The Myopic Choice Between Fixed and Adjustable Rate
Mortgages in Flanders
Damen Sven and Buyst Erik
February 4, 2015
Abstract
Using data from Flanders (Belgium), we study the choice between …xed rate mortgages (FRM)
and adjustable rate mortgages (ARM). The results indicate that households base their decision on
initial mortgage payments and do not consider expected changes in ARM rates. This results in
a remarkable large variation in mortgage structure over time in comparison to other countries. An
important …nding is, however, that borrowers who expect to move in the near future are more likely to
choose the FRM. As all mortgages are portable by law, mobile borrowers prefer to reduce variability
in mortgage payments, which was not observed in previous empirical studies.
JEL classi…cation: D14, E43, G11, G21
Keywords: mortgage choice, …xed and adjustable rate mortgages, household …nance, bond risk premia
Sven Damen:
KU Leuven, Center for Economic Studies, Naamsestraat 69, 3000 Leuven, Belgium,
[email protected] (corresponding author). Prof. Dr. Erik Buyst: KU Leuven, Center for Economic Studies,
Naamsestraat 69, 3000 Leuven, Belgium.
1
Introduction
Many households have only one major asset, a house, which is usually …nanced through a mortgage
contract. One of the most important …nancial decisions a household has to make is therefore the choice
of mortgage contract. Most consumers make this complex choice only a few times in their lives. Borrowers
are thus likely to be ill-informed, which can lead to an ine¢ cient equilibrium if they make suboptimal
decisions (Woodward and Hall 2012). The major decision borrowers have to make is the choice between
the …xed rate mortgage (FRM) and the adjustable rate mortgage (ARM). The former has known nominal
payments over the whole mortgage term, whereas the latter exposes the borrower to interest-rate risk.
The resulting aggregate interest risk from individual mortgage choices has important policy implications for the transmission of monetary shocks through the e¤ect of mortgage payments on household
budgets. Miles (2004), Maclennan et al. (1998) and others have argued that a larger share of adjustable
rate mortgages causes a higher exposure to changes in the interest rate. The International Monetary Fund
(2004) pointed out that countries with high ARM shares typically have more volatile housing markets
and for given monetary policy, a higher proportion of …xed rate mortgages is welfare enhancing according
to the new Keynesian dynamic stochastic general equilibrium model by Rubio (2011). The dominant
share of adjustable rate mortgages in the UK renders their economy susceptible to di¤erent shocks than
other European countries (HM Treasury 2003a). The resulting asymmetric e¤ect of monetary policy was
mentioned as an important reason by the United Kingdom’s (UK) economics and …nance ministry (HM
Treasury 2003a, 2003b) for opting out in the third stage of the Economic Monetary Union to introduce
the euro.
Notwithstanding the di¤erent ways shocks are propagated, there still remains large cross-country
heterogeneity in mortgage structure within the Eurozone as shown in Figure 1. FRMs are traditionally
predominant in France and Germany1 , whereas Italy and Spain2 are often highlighted as countries with
adjustable rate mortgage debt. Country-speci…c institutional features play an important role in the
historical structure of the mortgage market in a particular country according to the European Mortgage Federation (2006