregularly-

spaced financial transactions data using a new multinomial autoregressive con-
ditional multinomial model. Economics Working Paper Series, 10, University of
California, San Diego.

Russell JR, Engle RF. 2005. A discrete-state continuous-time model of finan-

cial transactions prices and times: The autoregressive conditional multinomial-
autoregressive conditional duration model. Journal of Business and Economic
Statistics 23: 166–180.

Rydberg T, Shephard N. 2003. Dynamics of trade-by-trade price movements: De-

composition and models. Journal of Financial Econometrics 1: 2–25.

36
Shephard N. 1995. Generalized linear autoregressions. Unpublished manuscript,
Nuffield College, Oxford.

Startz R. 2008. Binomial autoregressive moving average models with an applica-

tion to U.S. Recessions. Journal of Business and Economic Statistics 26: 1–8.

Stock JH, Watson MW. 1989. New indexes of coincident and leading economic
indicators. NBER Macroeconomics Annual 4: 351–409.

Stock JH, Watson MW. 2003. Forecasting output and inflation: The role of asset
prices. Journal of Economic Literature 41: 788–829.

Tong H. 1990. Non-linear time series: A Dynamic System Approach, Oxford Uni-
versity Press, London.

Wooldridge JM. 2002. Econometric Analysis of Cross Section and Panel Data.
MIT press, Cambridge Massachusetts.

Zarnowitz V, Ozyildirim A. 2006. Time series decomposition and measurement

of business cycles, trends and growth cycles. Journal of Monetary Economics 53:
1717–1739.

Zeger SL. 1988. A regression model for time series of counts. Biometrika 75:
621–629.

Zeger SL, Qaqish B. 1988. Markov regression models for time series: A quasi-
likelihood approach. Biometrics 44: 1019–1031.

37
Chapter 2

Dynamic Probit Models and

Financial Variables in Recession
Forecasting

Abstract1

In this chapter, various financial variables are examined as predictors of the prob-
ability of a recession in the U.S. and Germany. We propose a new dynamic probit
model that outperforms the standard static model, giving accurate out-of-sample
forecasts in both countries for the recession period that began in 2001, as well
as the beginning of the recession in 2008. In accordance with previous findings,
the domestic term spread proves to be an important predictive variable, but stock
market returns and the foreign term spread also have predictive power in both
countries. In the case of Germany, the interest rate differential between the U.S.
and Germany is also a useful additional predictor.

1
A paper “Dynamic Probit Models and Financial Variables in Recession Forecasting” based
on this chapter has been published in the Journal of Forecasting, 29, 215–230, 2010, Wiley-
c
Blackwell, [2009], John Wiley & Sons, Ltd.

39
2.1 Introduction
A substantial amount of research has considered the predictive ability of various
financial variables to predict the economic growth and recession periods in dif-
ferent countries. Much of the previous analysis is focused on time series models
where the dependent variable is “continuous”, such as the growth rate of real GDP
(see, for example, the survey by Stock and Watson, 2003). However, in the re-
cent econometric literature, forecasting a binary recession indicator with probit or
logit models has attracted attention and, consequently, new time series models for
binary dependent variables have been introduced. Dueker (2002, 2005), Chauvet
and Potter (2005) and Startz (2008), among others, have proposed dynamic exten-
sions to the standard static probit model used by Estrella and Hardouvelis (1991),
Bernard and Gerlach (1998), and Estrella and Mishkin (1998), among others, to
predict recession periods. The main objective of this study is to apply the dynamic
models suggested by Kauppi and Saikkonen (2008) to predict monthly recession
periods in the U.S. and Germany.
Among various financial explanatory variables considered, the term spread,
which is the difference between the long-term and short-term interest rate, has
proved to be a useful predictor of future economic growth and recession periods
(see, for example, Estrella and Mishkin, 1998; Estrella, 2005a). However, other fi-
nancial predictors have also been suggested. For instance, if the domestic spread is
a useful predictor, then the foreign spread may also have predictive power (Bernard
and Gerlach, 1998). A potentially useful alternative is the interest rate differential
between the considered countries, which to the best of our knowledge has not been
used in recession forecasting prior to this study. Furthermore, as a forward-looking
variable, the stock market return should also have additional predictive power in
addition to interest rate-based predictive variables (Estrella and Mishkin, 1998).
Our findings extend the earlier literature in several ways. We confirm that
the domestic term spread is the primary predictive variable, but we also find
stock returns to have statistically significant predictive power for both countries.
Furthermore, in the case of German recessions the interest rate differential between

40
the U.S. and Germany is also a useful predictor, whereas the German term spread
helps predict the U.S. recessions. The U.S. term spread is also a statistically
significant explanatory variable in all predictive models fitted for German recession
periods, but its out-of-sample predictive performance seems to be poor. Out-of-
sample forecasts also lend some support to an asymmetric impact of the term
spread on the recession probability dependent on the state of the economy. Overall,
dynamic probit models outperform the standard static recession prediction models
in terms of both in-sample and out-of-sample predictions. The best models also
provide accurate out-of-sample forecasts and recession signals for the beginning of
the recession in 2008.
The rest of this chapter is organized as follows. Section 2.2 presents the probit
models to be used in forecasting, and provides a brief discussion on multiperiod
forecasts of the recession indicator. In Section 2.3 the results of the in-sample
and out-of-sample predictions of recession periods in the U.S. and Germany are
provided. Finally, Section 2.4 concludes.

2.2 Dynamic Probit Models

2.2.1 Models

In binary time series analysis, the dependent variable yt , t = 1, 2, ..., T , is a real-

ization of a stochastic process that only takes on values one and zero. In recession
forecasting, the value of an observable binary recession indicator depends on the
state of the economy in the following way

 1, if the economy is in a recessionary state at time t,
yt = (2.1)
 0, if the economy is in an expansionary state at time t.

In other words, conditional on the information set Ωt−1 , yt has a Bernoulli distri-
bution
yt |Ωt−1 ∼ B(pt ). (2.2)

Let Et−1 (·) and Pt−1 (·) denote the conditional expectation and conditional proba-
bility given the information set Ωt−1 , respectively. In the probit model the condi-

41
tional probability that yt takes the value 1 can be written as

pt = Et−1 (yt ) = Pt−1 (yt = 1) = Φ(πt ), (2.3)

where πt is a linear function of variables included in the information set Ωt−1 and
Φ(·) is a standard normal cumulative distribution function.
In the previous recession forecasting research, the standard ‘static” model has
been the most commonly used model (see, for example, Estrella and Hardouvelis,
1991; Estrella and Mishkin, 1998; Bernard and Gerlach, 1998). That is,
′
πt = ω + xt−k β, (2.4)

where xt−k is the vector of explanatory variables. In xt−k , the i th (i = 1, ..., n)

explanatory variable, xi, t−k , should satisfy the condition k ≥ h, where h is the
forecast horizon and the employed lag k may be different in different predictive
variables.
A major shortcoming of the static model (2.4) is that it does not take the
autocorrelation structure of the binary time series into account (Dueker, 1997). In
recession forecasting, this means that the previous states of the economy are not
included in the model. Therefore, a natural dynamic extension to the static model
(2.4) is obtained by adding a lagged value of the dependent time series, yt−l , to
the right hand side of (2.4). This yields the “dynamic” probit model2
′
πt = ω + δ1 yt−l + xt−k β, (2.5)

where l ≥ 1. Kauppi and Saikkonen (2008) extend this model by adding a lagged
value of πt . The resulting “dynamic autoregressive” model is given by
′
πt = ω + α1 πt−1 + δ1 yt−l + xt−k β. (2.6)

One can also consider an “autoregressive” model obtained by restricting the coef-
ficient δ1 in (2.6) to zero, that is,
′
πt = ω + α1 πt−1 + xt−k β. (2.7)
2
We use the terminology of Kauppi and Saikkonen (2008). All extensions of the static model
(2.4) are called dynamic models although model (2.5) is referred to as the “dynamic” probit
model.

42
When |α1 | < 1, it can be seen that by recursive substitution
∞
X ∞
X ∞
X ′
πt = α1i−1 ω + δ1 α1i−1 yt−l−i+1 + α1i−1 xt−k−i+1 β,
i=1 i=1 i=1

so that the dynamic autoregressive model (2.6) is an “infinite”-order extension

of the dynamic model (2.5). This presentation indicates that the autoregressive
models, (2.6) and (2.7), may be useful and parsimonious specifications if a large
number of explanatory variables are helpful in forecasting. Rydberg and Shephard
(2003) proposed a model somewhat similar to (2.6), but their model does not imply
dependence on the infinite history of the explanatory variables. Throughout this
chapter, only one lagged value of πt and of the recession indicator yt are assumed,
but including several lags is of course possible.
An interesting extension of model (2.7) is obtained by including an interaction
term
′ ′
πt = ω + α1 πt−1 + xt−k β + yt−a z t−k γ, (2.8)

where a ≥ 1. Note that the explanatory variables included in z t−k may be different
from those in xt−k . If z t−k = xt−k , the impact of the explanatory variables in xt−k
is allowed to depend on the state of the economy (cf. Kauppi and Saikkonen,
2008). Of course, it is also possible to augment model (2.8) by the lagged value
yt−l , l ≥ 1.
The parameters of models (2.4)–(2.8) can be estimated by the method of maxi-
mum likelihood (ML).3 Unfortunately, there is no formal proof of the asymptotic
properties of the maximum likelihood estimator in models (2.6)–(2.8) with an au-
toregressive structure. However, the results of Estrella and Rodrigues (1998) and
de Jong and Woutersen (in press) indicate that under reasonable regularity condi-
tions, such as the stationarity of the specification of πt and explanatory variables,
the ML estimator is consistent and asymptotically normal. Robust standard errors
allowing for autocorrelation can be obtained as in Kauppi and Saikkonen (2008).
3
In models with the autoregressive structure ((2.6), (2.7) and (2.8)) a choice for the initial
value π0 is needed. As suggested by Kauppi and Saikkonen (2008), the initial value in model
′
(2.6), for example, is set to π0 = (ω + δ1 ȳ + x̄t−k β)/(1 − α1 ), where a bar is used to denote the
sample mean of the considered variables.

43
2.2.2 Forecasts for the Recession Indicator

Kauppi and Saikkonen (2008) show how one period and multiperiod forecasts in
models (2.4)–(2.8) can be constructed by explicit formulae. A practical problem
with recession forecasting is that realized values of the recession indicator yt de-
fined in (2.1) are known after a considerable delay. The initial announcements
of many major indicators of economic activity are preliminary and often subject
to substantial revision. Thus it is difficult to identify the turning points in real
time. For instance, the most recent announcements of business cycle peak and
trough months in the U.S. have taken place from five up to twenty months after
the business cycle turning point occurred.4
In this study, the “publication lag” in the recession indicator is assumed to be
nine months. Owing to this assumed delay, the forecast horizon h consists of two
periods. The first nine months h = 1, 2, ..., 9, are related to predictions of the most
recent past values and the current value of the recession indicator. The longer-
horizon forecasts (h ≥ 10) are presumably the most interesting ones. Later in this
study, this “ahead” forecast horizon is denoted by hf , and defined as hf = h − 9
for h ≥ 10, where the number 9 is the assumed publication lag.
Kauppi and Saikkonen (2008) propose two methods of computing multiperiod
recession forecasts, termed “direct” and “iterative” (cf. forecasts in continuous de-
pendent time series models in, for example, Marcellino, Stock and Watson, 2006).
A “direct” forecast is obtained by employing lagged values of the dependent vari-
able yt−l and explanatory variables xi,t−k when k ≥ hf , provided that k ≥ 1 and
l ≥ h. This forecast is direct in the sense that the right-hand side of model (2.6),
for example, gives the h-step forecast “directly”. An “iterative” forecast at time
t−h is obtained by accounting for all possible paths and their probabilities between
yt−h and yt using the same one period model iteratively. Typically, yt−1 is used in
the one-period model instead of forecast horizon-specific predictor yt−l employed
in direct forecasts.

4
See details at http://www.nber.org/cycles/cyclesmain.html [20 March 2009].

44
2.3 Empirical Analysis of Recession Periods in the
U.S. and Germany

2.3.1 Data and Predictive Variables

We consider the domestic and foreign term spreads (SPt ), stock market returns (rt )
and the interest rate differential between the U.S. and Germany (ISt ) as predictive
variables in probit models. The term spread, defined as the difference between
the long-term and the short-term interest rates, has been the most commonly
used predictor in recession forecasting. The study of Estrella and Hardouvelis
(1991) was among the first to find the term spread a useful predictor of economic
growth and recession periods in the U.S. Bernard and Gerlach (1998), for instance,
present similar evidence for Germany. Estrella (2005a, 2005b) and the references
therein provide an extensive literature review and the main theoretical basis for
the predictive power of the term spread.
Using the static probit model (2.4), Bernard and Gerlach (1998) show that, in
addition to the domestic term spread, foreign term spreads are also useful predic-
tors in some considered countries. Estrella and Mishkin (1998) find that the stock
return is the only variable that has out-of-sample predictive power beyond the do-
mestic term spread to predict U.S. recession periods in model (2.4). These variables
have not been considered as predictors in dynamic probit models previously. Davis
and Fagan (1997) include the interest rate differentials between EU countries to
predict the output growth, but the evidence in favor of its predictive ability is quite
weak. To the best of our knowledge, interest rate differentials between different
countries have not been considered previously in recession prediction models.
The data set includes values of the recession indicator yt and the considered
explanatory variables xt in the U.S. and Germany covering the period from Jan-
uary 1972 to December 2007. We adopt the recession periods defined by the
National Bureau of Economic Research (NBER) for the U.S. and the Economic
Cycle Research Institute (ECRI) for Germany. The data set of predictive variables

45
is collected from various sources.5

2.3.2 In-Sample Results and Model Selection

In the in-sample analysis, the sample period from January 1972 to December
1994 is used to examine the performance of different probit models with various
combinations of explanatory variables. In model evaluation, the main goodness-
of-fit measure is the pseudo-R2 measure suggested by Estrella (1998). Values of
some other statistical measures are also presented in Table 2.1. We experiment
with different lag orders k and l of the explanatory variables xt−k and the lagged
dependent variable yt−l , respectively, with k and l varying between one and 12. In
practice, it has been common to set k and l equal to the forecast horizon h. On
the other hand, Estrella and Mishkin (1998) and Kauppi and Saikkonen (2008)
have emphasized that the latest values of the predictive variables included in the
information set at the time the forecast is made are not necessarily the best ones
in terms of predictive power. This indicates that better results may be obtained
by employing lags supported by model selection.
Tables 2.1 and 2.2 show the estimation results of the best in-sample models
for both countries.6 The sixth lags of the domestic and foreign term spreads
performed consistently better, on the average, than the alternative lag orders in
different probit models for both countries. Based on the model selection, the best
lag orders for the stock returns and the interest rate differential are also used in
the estimation results presented in Tables 2.1 and 2.2.

5
Recession periods for the U.S. are obtained from
http://www.nber.org/cycles/cyclesmain.html and German recession periods from
http://www.businesscycle.com/resources/cycles. Interest rates are taken from
http://www.federalreserve.gov/releases/h15/data.htm (10-year Treasury Bond and three-
month Treasury Bill rate) and http://www.bundesbank.de/statistik/statistik (10-year interest
rate and three-month money market rate). Stock returns are log-differences from the S&P 500
index (http://www.finance.yahoo.com) and German MSCI index (http://www.mscibarra.com)
[20 March 2009].
6
Details on all model selection results are available upon request. All estimations have been
executed with Matlab 7.4.0 and its BFGS optimization routine in the Optimization Toolbox.

46
 Table 2.1: In-sample results from recession prediction models for the U.S.
model static(2.4) static(2.4) dynamic(2.5) dyn.auto(2.6) auto(2.7) auto.int(2.8)
constant -0.26 -0.54 -0.17 -2.37 -0.11 -0.06
(0.24) (0.28) (0.03) (0.46) (0.09) (0.08)
US
SPt−6 -0.61 -0.42 -0.41 -0.51 -0.15 -0.32
(0.13) (0.13) (0.14) (0.20) (0.04) (0.09)
GE
SPt−6 -0.38 -0.31 -0.31 -0.11 -0.07
(0.09) (0.09) (0.11) (0.04) (0.05)
US
rt−2 -0.05 -0.12 -0.12 -0.13 -0.16
(0.02) (0.02) (0.03) (0.03) (0.04)
US
rt−4 -0.10 -0.17 -0.17 -0.08 -0.08
(0.02) (0.03) (0.03) (0.02) (0.03)
US
rt−6 -0.08 -0.07 -0.07 -0.05 -0.02
(0.03) (0.04) (0.06) (0.03) (0.03)
πt−1 -0.13 0.79 0.80
(0.21) (0.03) (0.03)
US
yt−1 3.88 4.55
(0.53) (0.90)
US US
yt−1 SPt−6 0.47
(0.12)
log-L -84.64 -58.93 -16.99 -16.79 -28.88 -22.34
2
psR 0.29 0.49 0.83 0.84 0.73 0.79
adj - psR2 0.29 0.48 0.83 0.83 0.72 0.78
AIC 86.64 64.93 23.99 24.79 35.88 30.34
BIC 90.26 75.79 36.66 39.27 48.55 44.83
QP S 0.19 0.13 0.04 0.04 0.07 0.05
CR50% 0.89 0.90 0.98 0.97 0.95 0.96
25%
CR 0.81 0.89 0.97 0.97 0.94 0.95
Notes: The models are estimated using monthly observations of the recession indicator (2.1) and explanatory
variables from 1972 M1 to 1994 M12, T = 276. Robust standard errors suggested by Kauppi and Saikkonen
(2008) are reported in parentheses. In the table, psR2 reflects the pseudo-R2 and adj - psR2 the adjusted
T −1
pseudo-R2 (Estrella, 1998), which is calculated as 1 − (1 − psR2 ) T −K−1 , where K is the number of estimated
parameters. In addition, AIC and BIC are the values of the Akaike (1974) and Schwarz (1978) information
criteria, QP S is the quadratic probability score (Diebold and Rudebusch, 1989). Furthermore, CR50% and
CR25% indicate the ratio of correct predictions with 50 and 25 percent threshold values in the classification of
recession probabilities.

47
 Table 2.2: In-sample results from recession prediction models for Germany.
model static(2.4) static(2.4) dynamic(2.5) dyn.auto(2.6) auto(2.7) auto.int(2.8)
constant 0.13 0.66 -0.17 0.09 0.69 1.72
(0.22) (0.38) (0.07) (0.34) (0.22) (0.93)
US
SPt−6 -0.55 -2.67 -1.10 -0.56 -2.53
(0.20) (0.78) (0.68) (0.19) (0.68)
GE
SPt−6 -0.95 -0.94 -0.67 -0.46 -0.47 -0.70
(0.16) (0.21) (0.28) (0.14) (0.13) (0.45)
GE
rt−3 0.01 0.04 -0.06 -0.08 -0.15
(0.02) (0.04) (0.05) (0.03) (0.06)
GE
rt−6 -0.04 -0.17 -0.18 -0.10 -0.45
(0.03) (0.07) (0.06) (0.05) (0.22)
GE
rt−9 -0.06 -0.23 -0.35 -0.25 -0.73
(0.03) (0.07) (0.15) (0.06) (0.23)
ISt−6 -0.27 -0.79 -0.44 -0.26 -0.80
(0.10) (0.23) (0.22) (0.08) (0.21)
πt−1 0.64 0.80 0.77
(0.10) (0.01) (0.03)
GE
yt−1 7.48 2.38
(1.35) (1.76)
GE US
yt−1 SPt−6 1.65
(0.48)
log-L -72.14 -54.26 -11.98 -7.65 -12.56 -1.37
2
psR 0.69 0.78 0.97 0.98 0.97 0.99
adj - psR2 0.68 0.77 0.97 0.98 0.97 0.99
AIC 74.14 61.26 19.98 16.65 20.56 10.37
BIC 77.76 73.93 34.46 32.94 35.04 26.66
QP S 0.16 0.13 0.03 0.02 0.03 0.01
CR50% 0.71 0.90 0.98 0.99 0.98 1.00
25%
CR 0.70 0.89 0.97 0.99 0.97 1.00
Note: See notes to Table 2.1.

48
 The main findings are very much the same for both countries. According to the
model selection criteria, the first lag of the dependent variable yt−1 is superior for
both countries, and it is a highly statistically significant predictor. This is in line
with the findings of Kauppi and Saikkonen (2008), and it gives tentative evidence
that the iterative multiperiod forecasts could be superior to horizon-specific direct
forecasts in out-of-sample forecasting.
The domestic term spread is the primary financial explanatory variable, but
the foreign term spread and most stock return lags are also statistically significant
predictors. The signs of the estimated coefficients are negative as expected, indi-
cating an increased probability of recession when the values of the term spreads
are relatively low. Negative stock returns also increase the probability of recession,
and it appears that the predictive power is distributed among several preceding
stock market returns. The interest rate differential is a statistically significant
predictor in the case of Germany. Its negative coefficient means that the recession
probability increases when the short-term interest rate is higher in Germany than
in the U.S. However, in the U.S. the interest rate differential turned out to be a
statistically insignificant predictor and is, therefore, not included in the reported
models in Table 2.1.
Overall, based on the in-sample evidence for both countries, it is clear that the
foreign term spreads, several lagged stock returns and the interest rate differential
in Germany add significantly to the predictive power of a model that contains
only the domestic term spread as a single explanatory variable. The dynamic
models (2.5)–(2.8) outperform the static model (2.4) in terms of in-sample perfor-
mance. However, the static model augmented with the above-mentioned additional
explanatory variables also outperforms the traditional static model where the do-
mestic term spread is the only predictor (the first and the second models in Tables
2.1 and 2.2).
The best in-sample fit for the U.S. is obtained from the dynamic model (2.5)
with the yt−1 predictor. On the other hand, the “pure” autoregressive model (2.7)
also yields good in-sample fit with a relatively large and highly statistically sig-

49
nificant estimate of the autoregressive coefficient α1 , indicating that the statistical
improvement compared with the static model (2.4) is clear.
In the autoregressive interaction model (2.8) it seemed reasonable to use an
interaction term of the form yt−1 z t−k . The first lag of the dependent variable yt−1
as such was excluded because its inclusion rendered the interaction term statisti-
cally insignificant, reducing the model to the dynamic model (2.5). In the models
presented in Tables 2.1 and 2.2, the sixth lag of the term spread is used in both
z t−k and xt−k . For both countries the estimate of the interaction term coefficient
is statistically significant, suggesting that the U.S. term spread has an asymmetric
effect on recession probability, with the asymmetry depending on the state of the
economy. As a matter of fact, this model yields the best in-sample fit for Germany.
Interestingly, the U.S. term spread has a stronger asymmetric effect than the do-
mestic term spread on German recession periods. The evidence of the asymmetric
effect of the term spread is in accordance with monetary policy having a similar
asymmetric effect on the real economy (see, e.g., Morgan, 1993; Florio, 2004).
An important issue in specifying a model for recession forecasting is the stability
of the relationship between the explanatory variables and the recession indicator
(2.1). We examined this by using the LM test proposed by Andrews and Fair (1988)
and applied by Kauppi (2008) in the context of probit model when testing potential
structural break dates suggested by Estrella, Rodrigues and Schich (2003). We
found no evidence of structural breaks in the models presented in Tables 2.1 and
2.2 at conventional significance levels.7 The evidence is in line with the findings of
Estrella et al. (2003), Wright (2006) and Kauppi (2008).
Figure 2.1 depicts the in-sample recession probabilities of the static model
(2.4) and the autoregressive model (2.7) (the first and the fifth models in Tables
2.1 and 2.2). These models are used as such also in out-of-sample forecasting in
the following section.8 It can be seen that in the autoregressive model the recession
7
Further details are available upon request.
8
This is not the case in the models employing yt−1 as a predictive variable because those one-
period models are used iteratively in out-of-sample forecasting to obtain the iterative recession
forecast.

50
probability matches better with the realized values of the recession indicator than
in the static model. In recession periods, the recession probabilities are also higher
in the autoregressive model. When the economy is in an expansionary state,
the recession probability is constantly higher in the static model, whereas in the
autoregressive model, it is very close to zero, as it should be.
US
Static model (2.4), SP Autoregressive model (2.7)
1 1

0.9 0.9

0.8 0.8

0.7 0.7

0.6 0.6

Probability
Probability

0.5 0.5

0.4 0.4

0.3 0.3

0.2 0.2

0.1 0.1

0 0
1970 1975 1980 1985 1990 1995 1970 1975 1980 1985 1990 1995
Time Time

Static model (2.4), SPGE Autoregressive model (2.7)

1 1

0.9 0.9

0.8 0.8

0.7 0.7

0.6 0.6
Probability

Probability

0.5 0.5

0.4 0.4

0.3 0.3

0.2 0.2

0.1 0.1

0 0
1970 1975 1980 1985 1990 1995 1970 1975 1980 1985 1990 1995
Time Time

Figure 2.1: In-sample recession probabilities implied by the static model (2.4), with
the domestic term spread as the only predictor and the autoregressive model (2.7)
with additional explanatory variables for the U.S. (upper panel) and Germany
(lower panel) as given in Tables 2.1 and 2.2.

51
2.3.3 Out-of-Sample Forecasting Results

The in-sample evidence shows a great deal of predictability for recession periods
in the U.S and Germany. However, in-sample predictability does not necessarily
mean out-of-sample predictability. For instance, some of the static probit models
considered by Estrella and Mishkin (1998) for the U.S. recession periods provide
the best in-sample fitted values, but perform quite poorly out of sample.
In this chapter, the first out-of-sample predictions are made for January 1995
and last ones for March 2007. The forecast period thus contains the recession
period that began in both countries in 2001. In other months, both economies
are in an expansionary state. The parameters are estimated recursively. In other
words, after adding one month to the previous estimation period and re-estimating
the parameters, forecasts for the next month are computed. This procedure is
repeated recursively until the end of the forecast period.
We examine the out-of-sample predictive ability of the models that turned out
to be the best ones according to the in-sample results presented in Tables 2.1 and
2.2. Therefore, we employ the following explanatory variables,
 ′
xUt−k
S
= US
SPt−6 , GE
SPt−6 , US
rt−2 , US
rt−4 , US
rt−6 (2.9)

and
 ′
xGE
t−k
GE
= SPt−6 , US
SPt−6 , GE
rt−3 , GE
rt−6 , GE
rt−9 , ISt−6 . (2.10)

Models where the foreign term spread is excluded are also examined. In these
cases, the vectors of the explanatory variables are denoted by xUt−k
S∗
and xGE∗
t−k .

When the domestic term spread is the only predictor, the corresponding vectors
are denoted by v Ut−k
S
and v GE
t−k .
9

With the forecast horizon h, the lags in the explanatory variables should be
tailored so that only the information included in the information set Ωt−h at the
forecast time t − h is used. For example, if the forecast horizon is 16 months
(h = 16), and because the publication lag is assumed to be nine months, it means
9
The lag orders used in the explanatory variables in these two cases are the same as in the
vectors xUS GE
t−k and xt−k presented in (2.9) and (2.10).

52
that we are interested in forecasting the value of the U.S. recession indicator seven
months (hf = 7) ahead, and the vector xUt−k
S
is given by
 ′
US US GE US
xt−k = SPt−7 , SPt−7 , rt−7 .

Forecast accuracy is evaluated by using the same statistical goodness-of-fit

measures as in the in-sample analysis. In addition, 50% and 25% threshold values
are used to classify recession probabilities and to construct “strong”, “weak” and
“no” recession signals. For example, if the recession probability is between 25
and 50 percent, the model gives a “weak” recession signal. Related to these signal
forecasts an asymmetric “forecasting point” scheme is applied. The idea, illustrated
in Table 2.4, is to put greater emphasis on correct forecasts (cf. Dueker, 2002).
It also favors a false recession alarm compared with a missed recession month.
One rationale behind this is that firms or policymakers, for example, are possibly
willing to take a “recession insurance” and accept a possible false alarm rather than
be caught by an unexpected recession.
As discussed earlier in the context of multiperiod forecasting, the most inter-
esting forecasts are typically the future values of the recession indicator. Thus
we concentrate on forecasts with forecast horizon h ≥ 10 (hf ≥ 1). The results
concerning shorter horizons (h ≤ 9) are available upon request. It is worth noting
that, in practice, iterative forecasts are computationally very demanding when the
forecast horizon is as long as 21 months in Tables 2.3 and 2.4.10 This is a difficulty
of the iterative forecasting approach employed in the dynamic models (2.5) and
(2.8) with yt−1 as a predictor. Therefore, only forecasts based on the static model
(2.4) and the autoregressive model (2.7) are considered when the forecast horizon
is so long that iterative forecasting becomes computationally infeasible.
Based on the adjusted-pseudo-R2 the best predictive models yield good out-of-
sample forecasts for the state of the U.S. economy. In Table 2.3, the highest values
of the adjusted-pseudo-R2 in different probit models are obtained with the fore-
cast horizon of 15 months (hf = 6).11 At this forecast horizon, the autoregressive
10
In iterative forecasts with h = 21, 221 different paths need to be computed and the compu-
tational burden increases rapidly if longer publication lags are considered.
11
The values of the adjusted-pseudo-R2 are adjusted to the number of parameters estimated.

53
Table 2.3: Adjusted pseudo-R2 measures of out-of-sample predictive performance
for different models in the U.S.
h 10 11 12 13 14 15 16 21
model hf 1 2 3 4 5 6 7 12
static (2.4); xUS
t−k 0.31 0.31 0.26 0.29 0.26 0.26 0.26 0.28
dyn.iter (2.5); xUS
t−k 0.50 0.52 0.49 0.49 0.37 0.37 0.37 –
auto (2.7); xUS
t−k 0.46 0.45 0.46 0.46 0.45 0.45 0.44 0.26
auto.int (2.8); xUS US
t−k , SPt−6 0.33 0.30 0.51 0.51 0.53 0.53 0.53 –
auto.int (2.8); xUS US
t−k , SPt−8 0.10 0.05 0.45 0.44 0.53 0.53 0.54 –
static (2.4); xUS∗
t−k 0.18 0.18 0.16 0.16 0.10 0.10 0.15 0.17
dyn.iter (2.5); xUS∗
t−k 0.43 0.45 0.41 0.40 0.26 0.26 0.25 –
auto (2.7); xUS∗
t−k 0.42 0.42 0.39 0.39 0.38 0.38 0.34 0.14
auto.int (2.8); xUS∗ US
t−k , SPt−6 0.21 0.18 0.40 0.38 0.43 0.43 0.36 –
static (2.4); v US
t−k 0.08 0.08 0.08 0.07 0.07 0.08 0.13 0.18
dyn.iter (2.5); v US
t−k 0.24 0.24 0.24 0.24 0.24 0.24 0.23 –
auto (2.7); v US
t−k 0.26 0.26 0.26 0.26 0.26 0.26 0.25 0.16
auto.int (2.8); v US US
t−k , SPt−6 0.31 0.30 0.29 0.29 0.28 0.28 0.23 –
2
Notes: The table presents the adjusted pseudo-R values (see Table 2.1 and Estrella, 1998) of
different models in out-of-sample predictions. The probit model is denoted at the left with the
explanatory variables included in the model. As in (2.9), and in the subsequent discussion,
xUS US∗
t−k include all explanatory variables, in xt−k the German term spread is excluded and in

v US
t−k only the U.S. term spread is employed in the model. In the autoregressive interaction

model (2.8), the term spread that is used in the interaction term z t−k is also mentioned. In
dynamic model (2.5) and in autoregressive interaction model (2.8) the first lagged value of the
recession indicator yt−1 is used in the model.

model (2.7) and the autoregressive model with the interaction term (2.8) outper-
form the dynamic model (2.5), yielding the best out-of-sample forecasts given the
explanatory variables included in the vector xUt−k
S
in (2.9). In fact, the autoregres-
sive interaction model (2.8) is the best model when the forecast horizon is between
The evidence from different goodness-of-fit measures, such as forecasting points, information
criteria or QP S, is the same. Further, and also in the case of Germany in Table 2.4, results from
the dynamic model (2.5) with yt−h and the dynamic autoregressive model (2.6) are excluded
because forecasts from the restricted static model (2.4) and the dynamic model (2.5) with yt−1
yield almost the same or even better predictions than these more general models.

54
12 to 16 months, providing evidence that the asymmetric predictive power of the
U.S. term spread found in the in-sample analysis also shows up in out-of-sample
predictions.
Overall, the models with the U.S. stock return (xUt−k
S∗
) and the models that
also include the German term spread (xUt−k
S
) outperform the models with the U.S.
term spread (v Ut−k
S
) as the only predictor across all probit model specifications and
forecast horizons. This suggests that these additional financial variables have not
only in-sample but also out-of-sample predictive content for the U.S. recessions.
In Germany, the recession in the out-of-sample period lasted considerably
longer than in the U.S, but the essential conclusions between different predictive
models are parallel to those for the U.S. However, because the term spreads soared
immediately after the recession began, the recession probability decreased amidst
the recession period. Consequently, the negative values of the adjusted-pseudo-R2
were obtained for some models, making comparisons difficult. Therefore, the fore-
casting points presented in Table 2.4 are the main model evaluation measure for
Germany.
The results of Table 2.4 confirm the previous in-sample findings. Even out
of sample, the interest rate differential between the U.S. and Germany and the
German stock return clearly have additional predictive power beyond the German
term spread in all probit models. As in the case of the U.S., when the forecast
horizon increases towards 15 or 16 months the models with an autoregressive struc-
ture, (2.7) and (2.8), seem to outperform their competitors. Interestingly, the U.S.
term spread, which is a statistically significant predictor in sample, seems to be
a rather poor predictor out of sample, because the forecasting results are much
better without it. The statistical significance of the interest rate differential, how-
ever, suggests that the U.S. monetary policy has an impact on the probability of
recession in Germany via the U.S. short-term interest rate.
Paap, Segers and van Dijk (2009) propose a model that allows for asymmetries
such that, for example, the term spread has a different lead time in recession and
expansion. This issue can be examined with model (2.8) by selecting different lag
orders of explanatory variables included in z t−k and in xt−k . In Tables 2.3 and

55
Table 2.4: Out-of-sample forecasting points of employed predictive models for
Germany.
h 10 11 12 13 14 15 16 21
f
model h 1 2 3 4 5 6 7 12
static (2.4); xGE
t−k 0.53 0.49 0.49 0.50 0.48 0.50 0.48 0.50
dyn.iter (2.5); xGE
t−k 0.66 0.64 0.63 0.61 0.60 0.60 0.55 –
auto (2.7); xGE
t−k 0.74 0.74 0.74 0.73 0.73 0.70 0.71 0.52
auto.int (2.8); xGE US
t−k , SPt−6 0.65 0.63 0.60 0.60 0.60 0.60 0.49 –
auto.int (2.8); xGE GE
t−k , SPt−6 0.60 0.52 0.52 0.50 0.50 0.48 0.45 –
GE∗
static (2.4); xt−k 0.60 0.58 0.58 0.58 0.58 0.58 0.58 0.53
dyn.iter (2.5); xGE∗
t−k 0.72 0.71 0.67 0.69 0.67 0.67 0.61 –
auto (2.7); xGE∗
t−k 0.81 0.81 0.81 0.71 0.71 0.71 0.70 0.53
auto.int (2.8); xGE∗ GE
t−k , SPt−6 0.66 0.67 0.62 0.75 0.75 0.75 0.70 –
auto.int (2.8); xGE∗ GE
t−k , SPt−7 0.85 0.84 0.85 0.73 0.73 0.73 0.70 –
static (2.4); v GE
t−k 0.29 0.29 0.29 0.30 0.29 0.30 0.41 0.30
dyn.iter (2.5); v GE
t−k 0.56 0.55 0.52 0.48 0.49 0.48 0.43 –
auto (2.7); v GE
t−k 0.44 0.44 0.44 0.44 0.42 0.42 0.41 0.42
auto.int (2.8); v GE GE
t−k , SPt−6 0.52 0.52 0.52 0.52 0.52 0.52 0.50 –
Notes: Forecasting points are obtained from the point scheme presented below by dividing the
sum of individual points by the number of maximum points obtained when predicting the state
of the economy correctly in every month. Forecast point scheme is as follows:
signal recession (yt = 1) expansion (yt = 0)
“strong” recession signal pt ≥ 0.50 1 -1
“weak” recession signal 0.25 ≤ pt < 0.50 1/2 0
“no” recession signal pt < 0.25 -1 1/2,
where pt is the recession probability obtained from (2.3). As in (2.10), and in the subsequent
discussion, xGE GE∗
t−k include all explanatory variables, in xt−k the U.S. term spread is excluded and

in v GE
t−k only the German term spread is employed in the model. See also notes to Table 2.3.

2.4 two models are considered where the predictive lag of the term spread in z t−k
is selected based on the in-sample model selection. For instance, the selection
US
z t−k = SPt−8 seems to be the best one for the U.S. However, according to the out-
of-sample results of both countries, differences between the presented two versions
of model (2.8) using different lag orders in z t−k are minor.

56
 For both countries, the best autoregressive interaction model (2.8) seems to
generate somewhat better out-of-sample forecasts than the dynamic model (2.5).
This is in contrast with the findings of Kauppi and Saikkonen (2008) and may be
due to the additional financial explanatory variables used in this study. However,
it should be pointed out that the autoregressive model (2.7) produces almost as
good out-of-sample predictions as model (2.8) for both countries. Moreover, fore-
casts obtained with model (2.7) do not require computationally intensive iterative
methods. When the forecast horizon is 21 months, which is the longest horizon
considered, the static model (2.4) without any dynamics turns out to be an ade-
quate model. However, also with this forecast horizon, the additional explanatory
variables have useful predictive power.
Figure 2.2 illustrates the out-of-sample performance of the static model (2.4)
with only the domestic term spread and the autoregressive interaction model (2.8)
with additional explanatory variables. The forecast horizon is 15 months (h = 15).
The performance of the static model considered in many previous studies is inferior
to the autoregressive interaction model for both countries. The latter model has
predictive power, especially at the beginning and at the end of the latest recession
period for both countries.
In recession forecasting the probability of continued expansion (a time period
where the economy is in an expansionary state every month) is of particular interest
(see Chauvet and Potter, 2005). Continued expansion probabilities give a similar
impression of expansionary and recessionary periods as the month-to-month pre-
dictions discussed above.12 The predictive ability of the different models appears
to depend on the state of the economy. During expansion the static model seems to
overpredict the recession probability, whereas the dynamic models perform better
in this respect (see, for example, Figure 2.2). On the other hand, in recession peri-
ods the dynamic model (2.5) constantly gives the highest probabilities of continued
expansion. Thus also according to the continued expansion probabilities, the au-
toregressive probit models (2.7) and (2.8), including the domestic term spread and
12
The results are available upon request.

57
other explanatory variables, seem to yield the most reliable predictions.
Static model (2.4), vUS
t−k
Autoregressive interaction model (2.8), xUS
t−k
1 1

0.9 0.9

0.8 0.8

0.7 0.7

0.6 0.6
Probability

Probability
0.5 0.5

0.4 0.4

0.3 0.3

0.2 0.2

0.1 0.1

0 0
1994 1996 1998 2000 2002 2004 2006 2008 1994 1996 1998 2000 2002 2004 2006 2008
Time Time

Static model (2.4), vGE Autoregressive interaction model (2.8), xGE*

t−k t−k
1 1

0.9 0.9

0.8 0.8

0.7 0.7

0.6 0.6
Probability

Probability

0.5 0.5

0.4 0.4

0.3 0.3

0.2 0.2

0.1 0.1

0 0
1994 1996 1998 2000 2002 2004 2006 2008 1994 1996 1998 2000 2002 2004 2006 2008
Time Time

Figure 2.2: Out-of-sample recession forecasts (h = 15, hf = 6) for the U.S. (upper
panel) and Germany (lower panel) based on the standard static model (2.4) with
the domestic term spread as the only predictor (left) and from the autoregressive
interaction model (2.8) (right), where the employed explanatory variables are given
in (2.9) and (2.10), and in the subsequent discussion.

2.3.4 Recession Probabilities in 2006–2008

In this section, we consider out-of-sample recession forecasts up to June 2008. Fig-

ure 2.3 depicts the recession probabilities from the beginning of year 2006 for both

58
countries. The forecast horizon is again 15 months (h = 15, hf = 6), indicating
that the latest forecasts for June 2008 are based on predictive information from
December 2007. In December 2008 the NBER announced that a peak in U.S. eco-

Static model (2.4), vUS

t−k
Autoregressive interaction model (2.8), xUS
t−k
1 1

0.9 0.9

0.8 0.8

0.7 0.7

0.6 0.6
Probability

Probability
0.5 0.5

0.4 0.4

0.3 0.3

0.2 0.2

0.1 0.1

0 0
2006 2007 2008 2009 2006 2007 2008 2009
Time Time

Static model (2.4), vGE

t−k
Autoregressive interaction model (2.8), xGE*
t−k
1 1

0.9 0.9

0.8 0.8

0.7 0.7

0.6 0.6
Probability

Probability

0.5 0.5

0.4 0.4

0.3 0.3

0.2 0.2

0.1 0.1

0 0
2006 2007 2008 2009 2006 2007 2008 2009
Time Time

Figure 2.3: Recession forecasts (h = 15, hf = 6) for the U.S. (upper panel) and
Germany (lower panel) in 2006 M2–2008 M6. Forecasts from the static model
(2.4) (left) with only the domestic term spread employed and the autoregressive
interaction model (2.8) (right), where the employed explanatory variables are given
in (2.9) and (2.10), and in the subsequent discussion. The beginning of a recession
is indicated by a vertical line.

59
nomic activity had occurred in December 2007. Similarly, ECRI made an an-
nouncement that for Germany a peak had occurred in April 2008. After these
peak months, both countries have been in a recession.
As seen in Figure 2.3, the autoregressive interaction model (2.8) predicts the
beginning of the recession well for both countries, especially for Germany. For the
U.S., the recession forecast exceeded the 50% threshold value in August 2007 and
thereafter the recession probability has been relatively high. The static model (2.4)
with the domestic term spread as the only explanatory variable, which has been
the standard recession prediction model, does not give as precise recession and ex-
pansion signals as the autoregressive interaction model. As in the previous section,
the performance of this standard model appears disappointing in comparison with
different dynamic models, such as model (2.8).

2.4 Conclusions
We examine the performance of recession prediction models that include a number
of financial explanatory variables. The results indicate that, compared with the
standard static recession prediction model used in many previous studies, statis-
tically significant additional predictive power is obtained by allowing for dynamic
structures in the model. In particular, models with an autoregressive structure
outperformed the static model and they were also somewhat better than other dy-
namic models considered in terms of out-of-sample performance. The best model
for the U.S. and Germany turned out to be an autoregressive interaction model in
which the term spread between the long-term and short-term interest rate has an
asymmetric effect on recession probability, with the asymmetry depending on the
state of the economy.
In accordance with previous studies, the term spread is found a useful predictor
for both the U.S. and German recession periods, but for both countries, additional
predictive power is provided by stock returns. For Germany, the short-term interest
rate differential between the U.S and Germany also has substantial predictive
power in both in-sample and out-of-sample prediction. The same holds for the

60
German term spread when forecasting the U.S. recessions. Furthermore, the U.S.
term spread is a statistically significant predictor in the case of Germany, but its
out-of-sample predictive power appears poor.

References
Andrews DWK, Fair RC. 1988. Inference in non-linear econometric models with
structural change. Review of Economic Studies 55: 615–640.

Akaike H. 1974. A new look at statistical model identification. IEEE Transactions

on Automatic Control 19: 713–723.

Bernard H, Gerlach S. 1998. Does the term structure predict recessions? The in-
ternational evidence. International Journal of Finance and Economics 3: 195–215.

Chauvet M, Potter S. 2005. Forecasting recession using the yield curve. Journal
of Forecasting 24: 77–103.

Davis EP, Fagan G. 1997. Are financial spreads useful indicators of future infla-
tion and output growth in EU countries? Journal of Applied Econometrics 12:
701–714.

de Jong RM, Woutersen TM. In press. Dynamic time series binary choice. Econo-
metric Theory, forthcoming.

Diebold FX, Rudebusch GD. 1989. Scoring the leading indicators. Journal of
Business 62: 369–391.

Dueker MJ. 1997. Strengthening the case for the yield curve as a predictor of U.S.
recessions. Federal Reserve Bank of St.Louis Review 79: 41–51.

Dueker MJ. 2002. Regime-dependent recession forecasts and the 2001 recession.
Federal Reserve Bank of St.Louis Review 84: 29–36.

Dueker MJ. 2005. Dynamic forecasts of qualitative variables: A qual VAR model
of U.S. recessions. Journal of Business and Economic Statistics 23: 96–104.

61
Estrella A. 1998. A new measure of fit for equations with dichotomous dependent
variables. Journal of Business and Economic Statistics 16: 198–205.

Estrella A. 2005a. The yield curve as a leading indicator: Frequently asked ques-
tions. Federal Reserve Bank of New York. http:/www.newyorkfed.org/research/
capital_markets/ycfaq.pdf. [30 October 2008].

Estrella A. 2005b. Why does the yield curve predict output and inflation? Eco-
nomic Journal 115: 722–744.

Estrella A, Hardouvelis GA. 1991. The term structure as a predictor of real eco-
nomic activity. Journal of Finance 46: 555–576.

Estrella A, Mishkin FS. 1998. Predicting U.S. recessions: Financial variables as

leading indicators. Review of Economics and Statistics 80: 45–61.

Estrella A, Rodrigues AP. 1998. Consistent covariance matrix estimation in probit

models with autocorrelated disturbances. Federal Reserve Bank of New York Staff
Reports No. 39.

Estrella A, Rodrigues AP, Schich S. 2003. How stable is the predictive power of the
yield curve: Evidence from Germany and the United States. Review of Economics
and Statistics 85: 629–644.

Florio A. 2004. The asymmetric effects of monetary policy. Journal of Economic

Surveys 18: 409–426.

Kauppi H. 2008. Yield-Curve Based Probit Models for Forecasting U.S. Reces-
sions: Stability and Dynamics. HECER Discussion Paper, 221. Helsinki Center
of Economic Research.

Kauppi H, Saikkonen P. 2008. Predicting U.S. recessions with dynamic binary

response models. Review of Economics and Statistics 90: 777–791.

Marcellino M, Stock JH, Watson MW. 2006. A comparison of direct and iterated
AR methods for forecasting macroeconomic time series. Journal of Econometrics
135: 499–526.

62
Morgan DP. 1993. Asymmetric effects of monetary policy. Federal Reserve Bank
of Kansas City Economic Review 78: 21–33.

Paap R, Segers R, van Dijk D. 2009. Do Leading Indicators Lead Peaks More
Than Troughs? Journal of Business and Economic Statistics 27: 528–543.

Rydberg T, Shephard N. 2003. Dynamics of trade-by-trade price movements: De-

composition and models. Journal of Financial Econometrics 1: 2–25.

Schwarz G. 1978. Estimating the dimension of a model. Annals of Statistics 6:

461–464.

Startz R. 2008. Binomial autoregressive moving average models with an applica-

tion to U.S. Recessions. Journal of Business and Economic Statistics 26: 1–8.

Stock JH, Watson MW. 2003. Forecasting output and inflation: The role of asset
prices. Journal of Economic Literature 41: 788–829.

Wright JH. 2006. The yield curve and predicting recessions. Finance and Eco-
nomics Discussion Series No. 7, Board of Governors of the Federal Reserve Sys-
tem.

63
Chapter 3

Testing an Autoregressive Structure

in Binary Time Series Models

Abstract1

This chapter introduces a Lagrange Multiplier (LM) test for testing an autoregres-
sive structure in a binary time series model proposed by Kauppi and Saikkonen
(2008). Simulation results indicate that the two versions of the proposed LM test
have reasonable size and power properties when the sample size is large. A para-
metric bootstrap method is suggested to obtain approximately correct sizes also
in small samples. The use of the test is illustrated by an application to recession
forecasting models using monthly U.S. data.

1
An earlier version of this chapter has been published in HECER Discussion Papers, No.
243, 2008.

65
3.1 Introduction
Recently, Rydberg and Shephard (2003), Chauvet and Potter (2005) and Startz
(2008), among others, have introduced new time series models for binary dependent
variables. In this chapter, the “dynamic autoregressive” probit model suggested
by Kauppi and Saikkonen (2008) is considered. We develop Lagrange Multiplier
(LM) test which can be used to test the adequacy of a restricted model in which
the autoregressive structure is excluded.
The proposed LM test is attractive because it only requires estimates from the
restricted models, which can be obtained by using standard econometric software
packages. According to our simulations, the two versions of the LM test considered
have reasonable size and high power, especially in large samples. In small samples,
a parametric bootstrap method is proposed to obtain critical values which are
more reliable than the asymptotic ones. In an empirical application, the LM tests
are used to assess recession forecasting models for the U.S.
The plan of this chapter is as follows. The probit model is introduced in Section
3.2 and the LM tests are developed in Section 3.3. Results of the simulation and
bootstrap experiments are provided in Section 3.4 and the empirical example is
presented in Section 3.5. Finally, Section 3.6 concludes.

3.2 Model
Consider the binary valued stochastic process yt , t = 1, 2, ..., T , and let Et−1 (·)
and Pt−1 (·), respectively, signify the conditional expectation and the conditional
probability given the information set Ωt−1 . Conditional on Ωt−1 , yt has a Bernoulli
distribution, that is,
yt |Ωt−1 ∼ B(pt ). (3.1)

In the probit model

pt = Et−1 (yt ) = Pt−1 (yt = 1) = Φ(πt (θ)), (3.2)

66
where Φ(·) is a standard normal cumulative distribution function. The model πt (θ)
is a linear function of variables in the information set Ωt−1 and the parameter vector
θ.
In the previous literature, the “static” model
′
πt (θ) = ω + xt−1 β, (3.3)

has been the most commonly used specification. It has been employed in various
applications, such as forecasting the recession periods of an economy (see, e.g.,
Estrella and Mishkin, 1998). A natural extension of the static model (3.3) is the
dynamic specification
′
πt (θ) = ω + δ1 yt−1 + xt−1 β, (3.4)

where the lagged value of the dependent variable is also assumed to belong to the
information set (see, e.g., Cox, 1981).
Kauppi and Saikkonen (2008) generalize the dynamic model (3.4) by adding a
lagged value of πt (θ), giving
′
πt (θ) = ω + α1 πt−1 (θ) + δ1 yt−1 + xt−1 β, (3.5)

where |α1 | < 1.2 This induces a first-order autoregressive structure to the model
equation. It is worth noting that alternative, but very similar, models have been
proposed by Rydberg and Shephard (2003) and Kauppi (2008). The LM tests
developed in the next section can straightforwardly be extended to these models
as well.
The parameters of models (3.3)–(3.5) can conveniently be estimated by the
method of maximum likelihood (ML). Conditional on initial values, the log-likelihood
function is
T
X T 
X 
l(θ) = lt (θ) = yt log(Φ(πt (θ))) + (1 − yt ) log(1 − Φ(πt (θ))) , (3.6)
t=1 t=1

where lt (θ) is the log-likelihood for t:th observation. The score function is
T T
∂l(θ) X X yt − Φ(πt (θ)) ∂πt (θ) 
s(θ) = = st (θ) = φ(πt (θ)) , (3.7)
∂θ t=1 t=1
Φ(πt (θ))(1 − Φ(πt (θ))) ∂θ
2
For simplicity, only the first lags of the dependent variable, yt−1 , and πt (θ), πt−1 (θ), are
employed.

67
where φ(·) signifies the probability density function of the standard normal distri-
bution and an explicit expression of the derivative term ∂πt (θ)/∂θ will be given
in the next section. The ML estimator θ̂, which solves the first order condition
s(θ̂) = 0, is found by maximizing the log-likelihood function (3.6) with numerical
methods.

3.3 LM Tests
In applications, model (3.5) may be a superior to its restricted version (3.4) but,
on the other hand, its ML estimation is more complicated and no estimation
procedures are readily available in standard econometric software packages. Thus,
it is of interest to start with the simpler model (3.4) and check for its adequacy
by testing whether the autoregressive coefficient α1 in (3.5) is zero. The null
hypothesis of interest is therefore

H0 : α1 = 0. (3.8)

In this context, the LM test is attractive because it only requires the estimation
of the parameters of model (3.4). The general LM test statistic (see, e.g., Engle,
1984) for the null hypothesis (3.8) can be written as

′
LM = s(θ̃) I(θ̃)−1 s(θ̃), (3.9)

where θ̃ is the restricted ML estimate of θ restricted by (3.8), s(θ̃) is the score

vector (3.7) evaluated at θ̃, and I(θ̃) is a consistent estimate of the information
matrix I(θ). Under H0 , the test statistic (3.9) has an asymptotic χ21 distribution.
Following Davidson and MacKinnon (1984) we can construct two LM test
statistics for the null hypothesis (3.8). The first one is

′
 ′
−1 ′
LM1 = ι S(θ̃) S(θ̃) S(θ̃) S(θ̃) ι, (3.10)

where ι is a vector of ones and the matrix S(θ̃) is given by

 ′
S(θ̃) = s1 (θ̃) s2 (θ̃) ... sT (θ̃) .

68
Expression (3.10) can also be seen as the regression sum of squares from the arti-
ficial linear regression
ι = S(θ̃)a + error.

Using the symbols Φ̃t = Φ(πt (θ̃)) and φ̃t = φ(πt (θ̃)), a second LM test statistic
can be based on the artificial regression

r(θ̃) = R(θ̃)b + error, (3.11)

where
 ′ ′ ′
′
R(θ̃) = R1 (θ̃) R2 (θ̃) ... RT (θ̃)

with
 −1/2 ∂π (θ̃)
t
Rt (θ̃) = Φ̃t (1 − Φ̃t ) φ̃t
∂θ
and
 ′
r(θ̃) = r1 (θ̃) r2 (θ̃) ... rT (θ̃)

with
 1 − Φ̃ 1/2  Φ̃ 1/2
t t
rt (θ̃) = yt + (yt − 1)
Φ̃t 1 − Φ̃t
 −1/2  
= (1 − Φ̃t )Φ̃t yt − Φ̃t .

Running the artificial regression (3.11) and computing the regression sum of squares
yields the test statistic

′
 ′
−1 ′
LM2 = r(θ̃) R(θ̃) R(θ̃) R(θ̃) R(θ̃) r(θ̃). (3.12)

′ ′
Because R(θ̃) r(θ̃) = s(θ̃) = S(θ̃) ι, it can be seen that the test statistics LM1 and
LM2 only differ in the way the information matrix estimate I(θ̃) is constructed.
Note that the test statistics LM1 and LM2 can also be expressed as
 ∂π (θ̃)  ∂π (θ̃) ′ −1 X
T T
! T
 ∂π (θ̃) ′ X  ∂π (θ̃) 
t ˜ t t t
X
2
LM1 = d̃t dt d̃t ,
t=1
∂θ t=1
∂θ ∂θ t=1
∂θ

and
T T
!−1 T
X  ∂π (θ̃) ′
t
X φ̃2t  ∂π (θ̃)  ∂π (θ̃) ′
t t
X  ∂π (θ̃) 
t
LM2 = d̃t d̃t ,
t=1
∂θ t=1
Φ̃t (1 − Φ̃t ) ∂θ ∂θ t=1
∂θ

69
where
yt − Φ̃t
d̃t = φ̃t .
Φ̃t (1 − Φ̃t )
This shows that the derivative term ∂πt (θ)/∂θ evaluated at θ̃ is central for the
test statistics. From (3.5), the derivative is defined as
   
∂πt (θ)
1 + α1 ∂πt−1 (θ)
 ∂ω   ∂ω 
 ∂πt (θ)  
  πt−1 (θ) + α1 ∂π∂α t−1 (θ)

∂πt (θ)  ∂α1 

= =  1 
∂θ  ∂π t (θ)
 ∂δ   yt−1 + α1 ∂δ
  ∂π t−1 (θ) 

 1   1 
∂πt (θ) ∂πt−1 (θ)
∂β
xt−1 + α1 ∂β

and under H0 , it is
   
∂πt (θ̃)
∂ω
1
   
∂πt (θ̃)
  πt−1 (θ̃)
   
∂πt (θ̃)  ∂α1

= = .
∂θ  ∂πt (θ̃)
  yt−1
  
∂δ1
 
   
∂πt (θ̃)
∂β
xt−1

3.4 Simulation Results

The two LM tests described in the previous section are asymptotically equivalent.
In this section, their finite small-sample properties are studied by simulation.3
We simulated realizations from the Bernoulli distribution (3.1) using two different
models4
πt (θ) = −0.30 + α1 πt−1 (θ) + 0.50 yt−1 , (3.13)
3
Matlab version 7.5.0 and the BFGS algorithm in the Optimization Toolbox is used in
simulation and estimation. Eviews code for computing LM tests (3.10) and (3.12) is also available
upon request.
4
The initial value π0 (θ) in (3.5) is set to π0 (θ) = (ω + δ1 ȳ + x̄t−k β)/(1 − α1 ) with the
parameter values used in (3.13) and (3.14) (see Kauppi and Saikkonen, 2008). A bar is used to
denote the sample mean of the considered variables.

70
and
πt (θ) = −0.30 + α1 πt−1 (θ) + 1.00 yt−1 − 0.20 xt−1 . (3.14)

Since many macroeconomic and financial time series exhibit rather strong persis-
tence we assume the following AR(1) process for the explanatory variable xt ,

xt = 0.1 + 0.90xt−1 + εt , εt ∼ NID(0, 1).

Positive coefficients for the lagged yt−1 in (3.13) and (3.14) indicate that the re-
alized values of yt , i.e. zeros and ones, tend to cluster in the similar way as, for
example, recession periods of the economy (see Section 3.5, and Chapters 2 and 5
in this thesis).
We provide simulation evidence for sample sizes 150, 300, 500, 1000 and 2000.
For all generated series, 200 extra observations were simulated and discarded from
the beginning of every sample to avoid initialization effects. We report empirical
sizes of the models at 10%, 5% and 1% significance levels. All results are based
on 2000 replications. However, in some cases, a little more than 2000 replications
(about 20–30 extra replications) are needed because of numerical difficulties in the
optimization of the log-likelihood function (3.6).
Empirical sizes of the LM tests for selected parameter values in (3.13) and
(3.14) are presented in Tables 3.1 and 3.2. The empirical sizes range between
sample sizes. Both tests seem to be rather severely oversized in small samples
(T = 150, T = 300 and T = 500), but for larger samples, the empirical sizes are
rather close to the nominal levels.

Table 3.1: Empirical size of the LM1 and LM2 tests in the model (3.13).
T LM1 LM2
10% 5% 1% 10% 5% 1%
150 28.5 15.3 3.1 28.9 14.7 3.2
300 19.6 9.0 1.5 19.3 8.9 1.4
500 17.0 8.5 1.4 16.8 8.4 1.3
1000 14.3 6.6 1.1 14.3 6.6 1.1
2000 10.3 5.3 1.2 10.3 5.4 1.2
Notes: In size simulations, α1 = 0. The results are based on the 2000 replications.

71
 Table 3.2: Empirical size of the LM1 and LM2 tests in the model (3.14).
T LM1 LM2
10% 5% 1% 10% 5% 1%
150 42.8 26.3 7.1 41.6 23.0 5.0
300 30.1 15.2 3.3 28.4 14.5 2.3
500 22.0 10.8 2.2 21.0 10.3 2.0
1000 14.0 7.6 1.5 13.7 7.3 1.3
2000 11.4 5.7 0.9 11.4 5.3 0.9
Note: See notes to Table 3.1.

Rejection rates presented in Tables 3.1 and 3.2 are based on the critical values
from the asymptotic χ21 distribution. However, one can use a parametric bootstrap
method to obtain alternative, potentially more accurate, critical values than the
asymptotic ones. The employed procedure is the following. ML estimates θ̃ =
′
(ω̃ δ̃ β̃) and LM test statistics are computed under the null hypothesis α1 = 0.
Bootstrap samples yτb and the values of test statistics LM1b and LM2b , b = 1, 2, ..., B,
are then generated from the data-generating process

yτb ∼ B(Φ(πτb (θ̃))), (3.15)

where τ = 1, 2, ..., T , and

′
πτb (θ̃) = ω̃ + yτb −1 δ̃ + xτ −1 β̃.

Finally, bootstrap critical values at different significance levels are obtained from
the empirical distribution of the test statistics LM1b and LM2b . The number of
bootstrap replications B is set to 500 and the simulation is carried out for 500
replications.
As an illustration for the usefulness of the proposed bootstrap method, Table
3.3 presents the rejection rates based on the bootstrap critical values instead of
the asymptotic ones. Compared with the results shown in Table 3.2, the empirical
sizes of the LM tests are now much closer to the nominal values.

72
Table 3.3: Empirical size of the LM1 and LM2 tests using the model (3.14) and
bootstrap critical values.
LM1 LM2
T 10% 5% 1% 10% 5% 1%
150 9.0 4.2 1.0 9.6 6.2 1.0
300 9.6 5.4 1.6 9.0 6.2 1.0
500 12.2 5.0 1.2 11.2 6.4 0.4

Size-adjusted empirical power functions with different sample sizes T at the

5% level are depicted in Figures 3.1 and 3.2 using (3.13) and (3.14) with different
values of α1 . It is expected that in many applications the parameter α1 is non-
negative and, therefore, we concentrate on values from α1 = 0.00 up to α1 = 0.80.5
The power seems to increase rather quickly when the value of α1 increases, in
particular when the explanatory variable xt is employed in the model. It appears
that the power of LM2 is typically slightly higher that of LM1 in both cases when
α1 > 0. However, the differences are not very large.
Even for smaller sample sizes reasonable power is obtained although the power
of tests is slightly decreasing at very high values of α1 . As Kauppi and Saikkonen
(2008) note, a potential reason for this finding is that πt−1 (θ) and yt−1 may interact
in a complicated way which could affect the statistical significance of πt−1 (θ),
especially in small samples.

5
The evidence appears to be rather the same with the negative values of α1 , especially in
large samples.

73
 T=150 T=300
1 1

0.9 0.9

0.8 0.8

0.7 0.7

0.6 0.6
Power

Power
0.5 0.5

0.4 0.4

0.3 0.3

0.2 0.2
LM LM
0.1 1 0.1 1
LM LM
2 2
0 0
0 0.2 0.4 0.6 0.8 0 0.2 0.4 0.6 0.8
α α
1 1

T=500 T=1000
1 1

0.9 0.9

0.8 0.8

0.7 0.7

0.6 0.6
Power

Power

0.5 0.5

0.4 0.4

0.3 0.3

0.2 0.2
LM LM
0.1 1 0.1 1
LM LM
2 2
0 0
0 0.2 0.4 0.6 0.8 0 0.2 0.4 0.6 0.8
α1 α
1

T=2000
1

0.9

0.8

0.7

0.6
Power

0.5

0.4

0.3

0.2
LM1
0.1
LM2
0
0 0.2 0.4 0.6 0.8
α
1

Figure 3.1: Empirical power in the case of model (3.13).

74
 T=150 T=300
1 1

0.9 0.9

0.8 0.8

0.7 0.7

0.6 0.6
Power

Power
0.5 0.5

0.4 0.4

0.3 0.3

0.2 0.2
LM LM
0.1 1 0.1 1
LM LM
2 2
0 0
0 0.2 0.4 0.6 0.8 0 0.2 0.4 0.6 0.8
α α
1 1

T=500 T=1000
1 1

0.9 0.9

0.8 0.8

0.7 0.7

0.6 0.6
Power

Power
0.5 0.5

0.4 0.4

0.3 0.3

0.2 0.2
LM LM
0.1 1 0.1 1
LM LM
2 2
0 0
0 0.2 0.4 0.6 0.8 0 0.2 0.4 0.6 0.8
α1 α
1

T=2000
1

0.9

0.8

0.7

0.6
Power

0.5

0.4

0.3

0.2
LM
0.1 1
LM
2
0
0 0.2 0.4 0.6 0.8
α
1

Figure 3.2: Empirical power in the case of model (3.14).

75
3.5 Application: U.S. Recession Forecasting Mod-
els
Forecasting recession periods has been one of the most common empirical appli-
cations of binary time series models. Predicting the direction-of-change in stock
market returns is an example of another potential application (see, e.g., Leung,
Daouk and Chen, 2000; Rydberg and Shephard, 2003; Nyberg, in press).
In recession forecasting, the dependent variable is the recession indicator

 1, if the economy is in a recession state at time t,
yt = (3.16)
 0, otherwise.

Although in this chapter we are not interested in out-of-sample forecasting, we

consider forecasting models behind the “direct” (using horizon-specific predictor
yt−15 ) and “iterative” (yt−1 ) multi-step forecasts for the binary response (3.16)
(for details, see Kauppi and Saikkonen, 2008). The difference between “direct”
and “iterative” forecasts is similar to that in time series models for traditional
continuous variables (see, e.g., Marcellino, Stock, and Watson, 2006).
In this application, the recession periods identified by the National Bureau of
Economic Research (NBER) for the U.S. between January 1954 and December
2006 are employed. The term spread (SPt ) between the long-term and short-term
interest rate and the stock market return (rt ) are found to be useful predictive
variables (see, e.g., Estrella and Mishkin, 1998; Nyberg, 2010). Explanatory vari-
ables are described in more detail in Table 3.4. Table 3.5 presents the estimated
predictive models when the forecast horizon h is assumed to be six months (h = 6).
The fact the NBER business cycle turning points are announced with a delay is
taken into account in “direct” forecasting models.6

6
We assume that this “publication lag” is nine months. Thus, in direct forecasting models
the lag yt−15 is used given the six-month forecast horizon. For further details see, for example,
Kauppi and Saikkonen (2008), Kauppi (2008), and Nyberg (2010).

76
 Table 3.4: Explanatory variables.
Rt 10-year Treasury bond yield rate, constant maturity
it Three-month Treasury bill rate, secondary market
SPt Term spread, Rt − it
rt Monthly stock market return, log-difference of the S&P500 index
Notes: Interest rates are from http://www.federalreserve.gov/releases/h15/data.htm. S&P500
stock index is taken from http://finance.yahoo.com and http://www.econstats.com. [2 July
2009].

In a direct forecasting (Model 1) shown in the first column of Table 3.5, the p-
values of the two LM tests based on asymptotic critical values are zero indicating
that the inclusion of an autoregressive structure gives a better model. The same
conclusion is drawn by using bootstrap critical values. Further, when πt−1 (θ) is
included in the model, it is a statistically significant predictor (Model 2) according
to the Wald-type of test comparing the estimated coefficient of α1 and its ro-
bust standard error or using likelihood ratio test between the competitive models.
The values of the estimated log-likelihood function and the pseudo-R2 measure
(Estrella, 1998) are also substantially higher in the latter model.
When the “iterative” predictive model is considered (Model 3), the values of
the test statistics LM1 and LM2 are statistically insignificant at 5% significance
level compared with both asymptotic and bootstrap critical values. When the
autoregressive structure is imposed on the model the coefficient for πt−1 (θ) is
indeed statistically insignificant in the extended Model 4.
In conclusion, in these two examples, outcomes of the two LM tests are in
accordance with the Wald test when testing the statistical significance of the
autoregressive structure. The recommendation is that an autoregressive model
structure is worth considering as an alternative to the static recession prediction
model (see, e.g., Estrella and Mishkin, 1998), possibly augmented by the forecast
horizon-specific lagged value of yt as presented in the first two models of Table 3.5.
However, the lagged state of the economy, yt−1 , seems to be the main dynamic
part in the iterative forecasting model (Models 3 and 4).

77
 Table 3.5: Estimation results for the recession prediction models.
Model 1 2 3 4
constant -0.50 -0.02 -1.71 -2.00
(0.16) (0.04) (0.17) (0.24)
SPt−6 -0.61 -0.21 -0.58 -0.67
(0.13) (0.05) (0.14) (0.16)
rt−6 -0.05 -0.08 -0.01 -0.01
(0.02) (0.02) (0.03) (0.03)
πt−1 (θ) 0.81 -0.17
(0.03) (0.09)
yt−1 3.38 3.95
(0.25) (0.35)
yt−15 -0.41 -0.07
(0.36) (0.16)
log-L -185.94 -136.60 -55.63 -55.29
2
pseudo − R 0.192 0.367 0.689 0.691
LM1 25.25 2.75
p-value 0.000 0.097
LM2 36.18 0.65
p-value 0.000 0.420
Bootstrap
critical values
LM1 10% 3.06 5.10
5% 3.93 6.88
1% 5.90 9.61
LM2 10% 2.71 2.30
5% 3.75 3.07
1% 5.37 6.45
Notes: Models are estimated using U.S. data from 1954 M01 to 2006 M12 (T = 636). First 21
months are used as initial values. Robust standard errors (see Kauppi and Saikkonen, 2008) are
reported in parentheses. The estimated value of the log-likelihood function (3.6) and the
pseudo-R2 measure (Estrella, 1998) are also provided as well as the values of the LM1 and
LM2 test statistics, their p-values based on the asymptotic χ21 distribution, and critical values
obtained by bootstrap.

78
3.6 Conclusions
We have proposed LM tests for testing an autoregressive model structure in binary
time series models. Based on a limited simulation study, the tests appear to have
reasonable empirical size, especially in large samples, and high power. For small
samples, the proposed bootstrap simulation method provides improved empirical
sizes. An empirical example of recession forecasting models for the U.S. illustrates
that the inclusion of an autoregressive model structure may be a useful addition
to the recession prediction model.

References
Chauvet M, Potter S. 2005. Forecasting recession using the yield curve. Journal
of Forecasting 24: 77–103.

Cox DR. 1981. Statistical analysis of time series: Some recent developments.
Scandinavian Journal of Statistics 8: 93–115.

Davidson R, MacKinnon JG. 1984. Convenient specification tests for logit and
probit models. Journal of Econometrics 25: 241–262.

Engle RF. 1984. Wald, likelihood ratio and Lagrange Multiplier tests in econo-
metrics, in Griliches, Z. and Intriligator MD. (eds) in Handbook of Econometrics,
Vol. II. Amsterdam, North-Holland.

Estrella A. 1998. A new measure of fit for equations with dichotomous dependent
variables. Journal of Business and Economic Statistics 16: 198–205.

Estrella A, Mishkin FS. 1998. Predicting U.S. recessions: Financial variables as

leading indicators. Review of Economics and Statistics 80: 45–61.

Kauppi H. 2008. Yield-curve based probit models for forecasting U.S. recessions:
Stability and dynamics. HECER Discussion Paper, 221. Helsinki Center of Eco-
nomic Research.

79
Kauppi H, Saikkonen P. 2008. Predicting U.S. recessions with dynamic binary
response models. Review of Economics and Statistics 90: 777–791.

Leung MT, Daouk H, Chen AS. 2000. Forecasting stock indices: A comparison
of classification and level estimation models. International Journal of Forecasting
16: 173–190.

Marcellino M, Stock JH, Watson MW. 2006. A comparison of direct and iterated
AR methods for forecasting macroeconomic time series. Journal of Econometrics
135: 499–526.

Nyberg H. 2010. Dynamic probit models and financial variables in recession fore-
casting. Journal of Forecasting 29: 215–230.

Nyberg, H. In press. Forecasting the direction of the U.S. stock market with dy-
namic binary probit models. International Journal of Forecasting, forthcoming.

Rydberg T, Shephard N. 2003. Dynamics of trade-by-trade price movements: De-

composition and models. Journal of Financial Econometrics 1: 2–25.

Startz R. 2008. Binomial autoregressive moving average models with an applica-

tion to U.S. Recessions. Journal of Business and Economic Statistics 26: 1–8.

80
Chapter 4

Forecasting the Direction of the

U.S. Stock Market with Dynamic
Binary Probit Models

Abstract1

Several empirical studies have documented that the signs of excess stock returns
are, to some extent, predictable. In this chapter, we consider the predictive abil-
ity of the binary dependent dynamic probit model in predicting the direction of
monthly excess stock returns. The recession forecast obtained from the model
for a binary recession indicator appears to be the most useful predictive variable,
and once it is employed, the sign of the excess return is predictable in sample. A
new dynamic “error correction” probit model proposed in the chapter yields better
out-of-sample sign forecasts with the resulting average trading returns higher than
those of the buy-and-hold strategy or trading rules based on ARMAX models.

1
A paper based on this chapter has been accepted for publication in the International Journal
of Forecasting, Elsevier (forthcoming).

81
4.1 Introduction
In the financial econometric literature there is considerable evidence that excess
stock market returns are, to some extent, predictable. The main objective has
been to predict the overall level, the conditional mean, of excess stock returns. It
is emphasized that even though the predictability is statistically weak, it can be
economically meaningful.
However, many studies have documented that only the direction of excess stock
returns or other asset returns are predictable (see, among others, Breen, Glosten,
and Jagannathan, 1989; Hong and Chung, 2003; Christoffersen and Diebold, 2006).
A possible explanation for this is that the noise in the observed returns is too high
for the accurate forecasting of the overall return. Leitch and Tanner (1991) find
that the direction of the change is the best criterion for predictability because
traditional statistical summary statistics may not be closely related to the profits
that investors are seeking in the financial market. Directional predictability is also
important for market timing, which is crucial for asset allocation decisions between
stock and risk-free interest rate investments.
The previous findings of directional predictability are mainly based on time
series models for the excess stock return. For instance, Christoffersen and Diebold
(2006) and Christoffersen et al. (2007) have considered the theoretical connection
between asset return volatility and asset return sign predictability and verified
that volatility and higher-order conditional moments of returns have statistically
significant explanatory power in sign prediction. Even though there is not much
previous research, binary dependent time series models provide an another way to
forecast the direction of excess stock returns. Various classification-based qualita-
tive models, such as traditional static logit and probit models, were considered by
Leung, Daouk, and Chen (2000), whereas Hong and Chung (2003), Rydberg and
Shephard (2003), and Anatolyev and Gospodinov (2010) used the so-called auto-
logistic model to predict the return direction. In the last two papers the return is
decomposed into a sign component and an absolute value component, which are
modeled separately before the joint forecast is constructed.

82
 We consider various commonly used financial variables as explanatory vari-
ables for forecasting the signs of the one-month U.S. excess stock returns from the
S&P500 index and size-sorted small and large firms stock indices in probit models.
This chapter introduces a model in which the recession forecast constructed for
a binary recession indicator is used as an explanatory variable in the predictive
model. To the best of our knowledge, this kind of approach has not been applied
earlier to forecast the stock return sign. As a motivation for this model, for exam-
ple, Fama and French (1989) and Chen (1991) propose that business conditions
are important determinants of expected stock returns and, therefore, the recession
forecast may be a useful predictive variable in our model. Further, Chauvet and
Potter (2000) have stressed that the stock market “cycle” leads the business cycle.
This argument is based on the fact that the expectations about changes in future
economic activity could have important predictive power to predict excess stock
returns. If there are expectations of a coming recession, excess stock returns are
low, and after a recession period stock returns should be positive.
In this chapter, new dynamic probit models suggested by Kauppi and Saikko-
nen (2008) are employed and further extended. Since there is not much earlier
evidence on suitable explanatory variables in sign prediction with probit models,
various explanatory variables and their in-sample predictive performance are first
evaluated. After that, the out-of-sample directional predictability for the excess
stock return sign is considered. It is not evident, however, how much the in-sample
evidence should be emphasized in assessing overall return predictability because
it does not guarantee out-of-sample predictability, as emphasized in many previ-
ous studies (see discussion, for example, Goyal and Welch, 2008; Campbell and
Thompson, 2008).
The results show that the probit models have statistically significant in-sample
predictive power for the signs of excess stock returns. A proposed new “error correc-
tion” model outperforms the other probit and alternative predictive models, such
as “continuous” ARMAX models, out of sample. The received excess investment
returns over the buy-and-hold trading strategy are economically significant. Com-
parisons between different probit models indicate that the forecasting framework

83
based on the constructed recession forecasts yields more accurate sign predictions
than the models where only financial explanatory variables are employed. Espe-
cially in the case of small and large firms, the excess stock return signs seem to be
predictable also out of sample.
This chapter proceeds as follows. The employed forecasting model with reces-
sion forecasts, suggested dynamic probit models, and in particular, the new error
correction model is presented in Section 4.2. In Section 4.3, the goodness-of-fit
evaluation of sign forecasts and statistical tests for sign predictability are intro-
duced. The empirical evidence on the directional predictability of the U.S. excess
stock returns is reported in Section 4.4. Section 4.5 concludes.

4.2 Forecasting Model

4.2.1 Dynamic Probit Models in Directional Forecasting

Let rt be the excess stock return over the risk-free interest rate. In many studies,
the directional predictability of excess stock returns is examined by using models
for continuous dependent variables. For example, Christoffersen et al. (2007)
proposed a method of forecasting the direction of excess stock return, where they
first model the conditional variance σt2 and the conditional mean µt . Assuming
that the data generating process of rt is

rt = µt + σt εt ,

where εt ∼ IID(0, 1), the conditional probability of a positive return given the
information set Ωt−1 is

Pt−1 (rt > 0) = 1 − Pt−1 (rt ≤ 0)

 −µt 
= 1 − Pt−1 εt ≤
σt
 −µ 
t
= 1 − Fε , (4.1)
σt
where Fε (·) is the cumulative distribution function of the error term εt . If the
conditional probability of positive excess return (4.1) varies with the information
set Ωt−1 , then the sign of the return should be predictable.

84
 In this chapter, the main interest is to study the directional predictability using
probit models, where the dependent variable is the binary sign return indicator

 1, if r > 0,
t
It = (4.2)
 0, if r ≤ 0,
t

which takes the value one when the excess stock return is positive and zero other-
wise. Thus, It is a binary-valued stochastic process. Conditional on the informa-
tion set Ωt−1 , which includes the predictive variables and lagged values of the stock
indicator (4.2), it has a Bernoulli distribution with probability pIt , that is

It |Ωt−1 ∼ B(pIt ).

Let Et−1 denote the conditional expectation given the information set Ωt−1 . In the
probit model the conditional probability of positive excess stock return (It = 1)
satisfies
pIt = Et−1 (It ) = Pt−1 (It = 1) = Pt−1 (rt > 0) = Φ(πtI ), (4.3)

where Φ(·) is the standard normal cumulative distribution function. The condi-
tional probability is modeled by specifying a model for πtI which is supposed to be
a function of variables in the information set.2
Previous literature indicates that there is not much autocorrelation between
the two successive values of excess stock returns. Thus, the benchmark forecasting
model is the static model,
′
πtI = ω + xt−1 β, (4.4)

where the employed explanatory variables are collected in the vector xt−1 . Because
of the expected lack of correlation between It−1 and It , this static model, without
any dynamic structure, might be adequate. In order to investigate this, the value
of the lagged return indicator It−1 can be included in the model. This yields the
dynamic probit model
′
πtI = ω + δ1 It−1 + xt−1 β. (4.5)

If the coefficient δ1 is statistically significant, then the lagged direction of the stock
return is a useful predictor of the future direction of excess stock returns.
2
The superscript “I” in πtI refers to excess stock return sign forecasting.

85
 In the last few years, new binary time series models have been introduced. We
concentrate on the model variants suggested by Kauppi and Saikkonen (2008).
I
They add the lagged value πt−1 , referred to as autoregressive structure, to the
model equation. Thus the static model (4.4) and the dynamic model (4.5) are
extended to the autoregressive model3

′
πtI = ω + α1 πt−1
I
+ xt−1 β (4.6)

and to the dynamic autoregressive model

′
πtI = ω + α1 πt−1
I
+ δ1 It−1 + xt−1 β, (4.7)

respectively. By recursive substitution, and assuming |α1 | < 1, the latter model
can be rewritten as
∞
X ∞
X ∞
X ′
πtI = α1i−1 ω + δ1 α1i−1 It−i + α1i−1 xt−i β. (4.8)
i=1 i=1 i=1

Therefore, if several lagged values of the stock indicator (4.2) or explanatory vari-
ables xt are useful in forecasting, the autoregressive specifications (4.6) and (4.7)
could be useful as parsimonious forecasting models.
Parameters of the probit models (4.4)–(4.7), as well as the case of a new model
presented in the next section, can be estimated by the method of maximum likeli-
hood as described in de Jong and Woutersen (in press) and Kauppi and Saikkonen
(2008). We assume needed regularity conditions, such as the stationarity of πtI , so
that the usual results on maximum likelihood estimation holds.

4.2.2 An Error Correction Model

Based on, for example, the principles of efficient market theory, the lagged values
of the stock indicator (4.2) should not have predictive power to predict future
market directions. This indicates that the estimated coefficient of the lagged return
indicator δ1 may be zero or close to zero. Therefore, in the dynamic autoregressive
model (4.7), if δ1 = 0 and there are no explanatory variables in the model, that is
3
In this chapter, the same model attributes as Kauppi and Saikkonen (2008) are used.

86
β = 0, then the autoregressive parameter α1 is not identified as seen from equation
(4.8) by imposing the above-mentioned restrictions.4 Even if the coefficient δ1
is just close to zero, there is a potential identification problem that can affect
parameter estimation and have implications on the forecasting accuracy of excess
return sign predictions.
Imposing the restriction δ1 = 1 − α1 in the unrestricted dynamic autoregressive
model (4.7) and assuming |α1 | < 1, a new “restricted” dynamic autoregressive
model can be formulated as

′
πtI = ω + α1 πt−1
I
+ (1 − α1 )It−1 + xt−1 β. (4.9)

Because of the assumption |α1 | < 1, the coefficient for the lagged return indicator
It−1 , 1 − α1 , is always positive. In model (4.9), α1 can also be interpreted as a
I
“weight” between πt−1 and It−1 . It is expected that in our application α1 will be
positive and quite high (since δ1 ≈ 0). This leads to the fact that the predictive
power is distributed over the longer history of It and explanatory variables xt . On
the other hand, if α1 is “small”, then the first lag It−1 is more useful than in the
case of a higher value of α1 .
For simplicity, we refer to model (4.9) as an “error correction” model (ecm) in
I
this study. The reason is that by adding −πt−1 to both sides of equation (4.9), we
obtain the error correction form
′
∆πtI = ω + (1 − α1 )(It−1 − πt−1
I
) + xt−1 β, (4.10)

where ∆πtI = πtI − πt−1

I I
. Thus, the difference between It−1 and πt−1 measures
the long-run relationship between the value of the stock return indicator and the
I
transformed probability πt−1 = Φ−1 (pIt−1 ) in the probit model. Rewriting model
(4.10) as
′
πtI = ω + πt−1
I I
+ (1 − α1 )(It−1 − πt−1 ) + xt−1 β,

it can be seen that, for α1 close to one, the model can be expected to exhibit “near
unit root” behavior, implying rather strong persistence for the variable πtI . This
4
Note that when explanatory variables xt−1 are included in the model and β 6= 0, then there
is no such identification problem, even though δ1 = 0.

87
means that the conditional probability of positive excess stock return does not
change much between successive time periods.
As seen from equation (4.9), the parameter 1 − α1 is always positive, and it
I
can be interpreted as the proportion of the disequilibrium between It−1 and πt−1 in
I
period t−1. A positive value of the error correction term (It−1 −πt−1 ) increases the
probability of positive excess stock return in the following period and, of course,
vice versa if the the error correction term is negative. Later on, we will see that
the conditional probability of positive excess stock return pIt is typically close to
0.50 in most models, which means that πtI is close to zero.
It is worth noting that the error correction model (4.9) is somewhat the same as
the autoregressive conditional multinomial model (ACM) suggested by Russell and
I I
Engle (2005). In their model, the term (It−1 −πt−1 ) is replaced by (It−1 −Φ(πt−1 )).
The model (4.9) without the term x′t−1 β is also similar to the IGARCH model,
suggested by Engle and Bollerslev (1986), for conditional heteroskedasticity in
models for continuous variables.

4.2.3 Recession Forecast as an Explanatory Variable

A novel idea of this study is to consider whether recession forecasts have explana-
tory power to forecast the direction of excess stock returns. In the empirical fi-
nance literature, it is shown that as forward-looking variables, lagged stock returns
should provide information about the future evolution of economic activity and po-
tential recession periods (see, e.g., Pesaran and Timmermann, 1995; Estrella and
Mishkin, 1998; Nyberg, 2010). Therefore, if the expectations of future economic
activity are correct, the movements of the stock market should lead movements in
economic activity (see, e.g., Fama 1990). Theoretically, this relation can be justi-
fied by present value or discounted-cash-flow models, where the price of a stock is
equal to expected future dividends which are assumed to be related to the future
economic activity and the profitability of firms.
Our main goal is to forecast recession periods and use the potential explanatory
power of the obtained recession forecasts to make better forecasts for the sign of

88
the excess stock returns. This is done by using the binary recession indicator

 1, if the economy is in a recession at month t,
yt = (4.11)
 0, if the economy is in an expansion at month t.

In this chapter, recession dates defined by the NBER are used. As Chauvet and
Potter (2000) argue, one feature, but also a potential problem, with the NBER
recession dates is that they do not reflect short-lived contraction periods in the
economy, which could have notable explanatory power for predicting excess stock
returns. Further, Chauvet and Potter (2000) construct the transition probabilities
of the “bear” and “bull” state of the stock market using Markov chain methods.
They find that bear markets generally start a couple of months before an economic
slowdown or recession period and end some months before a recession period ends.
Thus, it seems evident that movements in the stock market should lead the business
cycle. Evidence of this kind can be seen in Figure 4.1.5 The U.S excess stock
returns are often negative before a recession period begins. On the other hand,
it seems that the returns are positive in the last few recession months, indicating
expectations about recovery in economic activity.
According to this idea, for example, in the general dynamic autoregressive
model (4.7), the estimated recession forecast pyt+5 , constructed using the model
(4.12), may be included in the vector xt−1 = (pyt+5
′
x̃t−1 ) , where x̃t−1 contains
other financial explanatory variables. Therefore, a predictive probit model contains
the fitted values of the binary explanatory recession indicator (4.11) (cf. Maddala
1983, 122–123). Parameter estimation and forecasting is carried out with a two-
step procedure where the recession and the stock return sign prediction models are
estimated separately. It is worth noting, however, that in this kind of model, the
usual asymptotic distribution of the maximum likelihood estimate may not apply
because the recession probability forecast included in the model is based on the
estimated model (cf. Pagan, 1984).
The forecast horizon in recession forecasting is assumed to be six months. A
six-month recession forecast for the value of the recession indicator (4.11) at time
5
Details on dataset are given in Table 4.1.

89
 15

10

Excess stock return rt 0

−5

−10

−15

−20

−25

−30
1965 1970 1975 1980 1985 1990 1995 2000 2005 2010
Time

Figure 4.1: U.S. excess stock returns rt and the NBER recession periods yt (shaded
areas) for a sample period of 1968 M1–2006 M12.

t + 5, based on the information set Ωt−1 at time t − 1, is the conditional probability

y
Et−1 (yt+5 ) = Pt−1 (yt+5 = 1) = Φ(πt+5 ) = pyt+5 .

In recession forecasting, an autoregressive probit model

y y
πt+5 = c + φπt+4 + z ′t−1 b (4.12)

is employed where, according to the findings in the recession forecasting literature,

domestic and foreign term spread and lagged nominal stock return are used as
predictors. The values of these variables are included in the vector z t−1 , that is
 ′
US n GE
z t−1 = SPt−1 rt−1 SPt−1 .

The usefulness of the domestic term spread (SPtU S ), defined as the spread between
the long-term and the short-term interest rates, to predict recession periods is
demonstrated in many studies (see, among others, Estrella and Mishkin, 1998;
Estrella, 2005). Using dynamic probit models, Nyberg (2010) also suggests that the
foreign term spread (SPtGE , the term spread of Germany) and stock market returns

90
(rtn ) can be used to forecast coming recession periods (see also, e.g., Bernard and
6
Gerlach, 1998; Estrella and Mishkin, 1998). In that study, the obtained recession
forecasts were quite accurate for at least six months ahead and the probit models
with the autoregressive structure (see model (4.2.3)) yield the best out-of-sample
forecasts compared with various other probit models. Therefore, in this study, we
take model (4.2.3) and the six-month-ahead forecast horizon as a given.
An advantage in model (4.12) is that it does not contain a lagged value of the
recession indicator (4.11). It is important to take into account that it takes several
months before, for example NBER, can be sure what the state of the economy really
is. Hence, the values of the recession indicator are known with a considerable delay,
which indicates that it is computationally easier to construct recession forecasts
without using the lagged values of the binary recession indicator (4.11) in an
estimated model. This is based on the fact that in model (4.12), no multiperiod
iterative forecasts (see Kauppi and Saikkonen, 2008) for the recession indicator are
made because all predictive power comes from the employed explanatory variables
z t . Thus, it is not needed to specify the assumed “publication lag” in the known
values of yt exactly.

4.3 Evaluation of Forecasts

4.3.1 Statistical and Economic Goodness-of-Fit Measures

Both in-sample and out-of-sample performance of predictive models are evaluated

with frequently used goodness-of-fit measures. One is Estrella’s (1998) pseudo-R2
measure
psR2 = 1 − (ˆlu /ˆlc )−(2/T )l̂c , (4.13)

where ˆlu is the maximum value of the estimated unconstrained log-likelihood func-
tion and ˆlc is its constrained counterpart in a model which only contains a constant
term. This measure takes on values between 0 and 1, and it can be interpreted
in the same way as the coefficient of determination in linear models. The value of
6
Further information on the explanatory variables is in Table 4.1 in Section 4.4.1.

91
the maximized log-likelihood function also enables the comparison of model per-
formance using model selection criteria such as the Schwarz information criterion
BIC (Schwarz, 1978).
The binary nature of the dependent variable leads to the question of what the
percentage of correct “matches” is of the realized values and the forecasts of the
stock indicator. This ratio is denoted by CR. By the hypothesis of no predictability
in excess stock return signs, the estimated value of CR should be close to 0.50,
which means that the employed model is unable to forecast the future market
directions correctly. It is desirable to specify a threshold value that translates
the probability forecasts into forecasting signals. The most commonly used and
natural threshold choice is 0.50, which is also used in this case.
For financial analysts and investors, the most important model evaluation cri-
terion is the return on their investment. There are many different kinds of trading
strategies that can be applied. Here a simple trading simulation similar to that in
Leung et al. (2000) is used. At the beginning of each month, the investor makes
an asset allocation decision. She can shift her assets either into stocks or into
risk-free Treasury Bills, and the money that has been invested in either of these
alternatives remains there until the next decision date. In this trading strategy,
the mentioned 50 percent threshold value is used. Then, the portfolio consists
of interest rate investment in Treasury Bills (Itf = 0), if pIt ≤ 0.50, and stocks
(Itf = 1), if pIt > 0.50. Here the superscript f refers to forecast.7
In this trading simulation, transaction costs are also taken into account. Fol-
lowing Granger and Pesaran (2000), the marginal cost of transaction for asset
allocation changes between stocks and interest rates will be denoted by ζs and
ζb , respectively. This means that, every time the asset allocation changes, the
amount of the transaction cost is subtracted from the final investment return. In
this study, the “low cost scenario” suggested by Pesaran and Timmermann (1995),
where ζs = 0.005 and ζb = 0.001, is applied. For example, when the risk-free
7
An alternative method for parameter estimation and determining the cutoff threshold, which
is assumed to be 50 percent in the above, is for example, the Maximum Utility estimation method
proposed by Elliott and Lieli (2007).

92
interest rate investments are switched to stocks, 0.50% of the whole amount of the
portfolio value is lost.
As Granger and Pesaran (2000) have shown, it is also possible to form non-
constant “payoff” probability ratios of switches between stocks and interest rates
as an alternative for this 50% threshold. However, in this study, these payoff ratios
are not very useful because, according to these threshold ratios, the asset allocation
decision is to stick to stocks almost all the time. Therefore, probability forecasts,
even if accurate, have little economic value.
When considering the predictability of excess stock return signs with different
trading rules, one important evaluation criterion is the overall portfolio return, de-
noted by RET . Nevertheless, as for example, Hong and Chung (2003) emphasize,
it is also worth considering risk-adjusted returns because different trading rules
involve different levels of risk. In this evaluation, one commonly used measure is
the Sharpe ratio (Sharpe, 1966 and 1994)
k rf
RET − RET
SR = , (4.14)
σ̂ k
k
where RET is the average portfolio return based on the model and trading rule
rf
k, RET is the average risk-free portfolio return (bond investment strategy), and
σ̂ k is the sample standard deviation of portfolio returns RET k . The higher the
Sharpe ratio is, the higher the return and the lower the volatility. Portfolios with
a high Sharpe ratio are preferable to those with a low Sharpe ratio.

4.3.2 Testing the Statistical Predictability

For the evaluation of the directional forecasting performance and market timing,
a test proposed by Pesaran and Timmermann (1992) is available. The null hy-
pothesis of this test is that the value of the correct prediction ratio, CR, does
not differ statistically significantly from the ratio that would be obtained in the
case of non-predictability, where the forecasts and the realized values of the return
indicator It are independent. Granger and Pesaran (2000) show that the market

93
timing test can be based on the test statistic
√
m KS
PT =  1/2 . (4.15)
P̄I (1−P̄I )
¯
I(1− ¯
I)

Here KS is the Kuipers score KS = HR − F R between the “hit rate”

Iˆuu
HR =
Iˆuu + Iˆdu
and the “false rate”
Iˆud
FR = ,
Iˆud + Iˆdd
where forecast classification is denoted by
m m
Iˆuu = = 1, It = 1), Iˆud =
X X
1(Itf 1(Itf = 1, It = 0),
t=1 t=1
m m
Iˆdu = 1(Itf = 0, It = 1), Iˆdd =
X X
1(Itf = 0, It = 0),
t=1 t=1

where f refers to forecast, u to an “up” signal (It = 1) and d to a “down” signal

(It = 0), and 1(·) is an indicator function. Furthermore, in the test statistic (4.15),
I¯ is the sample average of the sign indicator It values in the m-month sample period
¯
and P̄I = IHR ¯ R. Under the null hypothesis of non-predictability, the
+ (1 − I)F
P T test statistic has an asymptotic standard normal distribution.
The directional predictability of an underlying data generation process is, how-
ever, not the same thing as a successful trading strategy. To evaluate the forecasts
of the best forecasting models, we test the significance of the differences between
the investment returns on the best models and trading strategies. This is tested
by means of the Diebold-Mariano test (1995). Because the forecast horizon is one
month, h = 1, the test statistic is
√ ¯
md
DM = p , (4.16)
¯
var(d)

where d¯ is the average difference between the predicted excess returns of the consi-
dered models. As in the P T test, under the null hypothesis of equal forecast
accuracy, the DM statistic also has an asymptotic N(0, 1) distribution.

94
4.4 Empirical Results

4.4.1 Data and Previous Findings

The monthly data set contains financial variables which have been used to predict
overall level and the direction of excess stock returns in the previous literature.
The data set covers the period from January 1968 to December 2006, and it is
obtained from different sources mentioned in Table 4.1. The first 12 observations
are used as initial values. The total number of observations, T , is 468. In out-of-
sample forecasting, the data set is divided into two subsamples: the estimation and
the forecasting sample. The first out-of-sample forecasts will be made for January
1989 and the last for December 2006.
In out-of-sample forecasting, parameters are estimated recursively using an ex-
panding window of observations, where models are estimated using data from the
start of the data set through to the present time to obtain a new one-period fore-
cast. This procedure is repeated until the end of the forecasting sample. The use
of an alternative rolling estimation window is problematic, because there are not
many recession periods in the post-1970 time period and there would be estimation
samples with no recession periods at all.
The one-month excess stock return is defined as the continuously compounded
return of the price index Pt minus the risk-free interest rate rft
 P 
t
rt = 100 log − rft . (4.17)
Pt−1
Here Pt is the value of the S&P 500 stock index and the one-month risk-free re-
turn rft is approximated by the three-month U.S. Treasury Bill rate it . With
excess stock returns rt , the values of the binary stock return indicator described
in equation (4.2) can be constructed.
Several explanatory variables to forecast the direction of excess stock returns
will be considered. As confirmed by Leung et al. (2000), the majority of useful
information for forecasting stock returns is contained in interest rates and lagged
stock returns. Hence, the financial explanatory variables that are considered in the
predictive models are the short-term and long-term interest rates and their first

95
 Table 4.1: Data set of dependent and explanatory variables.
Variable Description
Pt Standard&Poors 500 U.S. stock index
PtS CRSP small size firms index, first decile
PtL CRSP large size firms index, tenth decile
rt , rtS , rtL One-month excess return over the risk-free return (see (4.17))
rtn One-month nominal stock return from the S&P 500 index
yt U.S. Recession periods (NBER)
it Three-month U.S. Treasury Bill rate, secondary market
Rt 10-year U.S. Treasury Bond rate, constant maturity
∆it , ∆Rt First differences of it and Rt
SPtUS U.S. term spread between Rt − it
SPtGE German term spread between German long-term and short-term interest rates
σt Sum of squared daily stock returns in the S&P500 index within one month
DPt Dividends over the past year divided by the current stock index value, DPt = Dt /Pt
EPt Earnings over the past year divided by the current stock index value, EPt = Et /Pt
Notes: The sample period is 1968 M1–2006 M12. Monthly and daily S&P500 index series are taken from
http://finance.yahoo.com and http://www.econstats.com. Size-sorted CRSP indices are obtained from the
Kenneth French Data Library (http://mba.tuck.dartmouth.edu/pages/faculty/ken.french/data_library.html).
Interest rates are from http://www.federalreserve.gov/releases/h15/data.htm. German term spread is
constructed as the difference between 10-year Federal security (series WZ9826, the missing values between 1971
M1–1972 M9 are replaced by the OECD 10-year interest rate) and the three-month money market rate (series
su0107, see http://www.bundesbank.de/statistik/statistik.). Data for log-dividends Dt and log-earnings Et were
obtained from the homepage of Robert Shiller’s book Irrational exuberance
(http://www.irrationalexuberance.com) [5 January 2009].

differences, the U.S term spread, earnings/price and dividends/price-variables, and

the realized volatility (see Table 4.1).
In previous studies, both the lagged excess stock returns and the lagged values
of the return indicator (4.2) are used as predictors. Leung et al. (2000) use first
differences of interest rates and lagged excess stock returns in their comparison
between the sign and the overall return forecasting models, and concluded that in
probit and logit models, several past returns should be included in the model. If
the explanatory power is distributed among many lags of past returns, then the
autoregressive models (4.6) and (4.7) could be useful in forecasting. On the other

96
hand, Anatolyev and Gospodinov (2010) use the lagged sign return indicator It−1
in their dynamic logit model for the direction of the future excess stock return.
The corresponding estimated regression coefficient was positive, but statistically
insignificant.
Interest rate spreads between different maturities may offer information about
future expectations in financial markets (see, for instance, Fama and French, 1989).
In recession forecasting, the term spread (SPtU S ) is expected to transmit the ex-
pectations for future monetary policy. The lower is the difference between the
long-term and short-term interest rates the more restrictive is the current mone-
tary policy. The term spread could also have its own impact on the stock market,
not only on the real economic activity.
Dividends (Dt ) and earnings (Et ), divided by the value of the price index
(Pt ), have been among the most commonly used explanatory variables (see, e.g.,
Campbell and Shiller, 1988; Cochrane, 1997). The dividend-price (earnings-price)
ratio is computed with the dividends (earnings) of S&P 500 stock index companies
over the past year. Since the monthly data of dividends and earnings are not
available, DPt and EPt are constructed as sums of dividends and earnings over
the past year divided by the current, monthly price level Pt .
Numerous studies have also documented a notable dependence of stock return
and stock return volatility with important implications for asset pricing. The
realized monthly volatility (σt ), based on the sum of squared daily observations
within one month (see Christoffersen et al., 2007), is examined as a predictor in
probit models.

4.4.2 In-Sample Results

Even though our main interest lies in the out-of-sample predictions of the direction
of future excess stock returns, first, the in-sample performance of different probit
models and combinations of explanatory variables were experimented using the
sample period from January 1968 until December 1988. Explanatory variables are

97
included one by one in the model.8
The main results and findings are as follows. According to the psR2 and CR
values, and the returns of trading strategies in the chosen in-sample period, the re-
cession forecast and the first difference of the short-term interest rate are the best
predictive variables. When employing these variables, there seems to be evidence
that the excess stock return signs are predictable in sample. The first difference
of the long-term interest rate and the realized volatility also have some predic-
tive power. Interestingly, the corporate earning and dividend variables, used in
many previous studies, are not particularly useful predictors.9 When the recession
forecast is employed with different financial explanatory variables in the model
equation, the evidence is very much the same as in the above. The first difference
of the short-term interest rate appears to be the best predictor with the recession
forecast in-sample.
Table 4.2 presents details of parameter estimates in different probit models
when the recession forecast (pyt+5 ) and the first difference of the short-term inter-
est rate (∆it−1 ) are used as explanatory variables. The robust standard errors
suggested by Kauppi and Saikkonen (2008) are presented. However, it should be
noted that these standard errors may be inaccurate because the estimated recession
forecast is employed in the model.10
It seems that the lagged stock indicator It−1 has no statistically significant
predictive ability for the sign of the stock return. In the error correction model,
I
the autoregressive coefficient πt−1 is clearly statistically significant, but in other
dynamic models, it is not. As expected, the estimated coefficients of the recession
8
Matlab 7.4.0 and its BFGS optimization routine in the Optimization Toolbox is used in
estimation and forecast computation. In models with the autoregressive structure (models (4.6),
(4.7) and (4.9)) the initial value π0 is set to similar way as suggested by Kauppi and Saikkonen
(2008). For example, in model (4.7), it is π0 = (ω + δ1 I¯t−1 + x̄t−1 β)/(1 − α1 ), where a bar
denotes the sample mean of It and xt−1 .
9
Further information on in-sample performance of the different models is available upon
request.
10
In addition, there are no formal proofs of the asymptotic distributions of the maximum
likelihood estimator in models (4.6), (4.7) and (4.9) presently available.

98
forecast and the first difference of the short-term interest rate are both negative.

Table 4.2: Estimation results of in-sample predictive models.

static dynamic auto. dyn.auto. ecm
model(4.4) model(4.5) model(4.6) model(4.7) model(4.9)
constant 0.07 0.02 0.07 0.03 -0.07
(0.10) (0.12) (0.09) (0.13) (0.03)
I
πt−1 0.04 -0.03 0.85
(0.24) (0.28) (0.05)
It−1 0.08 0.08
(0.11) (0.19)
∆it−1 -0.24 -0.30 -0.30 -0.29 -0.16
(0.04) (0.12) (0.12) (0.12) (0.07)
pyt+5 -0.50 -0.47 -0.49 -0.49 -0.02
(0.25) (0.25) (0.24) (0.25) (0.05)
log-L -161.33 -161.22 -160.72 -160.64 -162.10
psR2 0.041 0.041 0.046 0.046 0.034
BIC 169.55 172.18 171.68 174.34 173.06
CR 0.580 0.591 0.579 0.579 0.579
RET 10.50 9.98 9.88 9.57 8.12
SR 0.95 0.86 0.79 0.72 0.25
PT 0.003 0.002 0.006 0.006 0.008
DM 0.009 0.017 0.017 0.023 0.123
DMra 0.000 0.000 0.000 0.000 0.003
DM It =0 0.000 0.000 0.000 0.000 0.000
It =0
DMra 0.000 0.000 0.000 0.000 0.000
Notes: The models are estimated using the in-sample data from 1969 M1 to 1988 M12. Robust
standard errors, given in parentheses, are computed with procedures suggested by Kauppi and
Saikkonen (2008). RET is the average annualized in-sample portfolio return in the considered
model and SR is the corresponding Sharpe ratio (4.14). The in-sample return in the B&H
strategy RET is 3.60%. The corresponding Sharpe ratio is negative because the average return
rf
in the pure risk-free interest rate investment strategy is higher (RET = 7.16%). The p-values
of the market timing test (4.15) and the Diebold and Mariano (1995) test (4.16) are reported.
In DM tests, the buy-and-hold trading strategy is the benchmark. Further, ra means the
risk-adjusted returns, where returns are standardized with the standard deviation of the
returns. The values of test statistics DM It =0 and DMra
It =0
are obtained when only months
with negative excess stock return (It = 0) are considered.

99
In this case the recession forecast is not statistically significant in the error cor-
rection model, but the first difference of the short-term interest rate is. In other
models, both of these predictors are statistically significant according to presented
robust standard errors.
Figure 4.2 depicts the estimated probability of a positive excess stock return
in the static model (4.4) and in the error correction model (4.9), whose estimation
results are shown in Table 4.2 (the first and the fifth model). Both models seem
to give roughly the same in-sample predictions. In recession periods, both models
suggest investing in a risk-free interest rate. More or less the only significant
difference between the models is the time period between approximately 1976
to 1979. At that time, the probability forecast in the static model is typically
above the 0.50 threshold value, while in the error correction model it is below the
threshold.

0.8

0.7

0.6

0.5
Probability

0.4

0.3

0.2

0.1
1967 1970 1972 1975 1977 1980 1982 1985 1987 1990
Time

Figure 4.2: In-sample probabilities pIt (see (4.3)) of the static model (4.4) (solid
line) and the error correction model (4.9) (dashed line). The 50 percent threshold
is also depicted.

As the psR2 values in Table 4.2 indicate, the statistical predictive power for
the sign of the excess stock return is, as expected, quite low. Although the statis-

100
tical predictability is weak, the portfolio investment performance yields evidence
about a useful sign predictability in excess stock returns. The average rates of
returns for different models and trading strategies are higher than in the “passive”
buy-and-hold trading strategy (hereafter B&H strategy), where one is investing
only in stocks. This annualized benchmark return is 3.60%. In the best models,
the returns, including the transaction costs, are between 9.50% to 10.50%. The
presented error correction probit model seems to yield smaller in-sample returns
than its counterparts.
The statistical significance of return differences between an examined model
and B&H returns was tested. Table 4.2 presents the values of the statistical test
statistics introduced in Section 4.3.2. Since we are only interested in cases where
the proportion of correctly predicted signs and the portfolio returns in estimated
models are higher than under the null hypothesis of no predictability, only the
positive and statistically significant values of the P T and DM test statistics (see
(4.15) and (4.16)) provide evidence of predictability. The values of the market
timing test statistic P T are statistically significant at the 1% level under all ex-
perimented models in Table 4.2. Thus the null hypothesis of no predictability is
rejected, providing in-sample evidence that excess return signs are predictable. In
the DM tests, the null hypothesis of equal performance between the returns in the
considered model and the B&H strategy are rejected in all models at a 5% level
except in the error correction model where the p-value of the test statistic is 0.123.
When risk-adjusted returns are considered, the DMra test statistics are statisti-
cally significant in all models, providing evidence of profitable trading strategies
based on the forecasts from the probit models.
Although the unrestricted dynamic autoregressive probit model (4.7) gives a
better in-sample fit than the error correction model in the models presented in
Table 4.2, in some other models with different explanatory variables, the error
correction model gives higher psR2 and CR values. In contrast to the error cor-
rection model (4.9), in many other unrestricted dynamic autoregressive models,
the autoregressive coefficient α1 is typically negative. Thus, if the probability of
a positive excess stock return has been high in some period, it tends to be lower

101
in the next period. As an example, consider a model in which the first difference
of the short-term interest rate (∆it−1 ) is the only explanatory variable in xt−1 .
Figure 4.3 shows the estimated probabilities of positive excess stock returns in the
unrestricted dynamic autoregressive model (4.7) and in the error correction model
(4.9) in this example case. As the estimate for the autoregressive coefficient α1 is
negative in model (4.7), the probability of excess stock return fluctuates heavily
around the threshold value 0.50. On the other hand, in the right panel with a posi-
tive and high estimate of α1 , the probability forecasts follow a relatively persistent
swing. Therefore, it seems that the error correction model yields less transactions
between stocks and bonds, and consequently also less transaction costs, than the
unrestricted dynamic autoregressive model. This is particularly striking in mod-
els presented in Figure 4.3 and could be an important property in out-of-sample
forecasting.

0.9

0.8

0.7

0.6
Probability

0.5

0.4

0.3

0.2

0.1
1967 1970 1972 1975 1977 1980 1982 1985 1987 1990
Time

Figure 4.3: In-sample predictions of the dynamic autoregressive model (4.7) (solid
line) and the error correction model (4.9) (dashed line) when ∆it−1 is the only
predictive variable.

102
4.4.3 Out-of-Sample Results

When forecasting the signs of excess stock returns, it is important to compare

the different models out of sample. Previous results on the predictive models for
the overall excess stock return suggest that the in-sample predictability does not
necessarily imply out-of-sample predictability. For example, Han (2007) finds that
a statistically superior predictive VAR-GARCH model in sample does not consis-
tently outperform its competitors in terms of portfolio investment returns out of
sample. Goyal and Welch (2008) argue that traditional predictive models for the
excess return cannot beat the historical average return out of sample and that
there is no single variable that has theoretically meaningful and robust explana-
tory power. On the other hand, Campbell and Thompson (2008) show that some
predictors perform better than the historical average when restrictions on regres-
sion coefficients are imposed. They and, for instance, Anatolyev and Gospodinov
(2010) have stressed that while the out-of-sample predictive power is small, it can
be utilized in market timing decisions to earn economically higher excess stock
returns than the B&H strategy even out of sample.
In this study, the out-of-sample period consists of 216 months from January
1989 to December 2006. The out-of-sample recession forecast pyt+5 is constructed
before making any sign forecasts for excess stock return signs. As described in
Section 4.2.3, the parameters in the sign and in the recession prediction models
are estimated recursively. Table 4.3 shows the out-of-sample performance of the
best in-sample predictive models and also some other probit models. The idea is
to compare the predictive performance of different models when the same combi-
nations of explanatory variables, which turned out to be the best out-of-sample
predictive variables, are examined.11
According to commonly used statistical model evaluation measures, there is not
much out-of-sample predictability in excess stock return signs. The values of psR2
measures (see (4.13)) are, even in the best models, small or even negative.12 The
11
The models with other financial variables, presented in Table 4.1, were also considered and
the results of those models are available upon request.
12
Negative psR2 value means very poor out-of-sample forecasting performance (Estrella, 1998).

103
 Table 4.3: Out-of-sample performance of different probit models.
model xt−1 psR2 CR RET SR
B&H 8.07 0.79
static (4.4) pyt+5 0.015 0.593 7.63 0.75
static (4.4) ∆Rt−1 , pyt+5 0.013 0.588 8.02 0.86
static (4.4) ∆it−1 , pyt+5 0.014 0.588 7.02 0.61
dynamic (4.5) – neg. 0.528 3.92 neg.
dynamic (4.5) pyt+5 0.005 0.569 6.16 0.46
dynamic (4.5) ∆Rt−1 , pyt+5 0.002 0.574 7.32 0.76
dynamic (4.5) ∆it−1 , pyt+5 0.004 0.579 7.46 0.79
auto (4.6) pyt+5 neg. 0.514 2.30 neg.
auto (4.6) ∆Rt−1 , pyt+5 0.015 0.583 7.20 0.68
auto (4.6) ∆it−1 , pyt+5 0.014 0.579 6.53 0.50
dyn.auto (4.7) – neg. 0.514 4.11 neg.
dyn.auto (4.7) pyt+5 0.008 0.574 5.42 0.28
dyn.auto (4.7) ∆Rt−1 , pyt+5 0.006 0.565 7.03 0.63
dyn.auto (4.7) ∆it−1 , pyt+5 0.011 0.565 5.46 0.27
ecm (4.9) – 0.014 0.588 8.62 0.97
ecm (4.9) pyt+5 0.018 0.606 10.33 1.46
ecm (4.9) ∆Rt−1 , pyt+5 0.016 0.588 9.09 1.12
ecm (4.9) ∆it−1 , pyt+5 0.017 0.588 9.78 1.28
Notes: See also notes to Table 4.2. The average return of the buy-and-hold trading strategy
(B&H) is 8.07% (annual) with the corresponding Sharpe ratio SR= 0.79. The risk-free return
on interest rate investments is 4.21%. The note “neg” means a negative psR2 value and “–” in
xt−1 indicates that there are no explanatory variables in the model. Note that the transaction
costs are also taken into account in RET .

percentage of correct forecasts, CR, vary between 0.51 and 0.61. Contrary to the
employed statistical measures, the results of portfolio returns RET and Sharpe
ratios SR exhibit evidence of useful predictability for asset allocation decisions
even though average portfolio returns vary strongly between different models. As
in the in-sample evidence, the models with recession forecasts generate better
sign forecasts than models without these forecasts. It is worth noting that the
sign prediction models containing the constructed recession forecast outperform
the models including the variables used in recession forecasting, especially out of

104
sample.
It is interesting that the error correction model (4.9) clearly outperforms the
corresponding unrestricted dynamic autoregressive model (4.7) out of sample. As
seen in the best in-sample models in Table 4.2, the autoregressive model (4.6) and
the dynamic autoregressive model (4.7) outperform the error correction model
(4.9), but the out-of-sample evidence seems to be very different. In Table 4.3, the
psR2 values of the error correction models are clearly positive and the ratios of
correct predictions, CR, are higher than in the other probit models considered.
Above all, error correction models can generate more profitable trading strategies
than the other probit models. Perhaps the most striking finding is the performance
of the model with no explanatory variables (“–” in Table 4.3). The psR2 values,
CR ratio, average excess returns, and Sharpe ratios are clearly higher in the error
correction model. A potential identification problem in the dynamic autoregressive
model (4.7) discussed in Section 4.2.2 is a possible explanation for this superior
out-of-sample performance of the error correction model.
Overall, compared with the dynamic models (4.5)–(4.7), the static probit model
I
(4.4), without the autoregressive model structure πt−1 or the lagged It−1 , seems to
be an adequate model for the excess stock return sign. The error correction model
(4.9) appears to be the only dynamic model which yields better forecasts in this
data set than the static model.
The recession forecast is the main predictive variable in different models. The
first differences of the short-term and the long-term interest rates are also fairly
good predictors almost in all probit models and perform consistently better than
the other financial explanatory variables examined. For instance, the realized
volatility σt was a quite good predictive variable in sample, but its out-of-sample
performance is very poor in probit models.
Figure 4.4 depicts the out-of-sample probability forecasts for the positive excess
stock return in two models presented in Table 4.4. The most notable difference is
that in 2001–2003 the error correction model gives a signal to invest in a risk-free
interest rate when the monthly stock returns are most of the time negative. Fur-

105
ther, in Table 4.4, the values of the P T and DM test statistics in these best error
correction and static models, in terms of investment return (RET ), are shown. In
the error correction model, the p-value of the P T test statistic is 0.053 and the
p-value of the DM test statistic is 0.121. The risk-adjusted returns are statistically
significantly higher than the returns in the B&H strategy (p-value 0.000). On the
other hand, the p-values of test statistics in the best static probit model shows
that the excess stock return signs are not predictable with this model.

0.7

0.65

0.6

0.55
Probability

0.5

0.45

0.4

0.35
1987 1990 1992 1995 1997 2000 2002 2005 2007
Time

Figure 4.4: Out-of-sample predictions of the static (4.4) model with ∆Rt−1 and
pyt+5 (RET = 8.02%, solid line). The dashed line is the error correction model (4.9)
with pyt+5 (RET =10.33%).

It should be pointed out that all models mostly suggest investing in stocks.
For instance, in Figure 4.4, the conditional probabilities of a positive excess stock
return are typically above the 0.50 threshold. Thus, the return differences between
probit models and the B&H trading strategy are zero in most months. In addition,
because the probability of recession is principally almost zero when the economy
is in an expansionary state, the recession forecast should be a particularly useful
predictor of negative excess stock return months when the economic activity is
declining. Hence, the values of the DM test statistics are also calculated based on

106
only those months when the excess stock returns have been non-positive (that is
It = 0). There are 86 months with a negative excess return in the out-of-sample
period. In Table 4.4, the values of test statistics DM It =0 , and DMra
It =0
in the case
of risk-adjusted returns, are strongly statistically significant in the best models.
As seen, the in-sample results in Table 4.2 are similar.

Table 4.4: Statistical tests for the best error correction probit model and the best
static probit model.
model xt−1 CR RET SR PT DM DMra DM It =0 It =0
DMra
ecm (4.9) pyt+5 0.61 10.33 1.46 0.053 0.121 0.011 0.000 0.000
static (4.4) pyt+5 , ∆Rt−1 0.59 8.02 0.86 0.326 0.616 0.184 0.000 0.000
Notes: The best error correction model (4.9) (ecm) and the best static model (4.4), reported in
Table 4.3 and depicted in Figure 4.4, are presented. The p-values of the P T and DM tests are
reported. In the DM tests the B&H trading strategy is the alternative asset allocation
strategy. In the table, ra means the risk-adjusted returns and the test statistics DM It =0 and
It =0
DMra are obtained when only months with negative excess stock return (It = 0) are taken
into account.

Furthermore, when the investment returns of the best error correction model
and the best static model are compared, the p-value of the DM test statistic is
0.111 and 0.048, respectively, when the risk-adjusted returns are considered. Thus,
the error correction model yields higher returns, but the statistical significance
between return differences is relatively weak in the considered out-of-sample peri-
od. However, according to the “asymmetric” DM test statistics discussed above
(DM It =0 and DMra
It =0
), the best error correction model outperforms the best static
model on all traditional statistical significance levels.

4.4.4 Comparison Between Probit and Alternative Predic-

tive Models

It is interesting to make some comparisons between probit models and alternative

models, such as ARMAX models and models where forecasts for the asset return
volatility are employed to produce sign forecasts for excess stock returns. In fact,
there are not many previous studies that compare the predictive performance of

107
these models. Leung et al. (2000) find some evidence that qualitative response
models, including logit and probit models, outperform models for the continuous
dependent variables in their out-of-sample forecasting. They considered a sample
of U.S., U.K. and Japanese stock indices from January 1991 to December 1995. In
their study, the ratios of correct sign predictions and the investment returns from
qualitative dependent models are higher than in models for continuous variables.
In ARMAX models the same explanatory variables as in probit models are
considered. The dependent variable is the excess stock return rt and it is assumed
that a positive forecast gives the signal to buy stocks (i.e. Itf = 1). This is
consistent with the definition of the stock return indicator (4.2). As in probit
models, the in-sample predictive performance of different ARMAX models is first
analyzed.13 The estimated values of the BIC model selection criterion (Schwarz,
1978) suggested an ARMAX(2,0)14 model with the first difference of the short-term
interest rate (∆it−1 ) and the recession forecast (pyt+5 ) as explanatory variables. An
ARMAX(1,0) model with the recession forecast and the U.S. term spread (SPtU S )
generates the highest in-sample investment return.
Out-of-sample forecasting performance of the best in-sample models and some
other ARMAX models are shown in Table 4.5. As a whole, the percentage of the
correct forecasts among the best probit models appears to be somewhat higher than
in the best ARMAX models, but the investment return performance in particular
is clearly better among the best probit models. In Table 4.5, only the best two
ARMAX models in terms of RET values yield considerably higher returns than
other ARMAX models. These two models are also depicted in Figure 4.5. When
the return differences between the best error correction probit model and the best
ARMAX models are tested, the return differences are also statistically significant
at a 5% level based on the all considered DM test statistics shown in Table 4.6.
Therefore, the error correction probit model seems to be a superior predictive
model also against the alternative ARMAX models.

13
In the estimation of ARMAX models, the UCSD_GARCH toolbox package for Matlab is
used.
14
For instance, ARMAX(2,0) is the same as the AR(2) model with explanatory variables.

108
 Table 4.5: Out-of-sample performance of ARMAX and volatility models.
model xt−1 CR RET SR
B&H 8.07 0.79
ARMAX(1,0) pyt+5 0.565 7.12 0.67
ARMAX(1,0) pyt+5 , US
SPt−1 0.542 6.07 0.45
ARMAX(1,0) pyt+5 , ∆Rt−1 0.556 4.98 0.19
ARMAX(1,0) pyt+5 , ∆it−1 0.569 6.16 0.45
ARMAX(2,0) pyt+5 0.576 5.83 0.37
ARMAX(2,0) pyt+5 , SPt−1
US
0.514 4.36 0.04
ARMAX(2,0) pyt+5 , ∆Rt−1 0.532 4.76 0.13
ARMAX(2,0) pyt+5 , ∆it−1 0.588 6.28 0.47
Volatility models
Non-parametric 0.481 5.57 0.65
Extended 0.514 4.93 0.16
Pp ′
Notes: The ARMAX(p, 0) model for rt is rt = a + i=1 bi rt−i + xt−1 d. “Non-parametric” and
“Extended” refer to the predictive models proposed by Christoffersen et al. (2007), which are
based on the volatility forecasts σ̂t|t−1 . See also notes to Table 4.3.

1.5

1
Excess stock return

0.5

−0.5

−1

−1.5
1987 1990 1992 1995 1997 2000 2002 2005 2007
Time

Figure 4.5: Out-of-sample predictions of the ARMAX(1,0) model with pyt+5 (solid
line) and ARMAX(2,0) model (dashed line) with pyt+5 and ∆it−1 are depicted.

109
Table 4.6: Diebold-Mariano tests between the best error-correction probit model
and the best ARMAX models.
model xt−1 DM DMra DM It =0 It =0
DMra
ARMAX(1,0) pyt+5 0.046 0.019 0.000 0.000
ARMAX(2,0) pyt+5 , ∆it−1 0.015 0.006 0.000 0.000
Notes: The p-values of the Diebold and Mariano (1995) tests between the investment returns
from the error correction probit model presented in Table 4.4 and the ARMAX models
mentioned in the first column are reported.

As presented in equation (4.1), if the volatility σt , conditional on the infor-

mation at time t − 1, is predictable, then the signs of stock returns should be
predictable as well, provided that µt 6= 0, although the conditional mean µt could
be unpredictable. Using the same terminology as Christoffersen et al. (2007) for
predictive models based on volatility forecasts, a “non-parametric” model indicates
a model where the one-step-ahead volatility forecast is also used to compute the
conditional mean forecast, which then together with the volatility forecast deter-
mine the probability forecast for positive excess return. In an “extended” model,
the skewness and kurtosis of excess returns are also taken into account in the
model. The percentage of correct forecasts CR shown in Table 4.5 shows that the
volatility models do not produce out-of-sample sign predictability, and when com-
pared to the best error correction probit model, the latter produces higher value of
CR and higher investment returns. The p-values of DM test statistics between the
models are 0.012 (non-parametric model) and 0.008 (extended model), indicating
that the return differences are statistically significant at a 5% level.

4.4.5 Sign Predictability of Small and Large Size Firms’ Re-

turns

Finally, we extend our analysis by considering the sign predictability of small and
large size firms. Perez-Quiros and Timmermann (2000), among others, find that
there is a close link between stock returns of different size of firms and the state of
the economy. They propose that recession, which indicates, for instance, worsening

110
credit market conditions, is expected to affect the expected returns of small firms
more strongly than large size firms’ returns. In their model, Perez-Quiros and
Timmermann (2000) employed a Markov switching model where the continuous
excess stock return is modeled as a function of lagged Treasury Bill rate, default
spread, changes in the money stock growth and dividend yield.
In our limited study, we employ the presented error correction probit model
(4.9) with the six-month recession forecast employed for both return indicator
series. The values of the return indicators (4.2) are constructed by using the excess
stock returns rtS and rtL (see details in Table 4.1) from the size-sorted CRSP decile
portfolios. It is worth noting that there is a significant correspondence with the
binary values of stock indicator series between S&P500 and these size-sorted stock
indices, as expected. Especially in the case of large size firms the correspondence is
about 96%, as expected. In the returns from small size firms there is more variation
between the values of return indicators. Overall, the mean and the volatility of
small size firms’ returns are higher than in the case of large size firms.
Table 4.7 presents the out-of-sample forecasting performance of considered error
correction probit models. The p-values of the P T market timing test (4.15) are
statistically significant, which shows that the signs are predictable out of sample
in both cases. The percentage of correct forecasts, CR, is even higher than in the
case of S&P500 returns. Further, the sign predictability does also convert to the
higher investment returns in our simple trading simulation compared with B&H
strategy. However, the differences are statistically significant only in terms of risk-
adjusted returns where the standard deviation of the portfolio returns are taken
into account.

Table 4.7: Out-of-sample sign predictions for small and large size firms’ returns.
It =0
Firm size model xt−1 psR2 CR RET SR PT DM DMra DM It =0 DMra
Small B&H 13.76 1.37
ecm (4.9) pyt+5 0.029 0.625 15.72 2.24 0.002 0.267 0.014 0.000 0.000
Large B&H 10.89 1.36
ecm (4.9) pyt+5 0.007 0.634 12.91 2.07 0.014 0.116 0.009 0.000 0.000
Notes: As in Tables 4.3 and 4.5, the average of annual risk-free interest rate return is 4.21% in both series. See
also notes to Table 4.3.

111
 It seems that there is not much difference between the results of large and
small size firms in terms of predictive accuracy. For small size firms the error
correction model produces a higher value of psR2 measure. On the other hand, the
percentage of correct forecasts are higher in the case of large size firms. However,
one significant difference between the models can be seen in Figure 4.6. There
the conditional probability forecast for small firms fluctuates much more than in
the case of large size firms. In that model the estimated value for coefficient α1
is between 0.30–0.50 for the out-of-sample period, whereas for large size firms it
is about the same magnitude as in the analysis of the S&P500 index (α1 ≈ 0.90).
Despite the fact that this fluctuation around the 50% threshold indicates more
transaction costs (about a 2 percent deficit in investment returns compared with
the returns without transaction costs), as discussed in Section 4.2.2, the investment
returns from our trading simulation are higher than in the B&H strategy.

0.75

0.7

0.65

0.6
Probability

0.55

0.5

0.45

0.4

0.35

1987 1990 1992 1995 1997 2000 2002 2005 2007

Time

Figure 4.6: Out-of-sample predictions of the error correction probit models (4.9)
with the recession forecast pyt+5 for small (solid line) and large (dashed line) size
firms.

112
4.5 Conclusions
We examine the predictability of the U.S. excess stock return signs by using dy-
namic binary probit models. The proposed forecasting method, where the six-
month recession forecast for the recession indicator is used as an explanatory vari-
able, seems to outperform other predictive models. Using the S&P500 stock index,
the direction of the excess stock return is predictable and it is possible to earn sta-
tistically significant higher investment returns compared with the buy-and-hold
trading strategy in sample. However, out-of-sample predictability turns out to be
weaker. This is in line with previous findings related to stock return forecasting.
In fact, in out-of-sample forecasting, the best dynamic probit model appears to be
the error correction model proposed in this chapter. Using this model, the number
of correct sign predictions and investment returns are higher than in other probit
models, ARMAX models, or predictive models based on volatility forecasts. In the
best error correction model the average investment returns are also higher than
in the buy-and-hold trading strategy. Compared with the evidence from S&P500
returns, it appears that the out-of-sample sign predictability is higher in small
and large size sorted firms’ returns using the error correction probit model with
recession forecasts.
The analysis can be extended in various ways. A system analysis in which the
recession and sign forecasts are determined simultaneously in the same model is
of particular interest. In this chapter, the six-month ahead recession forecast is
taken as given, but this selection need not be optimal in terms of predictive power
in sign predictions. It could also be interesting to form a somewhat more compli-
cated trading strategy rule that could take the sign predictability, and perhaps the
risk related to different models, even better into account in investment allocation
decisions.

113
References
Anatolyev S, Gospodinov N. 2010. Modeling financial return dynamics via decom-
position. Journal of Business and Economic Statistics 28: 232–245.

Bernard H, Gerlach S. 1998. Does the term structure predict recessions? The in-
ternational evidence. International Journal of Finance and Economics 3: 195–215.

Breen W, Glosten LR, Jagannathan R. 1989. Economic significance of predictable

variations in stock index returns. Journal of Finance 44: 1177–1189.

Campbell JY, Shiller RJ. 1988. Stock Prices, earnings, and expected dividends.
Journal of Finance 43: 661–676.

Campbell JY, Thompson SB. 2008. Predicting excess stock returns out of sample:
Can anything beat the historical average. Review of Financial Studies 21: 1509–
1532.

Chauvet M, Potter S. 2000. Coincident and leading indicators of the stock market.
Journal of Empirical Finance 7: 87–111.

Chen N. 1991. Financial investment opportunities and the macroeconomy. Jour-

nal of Finance 46: 529–554.

Cochrane JH. 1997. Where is the market going? Uncertain facts and novel theo-
ries. Economic Perspectives, Federal Reserve Bank of Chicago 21: 3–37.

Christoffersen PF, Diebold FX. 2006. Financial asset returns, direction-of-change

forecasting, and volatility dynamics. Management Science 52: 1273–1287.

Christoffersen PF, Diebold FX, Mariano RS, Tay AS, Tse YK. 2007. Direction-of-
change forecasts based on conditional variance, skewness and kurtosis dynamics:
International evidence. Journal of Financial Forecasting 1: 3–24.

de Jong RM, Woutersen TM. In press. Dynamic time series binary choice. Econo-
metric Theory, forthcoming.

114
Diebold FX, Mariano RS. 1995. Comparing predictive accuracy. Journal of Busi-
ness and Economic Statistics 13: 253–263.

Elliott G, Lieli RP. 2007. Predicting binary outcomes. Unpublished manuscript,

University of Texas.

Engle RF, Bollerslev T. 1986. Modeling the persistence of conditional variances.

Econometric Reviews 5: 1–50.

Estrella A. 1998. A new measure of fit for equations with dichotomous dependent
variables. Journal of Business and Economic Statistics 16: 198–205.

Estrella A. 2005. The yield curve as a leading indicator: Frequently asked ques-
tions. Federal Reserve Bank of New York. http://www.newyorkfed.org/research/
capital_markets/ycfaq.pdf. [5 January 2009].

Estrella A, Mishkin FS. 1998. Predicting U.S. recessions: Financial variables as

leading indicators. Review of Economics and Statistics 80: 45–61.

Fama EF. 1990. Stock returns, expected returns, and real activity. Journal of
Finance 45: 1089–1108.

Fama EF, French KR. 1989. Business conditions and expected returns on stocks
and bonds. Journal of Financial Economics 25: 23–49.

Goyal A, Welch I. 2008. A comprehensive look at the empirical performance of

equity premium prediction. Review of Financial Studies 21: 1455–1508.

Granger CWJ, Pesaran HM. 2000. Economic and statistical measures of forecast
accuracy. Journal of Forecasting 19: 537–560.

Han Y. 2007. Return predictability, economic profits, and model mis-specification:

How important are the better specified models? Tulane University.
http://ssrn.com/abstract=967564. [2 July 2009].

Hong Y, Chung J. 2003. Are the directions of stock price changes predictable?
Statistical theory and evidence. Manuscript, Cornell University.

115
Kauppi H, Saikkonen P. 2008. Predicting U.S. recessions with dynamic binary
response models. Review of Economics and Statistics 90: 777–791.

Leitch G, Tanner J. 1991. Economic forecast evaluation: Profits versus the con-
ventional error measures. American Economic Review 81: 580–590.

Leung MT, Daouk H, Chen AS. 2000. Forecasting stock indices: A comparison
of classification and level estimation models. International Journal of Forecasting
16: 173–190.

Maddala GS. 1983. Limited-Dependent and Qualitative Variables in Econometrics.

Cambridge University Press, New York.

Nyberg H. 2010. Dynamic probit models and financial variables in recession fore-
casting. Journal of Forecasting 29: 215–230.

Pagan A. 1984. Econometric issues in the analysis of regressions with generated

regressors. International Economic Review 25: 221–247.

Perez-Quiros G, Timmermann A. 2000. Firm size and cyclical variations in stock

returns. Journal of Finance 55: 1229–1262.

Pesaran HM, Timmermann A. 1992. A simple nonparametric test of predictive

performance. Journal of Business and Economic Statistics 10: 461–465.

Pesaran HM, Timmermann A. 1995. Predictability of stock returns: Robustness

and economic significance. Journal of Finance 50: 1201–1228.

Russell JR, Engle RF. 2005. A discrete-state continuous-time model of finan-

cial transactions prices and times: The autoregressive conditional multinomial-
autoregressive conditional duration model. Journal of Business and Economic
Statistics 23: 166–180.

Rydberg T, Shephard N. 2003. Dynamics of trade-by-trade price movements: De-

composition and models. Journal of Financial Econometrics 1: 2–25.

Schwarz G. 1978. Estimating the dimension of a model. Annals of Statistics 6:

461–464.

116
Sharpe WF. 1966. Mutual fund performance. Journal of Business 39: 119–138.

Sharpe WF. 1994. The Sharpe ratio. Journal of Portfolio Management 21: 49–58.

117
Chapter 5

A Bivariate Autoregressive Probit

Model: Predicting U.S. Business
Cycle and Growth Rate Cycle
Recessions

Abstract1

We propose a new bivariate autoregressive probit model for binary time series.
The model nests various special cases, such as two separate univariate models,
for which a LM test against the unrestricted bivariate model is developed. The
parameters of the model are estimated by the method of maximum likelihood and
forecasts can be computed by using explicit formulae. The model is applied to
predict the current state of the U.S. business cycle and growth rate cycle recession
periods. Evidence of predictability of both recession periods is obtained by using
financial variables as predictors. The bivariate model is found to outperform the
univariate models built separately for each cycle indicator.

1
An earlier version of this chapter has been published in HECER Discussion Papers, No.
272, 2009.

119
5.1 Introduction
In the previous literature on time series models for binary dependent variables,
the models have typically been univariate. Given the importance of vector au-
toregressive models for continuous dependent variables, it is of interest to study
multivariate binary time series models, where the probabilities of different binary
outcomes are modeled jointly.
In this chapter, we present a bivariate autoregressive probit model as an ex-
tension to the univariate autoregressive probit model of Kauppi and Saikkonen
(2008). The model can also be seen as an extension of the “static” bivariate probit
model of Ashford and Sowden (1970), where the dependence between two binary
time series is modeled by using a bivariate cumulative normal distribution func-
tion. In our bivariate model, the static model is extended by the inclusion of the
autoregressive model structure.
In the previous literature, only few bivariate and multivariate models have
been considered. Those models have mainly been based on the latent variable
formulation, where the values of binary time series are realizations of corresponding
continuous latent variables (see, e.g., Chib and Greenberg, 1998; Mosconi and Seri,
2006). In this chapter, the latent variable approach is not used. An advantage
of our model is that parameter estimation can conveniently be carried out by
the method of maximum likelihood and forecasts can be computed using explicit
formulae. This is not typically the case in dynamic models based on the latent
variables, such as the dynamic univariate model by Chauvet and Potter (2005) and
the Qual VAR model of Dueker (2005). Our bivariate model is closely related to
the model proposed by Anatolyev (2009), but we model the dependence between
the two binary time series in a different way.
As an empirical application, we consider several alternative specifications of
the proposed bivariate autoregressive probit model to nowcast the current state
of the U.S. economy. We measure the state of the economy in terms of reces-
sion periods defined by the business cycle and the growth rate cycle indicators.
Predicting business cycle recession periods with univariate probit models has at-

120
tracted considerable attention in the literature (see, e.g., Estrella and Hardouvelis,
1991; Estrella and Mishkin, 1998; Chauvet and Potter, 2005), where the growth
rate cycle indicator has hardly been considered at all. Given the fact that some
economic slowdown periods do not turn into business cycle recessions, it is also
of interest to consider the binary growth rate cycle indicator. To the best of our
knowledge, this type of bivariate framework of two cycle indicators has not been
considered in the previous literature.
A growth rate cycle is defined in terms of periods of increasing and decreasing
growth rate in economic activity (see details in Banerji and Hiris, 2001; Osborn,
Sensier and van Dijk, 2004). While “classical” business cycle recession periods are
associated with the level of economic activity, a growth rate recession may occur
without a decline in the level of economic activity. Therefore, growth rate cycle
recessions are more numerous than classical business cycle recessions, but from the
viewpoint of economic policy, they may be at least equally important and infor-
mative. For example, monetary policy decisions made by central banks are based
on the real time assessment of the current, and also expected future, economic
conditions using the data available at the time the decision is made. As Osborn
et al. (2004) point out, growth rate cycles are closely related to the estimated
output gap, which is supposedly an important variable affecting monetary policy
decisions.
We concentrate on the predictive power of financial variables for business cycle
and growth rate cycle recessions. The advantage of those variables is that they are
available on a continuous basis without revisions. A difficulty with macroeconomic
predictive variables, such as initial estimates of the real GDP or the estimated
output gap, in contrast, is that they face substantial revisions during subsequent
months and observations of some variables are not even available on a monthly ba-
sis. These properties of macroeconomic variables, which ultimately determine the
values of both cycle indicators, also mean that the real-time state of the economy
is always uncertain to some extent. Therefore, nowcasting the business cycle and
growth rate cycle indicators is of interest, and the real-time availability supports

121
financial variables as predictors.
Our results demonstrate the advantages of modeling the probabilities of busi-
ness cycle and growth rate cycle recessions jointly. As a matter of fact, among the
considered univariate and bivariate specifications, the proposed unrestricted bivari-
ate autoregressive probit model yields the best in-sample, but also out-of-sample
predictions. The lagged first difference of the Federal funds rate and monthly stock
market returns turn out to be the best predictive variables for the U.S. growth rate
cycle. As suggested in many previous studies, the U.S. term spread is an impor-
tant predictive variable for predicting business cycle recessions, but its predictive
power for growth rate cycle periods is limited.
The remainder of this chapter is organized as follows. The bivariate autore-
gressive probit model is introduced in Section 5.2. Issues of parameter estimation,
testing, and forecasting are discussed in Section 5.3. Section 5.4 presents the
empirical results. Section 5.5 concludes.

5.2 Bivariate Autoregressive Probit Model

Consider two binary time series, y1t and y2t , t = 1, 2, ..., T . Let us assume that
conditional on information set Ωt−1 the random vector (y1t , y2t ) follows a bivariate
Bernoulli distribution,

(y1t , y2t )|Ωt−1 ∼ B2 (P11,t , P10,t , P01,t , P00,t ), (5.1)

where
Pij,t = Pt−1 (y1t = i, y2t = j), i, j = 0, 1, (5.2)

and
P11,t + P10,t + P01,t + P00,t = 1. (5.3)

Hence, the conditional marginal probabilities of the separate outcomes y1t = 1 and
y2t = 1 are equal to
P1t = P11,t + P10,t , (5.4)

and
P2t = P11,t + P01,t , (5.5)

122
respectively.
A bivariate probit model was first proposed by Ashford and Sowden (1970) for
analyzing cross-sectional data. In their model the joint probabilities for different
outcomes of the vector (y1t , y2t ) are determined as

P11,t = Pt−1 (y1t = 1, y2t = 1) = Φ2 (π1t , π2t , ρ),

P10,t = Pt−1 (y1t = 1, y2t = 0) = Φ2 (π1t , −π2t , −ρ),

P01,t = Pt−1 (y1t = 0, y2t = 1) = Φ2 (−π1t , π2t , −ρ),

P00,t = Pt−1 (y1t = 0, y2t = 0) = Φ2 (−π1t , −π2t , ρ), (5.6)

where Φ2 (·) is the cumulative distribution function of the bivariate normal dis-
tribution with zero means, unit variances and correlation coefficient ρ, |ρ| < 1,
and π1t and π2t are assumed to be linear functions of variables x1,t−k and x2,t−k
included in the information set Ωt−1 , respectively. The sign changes in the ar-
guments of the bivariate cumulative normal distribution function are needed to
guarantee that condition (5.3) holds (see, for example, Greene, 2000, 849–850).
To complete the bivariate probit model, a parametrization for π1t and π2t needs
to be specified. Ashford and Sowden (1970) introduced the following parametriza-
tion,       
′
π1t ω1 x1,t−k 0 β1
 = +
′
 , (5.7)
π2t ω2 0 x2,t−k β2
where ω1 and ω2 are constant terms and β 1 and β 2 are coefficient vectors of the
′ ′
lagged explanatory variables included in x1,t−k and x2,t−k , respectively. Note that
using the same lag k in all explanatory variables is only for notational convenience
and can easily be relaxed in practice. Equations (5.6) and (5.7) together define
the static bivariate probit model.2
Dynamic extensions of the static model (5.7) can be obtained in various ways.
We propose the following “bivariate autoregressive probit model”,
         
′
π1t ω1 α11 α12 π1,t−1 x1,t−k 0 β1
 = +  +   , (5.8)
′
π2t ω2 α21 α22 π2,t−1 0 x2,t−k β2
2
The corresponding multivariate model is considered by Ashford and Sowden (1970), and
Chib and Greenberg (1998), among others.

123
where π1t and π2t are specified as linear functions of their lags and the lagged
′ ′
values of the explanatory variables included in x1,t−k and x2,t−k . Model (5.8) can
compactly be written as

′
π t = ω + Aπ t−1 + xt−k β, (5.9)
 ′  ′  ′
′ ′
where π t = π1t π2t , xt−k = diag x1,t−k x2,t−k , ω = ω1 ω2 , β =
 ′ ′
′
β 1 β 2 , and
 
α11 α12
A= .
α21 α22
In this bivariate autoregressive model we explicitly allow for the possibility that
′ ′
different explanatory variables can be included in x1,t−k and x2,t−k . If the param-
eter matrix A is unrestricted, both π1t and π2t can depend on the lagged values
of π1t and π2t . Thus, even if ρ = 0 in (5.6), the coefficients α12 and α21 provide
a linkage between the variables π1t and π2t in model (5.8). Note that the static
model is obtained from (5.9) with the restriction A = 0. Furthermore, it is only in
the special case where ρ = 0 and α12 = α21 = 0 that our bivariate autoregressive
probit model reduces to two independent univariate autoregressive probit models.
Model (5.8) is somewhat similar to the multivariate dynamic binary model of
Anatolyev (2009). The main difference is that Anatolyev (2009) suggests using the
so called “dependence ratios” (cf. Ekholm, Smith, and McDonald, 1995) between
the dependent variables to construct the conditional joint probabilities of the dif-
ferent outcomes of (y1t , y2t ). In parameter estimation the dependence ratios and
marginal probabilities for the variables y1t and y2t are handled separately by using
a logistic function. In our model, the dependence between y1t and y2t is instead
modeled by using the autoregressive specification (5.8) and the bivariate cumula-
tive normal distribution function, where the correlation coefficient ρ is allowed to
be nonzero. In addition, in the bivariate autoregressive probit model, parameter
estimation can be carried out within the same system without dependence ratios.
Note that if the roots of det(I2 − Az) lie outside the unit circle, we obtain by

124
recursive substitution of (5.9) the following representation,
∞
X ∞
X
j−1 ′
πt = A ω+ Aj−1 xt−k−j+1 β. (5.10)
j=1 j=1

This shows that in the bivariate autoregressive probit model (5.8) π1t and π2t
depend on the whole infinite history of the explanatory variables in a parsimonious
way and, therefore, the model can be interpreted as an “infinite order” extension
of the static model (5.7). Furthermore, assuming that the explanatory variables
included in xt−k are stationary, also π t is stationary.
It is worth noting that because of the characteristics of our empirical application
(see Section 5.4.2), the lagged values of y1t and y2t included in Anatolyev’s (2009)
model are excluded. However, that would be a possible extension of model (5.8).
This extension can be based on the univariate model of Kauppi and Saikkonen
(2008), where the lag yt−1 is also included in the right hand side of the model.

5.3 Parameter Estimation, Testing and Forecast-

ing

5.3.1 Maximum Likelihood Estimation

As in corresponding univariate models, parameter estimation in the bivariate au-

toregressive model defined by (5.6) and (5.8), as well as its special cases, can
conveniently be carried out by the method of maximum likelihood (ML). Using
the conditional probabilities in (5.6), one can write the likelihood function and
obtain the maximum likelihood estimate by using numerical methods.
Following Greene’s (2000, 849–850) notation, the log-likelihood function can
be constructed as follows. Define qjt = 2yjt − 1 and µjt = qjt πjt , j = 1, 2, so that

 1 if yjt = 1,
qjt =
 −1 if y = 0,
jt

and 
 π
jt if yjt = 1,
µjt =
 −π
jt if yjt = 0.

125
Furthermore, set
ρ∗t = q1t q2t ρ.

The conditional probabilities of the different outcomes of (y1t ,y2t ) can be expressed
as (cf. (5.6))
Pt−1 (y1t , y2t ) = Φ2 (µ1t , µ2t , ρ∗t ).
 ′
′
′ ′
Let θ = vec(A) ω β ρ denote the vector of the parameters of the
bivariate autoregressive probit model. The log-likelihood function, conditional on
initial values, is the sum of the individual log-likelihood functions lt (θ)3 ,
T
X T
X  
∗
l(θ) = lt (θ) = log Φ2 (µ1t , µ2t , ρt )
t=1 t=1
XT 
= y1t y2t log(P11,t ) + y1t (1 − y2t ) log(P10,t )
t=1

+ (1 − y1t )y2t log(P01,t ) + (1 − y1t )(1 − y2t ) log(P00,t ) . (5.11)

It is worth noting that if the correlation coefficient ρ in (5.6) is zero, the conditional
probabilities in (5.6) are products of the marginal probabilities (5.4) and (5.5). For
instance, in that case the conditional probability of the outcome (y1t = 1, y2t = 1)
is

P11,t = Pt−1 (y1t = 1, y2t = 1) = Pt−1 (y1t )Pt−1 (y2t ) = Φ(π1t )Φ(π2t ). (5.12)

The score vector of the log-likelihood function (5.11) is

T T
∂l(θ) X X ∂lt (θ)
s(θ) = = st (θ) = , (5.13)
∂θ t=1 t=1
∂θ

where

∂lt (θ) 1 ∂Φ2 (µ1t , µ2t , ρ∗t )

= .
∂θ Φ2 (µ1t , µ2t , ρ∗t ) ∂θ
′
3
In the bivariate autoregressive model (5.8) the selection of the initial value π0 = (π01 π02 )
is also needed to construct the log-likelihood function (5.11). Following Kauppi and Saikkonen
(2008) one way to obtain the initial values π0i , i = 1, 2, is to select π0i = (ωi + x̄i,t−k β i )/(1 − αii ),
where a bar indicates the sample mean of variables included in xi,t−k .

126
At this point it is convenient to split the parameter vector into three disjoint
′ ′ ′
components, namely θ = (θ 1 θ2 ρ) , where the parameters in θ 1 and θ 2 are
related to the specifications of π1t and π2t , respectively. The score vector can be
partitioned accordingly as
 ′ ′
′
st (θ) = s1t (θ 1 ) s2t (θ 2 ) s3t (ρ) . (5.14)

The first component of st (θ) can be written as

∂lt (θ) 1 ∂Φ2 (µ1t , µ2t , ρ∗t )
s1t (θ 1 ) = =
∂θ 1 Φ2 (µ1t , µ2t , ρ∗t ) ∂θ 1
1 ∂Φ2 (µ1t , µ2t , ρ∗t ) ∂µ1t
=
Φ2 (µ1t , µ2t , ρ∗t ) ∂µ1t ∂θ 1
1  µ − µ ρ∗  ∂π
2t 1t t 1t
= φ(µ1t )Φ p q1t ,
Φ2 (µ1t , µ2t , ρt )
∗
1 − ρt ∗2 ∂θ 1

where φ(·) and Φ(·) are the density function and the cumulative distribution func-
tion of the standard normal distribution, respectively. The value of s1t (θ 1 ) depends
on the realized values of y1t and y2t . For instance, if (y1t = 1, y2t = 1), then by
the definitions of µjt and q1t ,
1  π − π ρ  ∂π
2t 1t 1t
s1t (θ 1 ) = φ(π1t )Φ p .
Φ2 (π1t , π2t , ρ) 1−ρ 2 ∂θ 1
It can be seen that the main difference between the score vector of the static model
(A = 0) and model (5.8) is in the derivative term ∂π1t /∂θ 1 . In model (5.8),
   
∂π1t ∂π1,t−1 ∂π1,t−2
1 + α11 ∂ω1 + α12 α21 ∂ω1
 ∂ω1   
 ∂π1t   ∂π1,t−1 ∂π1,t−2 
∂π1t  ∂α11   1,t−1
   π + α11 ∂α11 + α α
12 21 ∂α11 
= = ,
∂θ 1 ∂π ∂π1,t−1
 ∂α   π2,t−1 + α11 ∂α + α12 α21 ∂α
1t
  ∂π1,t−2 

 12   12 12 
∂π1t ∂π1,t−1 ∂π1,t−2
∂β
x 1,t−k + α 11 ∂β + α α
12 21 ∂β
1 1 1

 ′
whereas, in the static model, it reduces to 1 x1,t−k . The derivative ∂lt (θ)/∂θ 2
is obtained in the same way by replacing ∂π1t /∂θ 1 in the definition of st (θ) by
∂π2t  ∂π2t ∂π2t ∂π2t ∂π2t ′
= .
∂θ 2 ∂ω2 ∂α22 ∂α21 ∂β 2
Given the result
∂Φ2 (µ1t , µ2t , ρ∗t )
= φ2 (µ1t , µ2t , ρ∗t ),
∂ρ∗t

127
the derivative with respect ρ is
∂Φ2 (µ1t , µ2t , ρ∗t ) ∂Φ2 (µ1t , µ2t , ρ∗t ) ∂ρ∗t
= = φ2 (µ1t , µ2t , ρ∗t )q1t q2t .
∂ρ ∂ρ∗t ∂ρ
Therefore, the score of the correlation coefficient ρ becomes
∂lt (θ) 1 ∂Φ2 (µ1t , µ2t , ρ∗t ) ∂ρ∗t φ2 (µ1t , µ2t , ρ∗t )
s3t (ρ) = = = q1t q2t .
∂ρ Φ2 (µ1t , µ2t , ρ∗t ) ∂ρ∗t ∂ρ Φ2 (µ1t , µ2t , ρ∗t )
The value of s3t (ρ) depends on realized values of the dependent variables. For
example, if y1t = 1 and y2t = 1,
φ2 (π1t , π2t , ρ)
s3t (ρ) = ,
Φ2 (π1t , π2t , ρ)
and if y1t = 1 and y2t = 0,
φ2 (π1t , −π2t , −ρ)
s3t (ρ) = − .
Φ2 (π1t , −π2t , −ρ)
Maximization of the log-likelihood function (5.11) yields the maximum like-
lihood estimate θ̂, which solves the first order condition s(θ̂) = 0. Under ap-
propriate regularity conditions, including the stationarity of π t and explanatory
variables and the correctness of the probit model specification, the conventional
large sample theory of ML estimation gives the usual asymptotic distribution
L
T 1/2 (θ̂ − θ) −→ N(0, I(θ)−1 ), (5.15)
′
where I(θ) = plim T −1 ∂ 2 l(θ)/∂θ∂θ exists and is positive definite.
A practical difficulty with the bivariate autoregressive probit model (5.8) is
that the number of parameters can become large if many explanatory variables
are included. ML estimation is considerably simplified if the correlation coefficient
ρ is restricted to zero because then the bivariate probabilities in the log-likelihood
function (5.11) factor into products of marginal probabilities, as in (5.12). Thus,
it is of interest to test for the hypothesis ρ = 0. In the next section, a LM test for
this purpose is developed.

5.3.2 LM Test for the Correlation Coefficient

For testing the significance of the correlation coefficient, the Lagrange Multiplier
test is attractive because it only requires ML estimation under the null hypoth-
esis ρ = 0. Kiefer (1982) has proposed a corresponding LM test for the static

128
bivariate probit model (5.7). In this section, the test is extended to the bivariate
autoregressive model (5.8).
′ ′
′
Let θ̃ = (θ̃ 1 θ̃ 2 0) be the restricted ML estimate of θ obtained by assuming

H0 : ρ = 0. (5.16)

The general form of the LM test statistic (see, for example, Engle, 1984) is

′
LM = s(θ̃) Ĩ(θ̃)−1 s(θ̃), (5.17)

where Ĩ(θ̃) is a consistent estimate of the information matrix I(θ) and s(θ̃) is
the score vector (5.13) evaluated at the restricted ML estimates θ̃. Under the null
hypothesis (5.16) the test statistic has an asymptotic χ21 distribution.
Due to the complexity of the second derivatives of the log-likelihood function
(5.11), the outer-product of the score is an attractive estimator of the information
matrix I(θ). The resulting test statistic is

′
 ′
−1 ′
LM ρ = ι S(θ̃) S(θ̃) S(θ̃) S(θ̃) ι, (5.18)

where ι is a vector of ones and the matrix S(θ̃) is given by

 ′
S(θ̃) = s1 (θ̃) s2 (θ̃) ... sT (θ̃) .

As in (5.14), the score vector st (θ̃), evaluated at θ̃, consists of three components.
The bivariate densities and probabilities factor into products of marginals and,
consequently, the components of the score reduce to

φ(µ̃1t ) ∂π̃1t
s1t (θ̃ 1 ) = q1t ,
Φ(µ̃1t ) ∂θ 1
φ(µ̃2t ) ∂π̃2t
s2t (θ̃ 2 ) = q2t ,
Φ(µ̃2t ) ∂θ 2

and
φ( µ̃1t )φ( µ̃2t )
s3t (0) = q1t q2t ,
Φ(µ̃1t )Φ(µ̃2t )
where “∼” on the right hand side means that the corresponding quantities are eval-
uated at θ = θ̃. The derivatives ∂π1t /∂θ 1 and ∂π2t /∂θ 2 depend on the considered
specifications of π1t and π2t (see Section 5.3.1).

129
5.3.3 Forecasting

As shown by Kauppi and Saikkonen (2008), explicit formulae can be used to obtain
one-period and multiperiod forecasts in the case of the univariate model. The
obtained forecasts are probability forecasts for different outcomes of (y1t , y2t ). In
the following we show that the same principles can also be applied in the proposed
bivariate model.
In the mean-square sense, the optimal h-period forecast based on the given
information available at time t − h, h ≥ 1, is the conditional expectation

Et−h (y1t , y2t ) = Φ2 (µ1t , µ2t , ρ∗t ). (5.19)

For example, the forecast for outcome (y1t = 1, y2t = 1) is given by

(h) (h)
Et−h (y1t = 1, y2t = 1) = Φ2 (π1t , π2t , ρ),

where, as shown in (5.10), by recursive substitution the bivariate system (5.9) can
be written as
h  
(h)
X ′
h
πt = A π t−h + Aj−1 ω + xt−k−j+1 β , (5.20)
j=1

 ′
(h) (h) (h)
where π t = π1t π2t and the vector π t−h is a function of values of the
explanatory variables and the initial values of π1,0 and π2,0 . In addition, the
condition k ≥ h for all predictors included in xt−k must hold indicating that
the employed lags of the predictive variables are always tailored to match the
information available at the time of forecasting. The usual case is obtained by
selecting k = h. The right hand side of (5.20) gives the h step forecast for the
outcome (y1t = 1, y2t = 1) “directly” using the information up to the forecast time
t − h. Forecasts for other outcomes of vector (y1t , y2t ) are obtained by imposing
necessary sign changes in bivariate normal cumulative distribution function (see
(5.6)).
Both in-sample and out-of-sample predictive performance of the employed mod-
els can be evaluated with goodness-of-fit measures commonly used for binary de-
pendent variables. Comparisons between different models can be based on the

130
value of the maximized log-likelihood function (5.11), denoted by logL below. It
can also be used to compute values of model selection criteria, such as the Schwarz
information criterion (Schwarz, 1978) defined as

log(T )
BIC = −logL + K , (5.21)
2

where K is the number of parameters in θ and T is the number of observations. An-

other goodness-of-fit measure is the quadratic probability score, QP S, suggested
by Diebold and Rudebusch (1989). Using the marginal conditional probability
forecasts P1t and P2t (see (5.4) and (5.5)), the quadratic probability score for vari-
able yjt is
T
1X  2
QP Sj = 2 yjt − Pjt ) , (5.22)
T t=1
where j = 1, 2. The values of the QP Sj lie on the interval [0,2] with the value 0
indicating a perfect fit. It can be seen as a counterpart of the mean square error
used with models for continuous variables.
Because of the binary nature of the dependent variable, the percentage of
correct predictions (CR) is a natural measure of predictive performance. However,
a threshold value must be specified that translates the probability forecasts into
signal forecasts (yjt = 1 or yjt = 0, j = 1, 2). The most commonly used and natural
threshold value is 0.50, which is also used in this study. When the signal forecasts
are constructed, a test proposed by Pesaran and Timmermann (1992) is available
for the evaluation of the directional predictive performance of a model. The null
hypothesis of the test is that the value of the correct prediction ratio does not differ
significantly from the ratio that would be obtained in the case of no predictability,
where the forecasts and realized values of yjt , j = 1, 2, are independent. Under the
null hypothesis of no predictability, the test statistic has an asymptotic standard
normal distribution.

131
5.4 Empirical Application: Predicting the Current
State of the U.S. Economy
We apply different bivariate probit model specifications to predict the state of
the U.S. economy. In this application, we predict, or more specifically “nowcast”,
values of the dependent U.S. business cycle and growth cycle indicators to be
discussed in more detail in Section 5.4.1 using the real-time information on financial
predictive variables. The monthly sample size covers the period from January 1971
to December 2005.
Knowledge of the current state of aggregate economic activity is important for
many economic agents in business and finance, as well as for policymakers, such as
central banks and government organizations. However, because of informational
lags and revisions of important macroeconomic variables, such as the real GDP,
the current state of the economy is always uncertain to some extent. In our
nowcasting exercise, the forecast horizon will therefore be one month, h = 1. Thus,
we are interested in predicting the probabilities of business cycle and growth cycle
recessions for month t using the information up to the end of the previous month
t − 1. In other words, the nowcasts are constructed at the beginning of month t.4

5.4.1 Binary Indicators for the Business and Growth Rate

Cycles

Forecasting the recession periods of the economy with various univariate binary
time series models has attracted considerable attention in many previous studies
(see, among others, Estrella and Mishkin, 1998; Chauvet and Potter, 2005; Kauppi
and Saikkonen, 2008; Nyberg, 2010). These recession periods are related to busi-
ness cycle fluctuations defined in terms of the level of economic activity. Thus, our
4
Matlab 7.8.0. and its BFGS optimization routine in the Optimization Toolbox is employed
in estimation.

132
first binary recession indicator is

 1, if the economy is in a recession at time t,
y1t = (5.23)
 0, if the economy is in an expansion at time t.

The best-known indicator for the U.S. is that one provided by the National Bureau
of Economic Research (NBER). It is based on the definition in which a recession
is “a significant decline in economic activity spread across the economy, lasting
more than a few months, normally visible in real GDP, real income, employment,
industrial production, and wholesale-retail sales.” 5 It is important to note that the
NBER uses a broader array of economic indicators than just, say, the real GDP,
to determine the recession periods.
In this chapter, we are mainly interested in predicting the growth rate cycles
of the U.S. economy. To the best of our knowledge, only Osborn et al. (2004)
have so far studied these cycle periods by means of binary time series models. In
contrast to classical business cycles characterized by the recession indicator (5.23),
growth rate cycles are related to the growth rate of aggregate economic activity.
We adopt the growth rate cycle periods defined by the Economic Cycle Research
Institute (ECRI). Based on their definition of “periods of cyclical upswings and
downswings in growth”, we introduce the binary indicator

 1, if the growth rate cycle is in a downswing state at time t,
y2t = (5.24)
 0, if the growth rate cycle is in an upswing state at time t.

In this chapter, the “downswing state” of the growth rate indicator (y2t = 1) is
referred to as a “growth rate recession”.6
ECRI determines the turning points in the growth rate cycle in a way analogous
to the “NBER approach”, where the co-movements and cyclical turns in various
measures of the aggregate macroeconomic activity are taken into account. Banerji
and Hiris (2001) provide a more detailed discussion on the business cycle and
growth rate cycle periods (see also Layton and Moore, 1989). It is expected that
the growth rate cycle indicator y2t exhibits more regime switches than the business
5
See details on http://www.nber.org/cycles/recessions.html [2 July 2009].
6
Osborn et al. (2004) refer these periods as “growth regimes”.

133
cycle indicator y1t . The reason is that a period with a lower growth rate may be
classified as a growth rate recession, but it is not necessarily defined as a business
cycle recession.

70 72 75 77 80 82 85 87 90 92 95 97 00 02 05 07
Time

Figure 5.1: U.S. business cycle recession periods (y1t = 1, line) since January 1972
until December 2005. Shaded areas indicate growth rate recessions (downswing
periods in the growth rate cycle, i.e. y2t = 1).

Figure 5.1 depicts the ECRI growth rate recession periods along with the NBER
recession periods. Table 5.1 shows the cross-tabulation of the realized values of
y1t and y2t defined in terms of these periods. As expected, growth rate cycles are
more numerous than “classical” business cycles. All slowdowns in the growth of
economic activity do not involve business cycle recessions. On the other hand,
the growth rate cycle recessions seem to lead the business cycle recessions: the
growth rate recession has typically started a few months before a business cycle
recession period. Furthermore, it should be pointed out that the rare outcome
(y1t = 1, y2t = 0), i.e. the economy is in a business cycle recession, but at the
same time in a growth rate expansion, is also possible. Table 5.1 shows that this
outcome has been occurred in five months in our data set. When taking a closer
look at the turning point chronologies of the NBER and ECRI, it can be seen that

134
those periods have been related, as expected, to the endpoints of the business cycle
recessions (y1t = 1), where the growth rate expansion (y2t = 0) has started before
the business cycle expansion (y1t = 0).

Table 5.1: Dependent variables and the cross-tabulation of realized values.

y2t
0 1
y1t 0 162 205
1 5 48
Notes: U.S. business cycle periods y1t (recession/expansion) and growth rate cycle periods y2t
(growth rate recession/growth rate expansion) are obtained from
http://www.nber.org/cycles/cyclesmain and http://www.businesscycle.com. The sample
period is 1972 M1–2005 M12. [30 April 2009]

Osborn et al. (2004) emphasize the relationship between the output gap (i.e.
the difference between the realized level and potential of output) and growth rate
cycle recessions and expansions periods (see Figure 5.1 and Section 1.3.2). The
output gap is of interest for many economic agents, especially central banks in
their setting of monetary policy. However, information on the value of the current
output gap is not available in real time (see, e.g., Orphanides and van Norden,
2002) because of the informational delays and revisions in real-GDP which is often
employed to determine the output gap. Thus, because of this above-mentioned
relationship, accurate predictions for the growth rate cycles can be very useful, for
example, in monetary policy decision making.

5.4.2 Data Set and Predictive Models

In addition to y1t and y2t , our data set consists of a number of financial variables,
such as interest rates and stock market returns, which are used as predictors in
bivariate probit models.7 Both levels and first differences of various interest rates
are considered. Assuming that monetary policy has an impact on real economic
activity and its growth rate, it is of interest to study which interest rate variable
7
The variables are described in more detail in Table 5.2.

135
is the most informative predictor. The Federal funds rate (F Ft ) is closely related
to the monetary policy in the U.S., so that it is a natural candidate variable (see,
e.g., Bernanke and Blinder, 1992).

Table 5.2: Explanatory variables.

rt Stock market return, log-difference of the S&P500 index
F Ft Federal funds rate
it Three-month Treasury bill rate, secondary market
Rt 10-year Treasury bond yield rate, constant maturity
SPt Term spread, Rt − it
∆F Ft First difference in Federal funds rate
∆it First difference in three-month Treasury bill rate
∆Rt First difference in 10-year Treasury bond yield
SPtGE German term spread between the long-term and short-term interest rate.
Notes: Interest rates are from Federal Reserve Statistical Release Historical Data set
(http://www.federalreserve.gov/releases/h15/data.htm). S&P500 stock market index is taken
from Yahoo Finance (http://finance.yahoo.com) and from http://www.econstats.com. German
term spread is constructed as the difference between 10-year Federal security (series WZ9826,
the missing values between 1971 M1-1972 M9 are replaced by the OECD 10-year interest rate)
and the three-month money market rate (series su0107, see
http://www.bundesbank.de/statistik/statistik) [30 April 2009].

The term spread (SPt ) between the long-term interest rate and the short-
term interest rate has often been found the most important predictor of business
cycle recessions (see, e.g., Estrella and Mishkin, 1998). Hence, it is of interest to
examine whether the term spread is also an important predictor for the growth
rate cycle periods. Furthermore, as a forward-looking variable and incorporating
expectations of future dividends and profitability of firms, stock market returns
(rt ) should also have predictive power.
We concentrate on the potential predictive information of financial variables,
which are available with no revisions or informational lags at the monthly fre-
quency. For example, the output gap, which Osborn et al. (2004) find the best
predictor of growth rate cycle recessions with the long-term and short-term inter-
est rates and stock market returns for European countries, is not available on a

136
real-time basis. In addition to the real-time availability of predictive variables, an-
other issue that should be taken into account in the specification of the predictive
model is the fact that the values of the NBER business cycle phases (y1t ), and
apparently also the growth cycle periods defined by the ECRI (y2t ), become avail-
able with very long delays. We call these delays as “publication lags”. Without
explicit assumptions concerning the publication lags it is difficult to use lagged
values of the cycle indicators in the predictive model. Overall, the publication lags
are typically so long that it is likely that the lagged values of the indicators are
statistically insignificant in estimated models (see the evidence on univariate mod-
els in Nyberg, 2010).8 Therefore, we only consider models excluding the lagged
values of the indicators y1t and y2t .
We consider four different model specifications obtained from the bivariate
autoregressive probit model given in (5.6) and (5.8). The models are defined as
follows:

Model 1 : α12 = 0, α21 = 0, ρ = 0,

Model 2 : α12 = 0, α21 = 0,

Model 3 : ρ = 0,

Model 4 : unrestricted model.

Model 1 consist of two independent autoregressive probit models. Model 2 is ob-

tained from Model 1 by removing the restriction that the correlation coefficient ρ is
zero. Note that Models 1 and 2 are already extensions of the static bivariate model
(5.7) because both π1t and π2t follow univariate autoregressive models. Model 4 is
the bivariate autoregressive probit model (5.8) without any restrictions, whereas
only the correlation coefficient ρ in (5.6) is restricted to zero in Model 3.

5.4.3 Model Selection and In-Sample Results

Model selection considered in this section is based on models estimated over the
entire sample period from January 1971 to December 2005. The first 12 observa-
8
The most recent publication lags of the NBER have varied from five up to twenty months
(see http://www.nber.org/cycles/cyclesmain.html) [2 July 2009].

137
tions are used as initial values in estimation. This section is also a starting point
for out-of-sample forecasts for the U.S. at time period 2006–2008 considered in
Section 5.4.5.
As mentioned in Section 5.4.1, unlike the business cycle recession periods, there
are few previous results on the predictability of growth rate cycle periods. In model
selection, therefore, we concentrate on predicting the growth rate cycle recessions
with various financial variables. For simplicity, we employ the same explanatory
variables as Nyberg (2010) in the predictive models of the U.S. business cycle
recession periods, i.e.,
 ′
GE
x1,t−k = SPt−6 rt−1 SPt−6 , (5.25)

where SPt is the U.S. term spread, rt is the monthly stock market return, and
SPtGE is the German term spread (see also the evidence in Estrella and Mishkin,
1998; Bernard and Gerlach, 1998). The German term spread is used as a “repre-
sentative” foreign term spread reflecting the state of the economy in the euro area.
Here the employed lags are the same as in Nyberg (2010) and, because the forecast
horizon is one month (h = 1), the condition k ≥ h (see Section 5.3.3) is satisfied.
We apply the following model selection procedure concerning the variables
included in x2,t−k . First, we estimate univariate autoregressive probit models
with one predictor. The considered predictive variables are listed in Table 5.2.
When the best single predictor based on BIC is found, it will be retained in the
model and different models with two predictors are estimated. We mainly restrict
ourselves to models with two predictors, but make some experiments with models
containing the third predictor as well. Finally, we consider bivariate models with
the predictors selected at the first stage.
Table 5.3 shows values of the Schwarz information criterion (5.21) for Model 1
with different explanatory variables, when the employed lag varies from one (k = 1)
to six (k = 6).9 Especially in more general bivariate probit models, such as Model
4, the number of parameters is quite large, indicating that ML estimation may
9
It appears that the evidence of predictive power of different explanatory variables is the
same when goodness-of-fit measures other than the BIC are used.

138
become difficult. Therefore, we prefer parsimonious models which makes BIC a
suitable model selection criterion.

Table 5.3: BIC values for Model 1 with different explanatory variables for the
growth rate cycle indicator.
k rt−k F Ft−k it−k Rt−k SPt−k ∆F Ft−k ∆it−k ∆Rt−k
1 308.77 312.76 315.26 331.25 300.52 335.83 335.83 335.41
2 314.75 312.20 314.66 330.95 300.51 280.41 286.39 335.48
3 319.02 314.11 314.92 330.27 304.30 280.34 282.19 336.05
4 321.64 317.55 318.36 330.21 313.58 289.94 286.34 335.90
5 327.01 321.22 320.82 330.55 318.26 298.25 286.87 319.66
6 331.01 323.71 323.53 330.80 323.52 304.70 308.30 325.52
Notes: Explanatory variable included in x2,t−k and its lag k are mentioned in the first row and
the first column of the table. Schwarz information criterion, BIC, is defined in (5.21).

Table 5.4: BIC values for Model 1 with the lagged first difference of the Federal
funds rate and other explanatory variables for the growth rate cycle indicator.
k rt−k F Ft−k it−k Rt−k SPt−k ∆it−k ∆Rt−k
∆F Ft−2 1 269.75 277.56 277.49 276.91 283.39 281.06 278.26
2 273.18 277.49 277.31 276.54 303.51 283.08 280.49
3 278.27 277.49 277.25 276.26 283.28 282.00 282.94
4 277.05 277.64 277.35 276.11 283.20 280.28 283.30
5 278.38 277.88 277.57 276.12 283.11 276.31 282.87
6 282.23 278.19 277.97 276.23 282.94 281.82 282.66
∆F Ft−3 1 270.27 278.18 277.90 277.49 283.32 282.57 280.85
2 274.09 278.43 278.08 277.22 303.44 281.84 281.95
3 278.64 278.62 278.73 276.84 307.16 282.85 282.64
4 278.61 278.62 278.30 276.83 282.92 283.23 283.24
5 279.22 278.63 278.30 276.55 282.87 280.26 283.34
6 282.07 278.78 278.51 276.53 282.72 282.28 282.93
Notes: The employed predictive variables in x2,t−k are the second, or the third, lag of the first
difference of the Federal funds rate and variable mentioned in the first row of the table. See
also notes to Table 5.3.

According to Table 5.3, the best single predictive variable seems to be the first
difference of the Federal funds rate lagged by two or three months (∆F Ft−2 or

139
∆F Ft−3 ). The first difference of the short-term interest rate (∆it−k ) also performs
quite well. Furthermore, as seen from Table 5.4, when models with two predict-
ive variables are considered, the lagged stock market return rt−1 combined with
∆F Ft−2 provides the lowest value of the BIC. Thus the vector x2,t−k is selected
as
 ′
x2,t−k = ∆F Ft−2 rt−1 . (5.26)

This selection is also meaningful from the viewpoint of the predictive ability of
financial markets because variables reflecting the effect of both the monetary policy
(∆F Ft−2 ) and the stock market returns (rt−1 ) are now included in the model.
Based on the evidence in Tables 5.3 and 5.4, the U.S. term spread has some
ability to predict the growth rate cycle periods. However, it is outperformed
by several other variables. The term spread is also found to be a statistically
insignificant predictor when used a third predictor in Models 1–4 together with
the explanatory variables given in (5.26).
Table 5.5 shows the estimation results for Models 1–4 with the explanatory
variables given in (5.25) and (5.26). The signs of the estimated coefficients are
as expected. In the case of the growth rate cycle periods, increasing values of
the differenced Federal funds rate and negative stock market returns increase the
probability of growth rate recession. The results thus indicate that the U.S. mone-
tary policy has a statistically significant predictive impact for the current state
of the growth rate cycle via the first difference of the Federal funds rate. Stock
market returns also have predictive power for the growth rate cycles as well as for
business cycle recession periods. In addition, the U.S. term spread (SPt−6 ) and
GE
the German term spread (SPt−6 ) are also statistically significant predictors.
In Model 2, the estimate of the correlation coefficient ρ is statistically signifi-
cant. The LM ρ test based on the restricted Model 1 yields the same conclusion.
A positive estimate of ρ is obtained, as expected, since the correlation between
the dependent variables is positive. Further, in Model 3, the estimates of the
off-diagonal elements of the matrix A, α12 and α21 , for π2,t−1 and π1,t−1 are statis-
tically significant, and according to the values of the BIC, Model 3 outperforms

140
 Table 5.5: Estimation results of different bivariate models.
model variable Model 1 Model 2 Model 3 Model 4
π1t constant1 0.06 0.06 -0.07 -0.16
(0.03) (0.03) (0.05) (0.05)
π1,t−1 0.85 0.86 0.88 0.86
(0.02) (0.02) (0.01) (0.01)
π2,t−1 0.17 0.18
(0.05) (0.05)
SPt−6 -0.19 -0.16 -0.08 -0.06
(0.03) (0.04) (0.03) (0.02)
rt−1 -0.11 -0.10 -0.10 -0.10
(0.02) (0.02) (0.02) (0.02)
GE
SPt−6 -0.08 -0.07 -0.13 -0.11
(0.03) (0.02) (0.03) (0.03)
π2t constant2 0.07 0.07 -0.02 -0.05
(0.01) (0.01) (0.01) (0.01)
π2,t−1 0.91 0.91 0.92 0.96
(0.02) (0.01) (0.01) (0.01)
π1,t−1 -0.03 -0.04
(0.01) (0.01)
∆F Ft−2 0.51 0.51 0.34 0.31
(0.18) (0.06) (0.06) (0.05)
rt−1 -0.03 -0.03 -0.06 -0.06
(0.01) (0.01) (0.01) (0.01)
ρ 0.58 0.53
(0.21) (0.21)
logL -242.70 -240.31 -223.38 -217.94
BIC 269.75 270.36 256.44 254.01
QP S1 0.061 0.063 0.061 0.062
QP S2 0.340 0.330 0.302 0.284
CR50%
1 0.963 0.961 0.961 0.963
CR50%
2 0.713 0.723 0.770 0.789
LM ρ 8.58 27.21
p-value 0.000 0.000
Notes: The models are estimated using the data from 1971 M1 to 2005 M12. The first 12 observations are used
as initial values. Standard errors of the estimated coefficients are given in parentheses. Estimated values of the
log-likelihood function (5.11), logL, and Schwarz (1978) information criteria, BIC, are reported, as well as the
values of quadratic probability scores, QP Sj , where j = 1, 2. Further, CR50%
j indicate the ratio of correct
prediction with using the 50% threshold value in the classification of probability forecasts. Lagrange Multiplier
test statistics LM ρ (see (5.18)) for the null hypothesis (5.16) and the corresponding p-values are also reported.

Model 2 as an extension of Model 1. However, both the LM ρ test based on Model

3 and the Wald test based on Model 4 point to a nonzero value of the correlation
coefficient ρ. Thus, Model 4 clearly yields the best in-sample fit. This is especially

141
the case when the models for the growth rate cycle periods are compared with the
values of QP S2 and CR2 .
The fact that Model 4 outperforms alternative bivariate probit models reflects
the fact that recession probabilities of the two cycle indicators are dependent on
both π1,t−1 and π2,t−1 . For instance, a positive estimate of α12 indicates that
the probability of a business cycle recession is high when the lagged probability
of growth rate recession is high (high value of π2t ). This is in line with the fact
that the growth rate recession appears to precede occurred business cycle recession
periods. Note that in Models 3 and 4, the U.S. term spread, which is a statistically
significant predictor for business cycle recession periods, has also an effect on the
growth rate cycle recession probability via the coefficient α21 for π1,t−1 in π2t .
As an extension of the results presented in Table 5.5, we consider the possibility
that there has been a structural break in the data generating process. In the
previous literature, it has been suggested that there might have been a structural
change in the U.S. economic activity in the mid-1980s. For example, McConnell
and Perez-Quiros (2000) and Blanchard and Simon (2001) have documented that
the variability of output, and also the variability of inflation, has declined after the
mid-1980s.10 Sensier and van Dijk (2004) also provide evidence that there have
been structural breaks in the unconditional volatility of many U.S. macroeconomic
time series around the years 1984 to 1986.
In our application, a parsimonious way to allow for the potential effect of a
structural break is the inclusion of an additional dummy variable in the model.
The dummy variable, denoted by 171−84 , takes the value one before the beginning
of the year 1985. The results for the augmented models are presented in Table
5.6, where the dummy variable turns out to be a statistically significant predictor.
The models are also superior to their counterparts in Table 5.5 according to BIC
values. Model 4 augmented with the dummy variable 171−84 yields clearly the best
in-sample predictions, especially for growth rate cycle recessions.

10
This time period after the mid-1980s is often referred to the “Great Moderation” period.

142
Table 5.6: Estimation results of different models with the additional dummy vari-
able 171−84 .
model variable Model 1 Model 2 Model 3 Model 4
π1t constant1 0.09 0.07 0.08 0.03
(0.05) (0.05) (0.04) (0.04)
171−84 -0.04 -0.00 -0.25 -0.21
(0.06) (0.05) (0.09) (0.08)
π1,t−1 0.85 0.86 0.89 0.88
(0.02) (0.02) (0.01) (0.01)
π2,t−1 0.11 0.11
(0.04) (0.03)
SPt−6 -0.20 -0.16 -0.11 -0.10
(0.04) (0.04) (0.03) (0.02)
rt−1 -0.11 -0.09 -0.09 -0.09
(0.04) (0.02) (0.02) (0.02)
GE
SPt−6 -0.08 -0.08 -0.12 -0.11
(0.03) (0.03) (0.03) (0.02)
π2t constant2 0.06 0.06 -0.07 -0.09
(0.01) (0.01) (0.02) (0.02)
171−84 0.06 0.06 0.09 0.08
(0.02) (0.02) (0.03) (0.02)
π2,t−1 0.91 0.91 0.93 0.94
(0.01) (0.01) (0.01) (0.01)
π1,t−1 -0.04 -0.04
(0.01) (0.01)
∆F Ft−2 0.55 0.55 0.40 0.38
(0.07) (0.06) (0.08) (0.07)
rt−1 -0.03 -0.03 -0.08 -0.07
(0.01) (0.01) (0.01) (0.01)
ρ 0.59 0.77
(0.22) (0.15)
logL -235.15 -232.76 -198.63 -192.46
BIC 268.21 268.83 237.71 234.54
QP S1 0.061 0.063 0.058 0.059
QP S2 0.327 0.310 0.256 0.254
CR50%
1 0.963 0.961 0.958 0.958
CR50%
2 0.725 0.740 0.816 0.826
LM ρ 8.14 26.21
p-value 0.000 0.000
Notes: See notes to Table 5.5. Variable 171−84 indicates a variable which takes value one at period from 1971
M1 to 1984 M12, and zero otherwise.

Although Model 4 seems to outperform its special cases, the estimated coef-
ficients of the explanatory variables are almost equal in all models in Tables 5.5

143
and 5.6. These findings confirm our earlier results that the changes in the Federal
funds rate and stock market returns appear to be the main predictive variables for
growth rate cycles also in the more general bivarite probit models.

1 1

0.9 0.9

0.8 0.8

0.7 0.7
Probability

0.6 0.6

Probability
0.5 0.5

0.4 0.4

0.3 0.3

0.2 0.2

0.1 0.1

0 0
70 72 75 77 80 82 85 87 90 92 95 97 00 02 05 07 70 72 75 77 80 82 85 87 90 92 95 97 00 02 05 07
Time Time

Figure 5.2: In-sample fitted values from the Model 1 presented in Table 5.6. Reces-
sion periods are depicted with shaded areas. Business cycle recession probabilities
are presented in the left panel, probabilities for growth rate cycle periods in the
right panel.

1 1

0.9 0.9

0.8 0.8

0.7 0.7
Probability

0.6 0.6
Probability

0.5 0.5

0.4 0.4

0.3 0.3

0.2 0.2

0.1 0.1

0 0
70 72 75 77 80 82 85 87 90 92 95 97 00 02 05 07 70 72 75 77 80 82 85 87 90 92 95 97 00 02 05 07
Time Time

Figure 5.3: In-sample fitted values from the Model 4 presented in Table 5.6. Reces-
sion periods are depicted with shaded areas. Business cycle recession probabilities
are presented in the left panel, probabilities for growth rate cycle periods in the
right panel.

Figures 5.2 and 5.3 depict the in-sample fitted values from Models 1 and 4
shown in Table 5.6. As the estimation results in Tables 5.5 and 5.6 suggest, the

144
business cycle recession periods predicted by these two models are almost identical.
For the growth rate recession periods, Model 4 gives somewhat more precise signals
around the years from 1984 to 1987 and before the year 2000.
In conclusion, especially the general bivariate autoregressive model (Model 4),
but also its special cases (Models 2 and 3), are superior to the independent uni-
variate autoregressive models for both cycle indicators (Model 1). These findings
indicate that superior predictive power can be found by using bivariate models
instead of univariate models.

5.4.4 Out-of-Sample Performance

In this section, we examine the out-of-sample predictive performance of different

models. The first out-of-sample nowcasts are made for January 2000 and the
last ones for December 2005. Notice that the out-of-sample period contains three
growth rate cycle recession periods, but only one business cycle recession in 2001
(see Figure 5.1). In addition to these out-of-sample nowcasts, in Section 5.4.5, the
best bivariate probit models, according to model selection criteria and in-sample
predictions provided in Section 5.4.3, are used to assess the state of U.S. economy
from 2006 to 2008.
Different predictive models are estimated using the data up to December 1999
assuming that the state of the economy is known at that time. To emulate real-
time forecasting, we should take into account the fact that the latest values of
the dependent variables are unknown at the time the forecast is made (see Section
5.4.2). Thus the out-of-sample exercise is carried out without updating the param-
eter estimates using the sample period up to December 1999. The same framework
is also applied in Section 5.4.5. It turns out that the model selection procedure
employed in Section 5.4.3, using the data set up to December 1999, yields the same
conclusions as obtained when using the whole sample in estimation. Therefore, the
differenced Federal funds rate and the stock market returns have predictive power
for the growth rate cycles also in this estimation period.
Table 5.7 shows the out-of-sample performance of the employed models. Out-

145
of-sample forecast accuracy is evaluated by the QP S and the percentage of correct
forecasts (CR). The first four models also include the dummy variable 171−84 . As
in Section 5.4.3, the bivariate autoregressive probit model (Model 4) yields the best
out-of-sample predictions although Model 3 also produces good forecasts. Based
on the predictability test of Pesaran and Timmermann (1992) the reported per-
centages of correct forecasts are statistically significant at the 1% level in models
including the 171−84 variable. Without the additional dummy variable the per-
centages of correct forecasts are lower and, consequently, the p-values are higher
and statistically insignificant at 5% level.

Table 5.7: Out-of-sample performance of different models.

QP S1 CR1 QP S2 CR2
Model 1 with 171−84 0.395 0.722 0.395 0.708
Model 2 with 171−84 0.435 0.722 0.395 0.708
Model 3 with 171−84 0.167 0.889 0.303 0.833
Model 4 with 171−84 0.130 0.903 0.307 0.847
Model 1 0.179 0.833 0.446 0.625
Model 2 0.228 0.819 0.446 0.625
Model 3 0.094 0.931 0.446 0.625
Model 4 0.108 0.944 0.608 0.611
Notes: The out-of-sample values of QP Sj and CRj , j = 1, 2, based on models mentioned on
the left. Explanatory variables included in the model are the same as in the estimation results
presented in Tables 5.5 and 5.6. Nowcasts from Model 4 with the dummy variable 171−84 (the
fourth model) are depicted in Figure 5.4.

Although the dummy variable 171−84 is useful in predicting the growth rate
cycle recessions, this is not the case for nowcasts of the business cycle recession
periods. However, the more general Models 3 and 4 outperform Models 1 and 2
even in this case. Note that the percentages of correct forecasts for business cycle
periods are statistically significant in all models presented in Table 5.7.
Figure 5.4 illustrates the out-of-sample performance of the unrestricted Model
4. It appears that the recession probabilities match well the realized values of
both cycle indicators. As depicted in the left panel, Model 4 predicts the business
cycle recession in 2001 really well, but the increase in the recession probability in

146
2002 weakens the performance of the model in terms of the values of the statistical
goodness-of-fit measures reported. The out-of-sample nowcasts depicted in the
right panel match the growth rate cycle periods well. For instance, when the 50%
threshold value for probabilitity forecasts to construct signal forecasts for growth
rate recessions and expansions is used, Model 4 gives the correct signal forecast
with approximately 85% accuracy (CR = 0.847).

1 1

0.9 0.9

0.8 0.8

0.7 0.7

0.6 0.6
Probability

Probability
0.5 0.5

0.4 0.4

0.3 0.3

0.2 0.2

0.1 0.1

0 0
99 00 01 02 03 04 05 06 99 00 01 02 03 04 05 06
Time Time

Figure 5.4: Out-of-sample nowcasts for business cycle recession (left panel) and
growth rate cycle (right panel) recession periods (shaded areas) using the bivariate
autoregressive probit model (Model 4) which contains also the dummy variable
171−84 given in Section 5.4.3.

In summary, the results confirm that the proposed general bivariate autoregres-
sive model (5.8) outperforms its restricted versions and the values of both cycle
indicators appear to be predictable also out of sample.

5.4.5 Predictions for 2006–2008

There has been great uncertainty about the state of the economy in the United
States during the last couple of years. Therefore, it is of great interest to consider
the probabilities of the business and growth rate cycle periods during 2006–2008.
Business cycle recession probabilities are of special interest due to the fact that
the NBER Business Cycle Dating Committee declared that a peak in the U.S.
economic activity occurred in December 2007 indicating that the value of the

147
recession indicator (5.23) has been one, at least in some months from January
2008.
In this section, we examine nowcasts of the state of the economy using the
best models found in Sections 5.4.3 and 5.4.4. The first predictions are made
for January 2006 and the last ones for November 2008. According to the recent
announcement of the NBER, the U.S. economy has been in business cycle recession
since the beginning of the year 2008. In the case of the growth rate cycle periods,
it is, however, not evident that there has been a constant downswing state (i.e. a
growth recession, y2t = 1) after January 2006. Thus, at the time this chapter is
written, the realized values of y2t are not known after January 2006.
Figures 5.5 and 5.6 depict the nowcasts from Models 1 and 4. The probabilities
in Figure 5.6 are based on the models including the additional dummy variable
171−84 . The evidence appears to be ambiguous between different models. It seems
that Model 1 produces greater business cycle recession probabilities than Model 4
at the beginning of the recession in January 2008 (see Figure 5.5). On the other
hand, in both cases, Model 1 gives higher “false” recession risks than model 4 for
some months before the recession started. Overall, the employed models seem to
nowcast the beginning of the recession period at 2008 reasonably well.
As discussed above, the latest values of the growth rate cycle indicator are
unknown and, thus, it is difficult to make comparisons between different models
with the currently available information. All in all, it seems that the growth rate
recession probability have been decreasing from mid-2006. As the business cycle
recession started at the beginning of 2008, it is likely that the U.S. economy has
been in a growth recession some months before. For those months, the predicted
probabilities in Figure 5.5 are quite high and exceed the 50% threshold value,
indicating a growth rate cycle recession. Note also that a decreasing probability
of growth rate recession affects the business cycle recession probability in Model
4. This is a potential reason why the business cycle recession probabilities in the
left panels of Figures 5.5 and 5.6 are lower than in the independent univariate
autoregressive model (Model 1).

148
 1 1

0.9 0.9

0.8 0.8

0.7 0.7

0.6 0.6
Probability

Probability
0.5 0.5

0.4 0.4

0.3 0.3

0.2 0.2

0.1 0.1

0 0
05 06 07 08 09 05 06 07 08 09
Time Time

Figure 5.5: Real-time predictive probabilities from Model 1 (dashed line) and from
Model 4 for the business cycle (left panel) and the growth rate cycle periods (right
panel) using the models described in Table 5.5. In the left panel, the business cycle
recession that began at 2008 is depicted with shaded area. In the right panel, the
shaded area corresponds to the time period of unknown values of the growth rate
cycle indicator y2t .

1 1

0.9 0.9

0.8 0.8

0.7 0.7

0.6 0.6
Probability

Probability

0.5 0.5

0.4 0.4

0.3 0.3

0.2 0.2

0.1 0.1

0 0
05 06 07 08 09 05 06 07 08 09
Time Time

Figure 5.6: Real-time recession probabilities from Model 1 (dashed line) and from
Model 4 for the business cycle (left panel) and the growth rate cycle periods (right
panel) using the models described in Table 5.6. In the left panel the business cycle
recession period that began at 2008 is depicted with shaded area. In the right
panel, the shaded area corresponds to the time period of unknown values of the
growth rate cycle indicator y2t .

149
5.5 Conclusions
We introduce a new bivariate time series model for binary dependent variables.
The bivariate autoregressive probit model is a bivariate extension of the univariate
autoregressive model of Kauppi and Saikkonen (2008), but it can also be considered
a dynamic extension of the static bivariate probit model of Ashford and Sowden
(1970).
The bivariate autoregressive probit model, and its special cases, are applied
to predict the current state of the U.S. economy using binary indicator variables
for the level and the growth rate of the U.S. economic activity. The proposed
bivariate model framework extends the traditional univariate analysis of business
cycle recession periods examined in the previous literature, where only business
cycle recessions have been considered.
We found strong in-sample and out-of-sample evidence in favor of the proposed
bivariate autoregressive probit model. The results suggest that it is possible to gain
additional predictive power by modeling the recession probabilities of the two cycle
indicators jointly instead of considering two independent univariate autoregressive
probit models. The lagged first difference of the Federal funds rate is the most
useful single predictor of the state of the growth rate cycle, but also monthly stock
market returns turn out to be statistically significant predictors for both cycle
indicators. As suggested in previous studies, a term spread between the long-term
and short-term interest rates is an important explanatory variable for business
cycle recession periods, but it turned out not to be the best predictive variable
for the growth rate cycle periods. We also found evidence that the probability of
growth rate recession was systematically higher in the 1971–1984 period than after
the mid-1980s.

References
Anatolyev S. 2009. Multi-market direction-of-change modeling using dependence
ratios. Studies in Nonlinear Dynamics and Econometrics 13, article 5.

150
Ashford JR, Sowden RR. 1970. Multivariate probit analysis. Biometrics 26: 535–
546.

Banerji A, Hiris L. 2001. A framework for measuring business cycles. International

Journal of Forecasting 17: 333–348.

Bernanke BS, Blinder AS. 1992. The Federal funds rate and the channels of mon-
etary transmission. American Economic Review 82: 901–921.

Bernard H, Gerlach S. 1998. Does the term structure predict recessions? The in-
ternational evidence. International Journal of Finance and Economics 3: 195–215.

Blanchard O, Simon J. 2001. The long and large decline in U.S. output volatility.
Brooking Papers on Economic Activity 1: 135–164.

Chauvet M, Potter S. 2005. Forecasting recession using the yield curve. Journal
of Forecasting 24: 77–103.

Chib S, Greenberg E. 1998. Analysis of multivariate probit models. Biometrika

85: 347–361.

Diebold FX, Rudebusch GD. 1989. Scoring the leading indicators. Journal of
Business 62: 369–391.

Dueker MJ. 2005. Dynamic forecasts of qualitative variables: A qual VAR model
of U.S. recessions. Journal of Business and Economic Statistics 23: 96–104.

Ekholm A, Smith PWJ, McDonald JW. 1995. Marginal regression analysis of a

multivariate binary response. Biometrika 82: 847–854.

Engle RF. 1984. Wald, likelihood ratio and Lagrange Multiplier tests in economet-
rics, in Griliches Z. and Intriligator MD. (eds.), Handbook of Econometrics, Vol.
II. Amsterdam, North-Holland.

Estrella A, Hardouvelis GA. 1991. The term structure as a predictor of real eco-
nomic activity. Journal of Finance 46: 555–576.

151
Estrella A, Mishkin FS. 1998. Predicting U.S. recessions: Financial variables as
leading indicators. Review of Economics and Statistics 80: 45–61.

Greene WH. 2000. Econometric Analysis. Fourth edition. Prentice-Hall Interna-

tional, London.

Kauppi H, Saikkonen P. 2008. Predicting U.S. recessions with dynamic binary

response models. Review of Economics and Statistics 90: 777–791.

Kiefer NM. 1982. Testing for dependence in multivariate probit models. Biometrika
69: 161–166.

Layton AP, Moore GH. 1989. Leading indicators for the service sector. Journal
of Business and Economic Statistics 7: 379–386.

McConnell MM, Perez-Quiros P. 2000. Output fluctuations in the United States:

What has changed since the early 1980’s? American Economic Review 90: 1464–
1476.

Mosconi R, Seri R. 2006. Non-causality in bivariate binary time series. Journal of

Econometrics 132: 379–407.

Nyberg H. 2010. Dynamic probit models and financial variables in recession fore-
casting. Journal of Forecasting 29: 215–230.

Osborn DR, Sensier M, van Dijk D. 2004. Predicting growth regimes for Euro-
pean countries, in Lucrezia Reichlin (eds.), The Euro Area Business Cycle: Stylized
Facts and Measurement Issues. Centre for Economic Policy Research.

Orphanides A, van Norden S. 2002. The unreliability of output-gap estimates in

real time. Review of Economics and Statistics 84: 569–583.

Pesaran HM, Timmermann A. 1992. A simple nonparametric test of predictive

performance. Journal of Business and Economic Statistics 10: 461–465.

Schwarz G. 1978. Estimating the dimension of a model. Annals of Statistics 6:

461–464.

152
Sensier M, van Dijk D. 2004. Testing for volatility changes in U.S. macroeconomic
time series. Review of Economics and Statistics 86: 833–839.

153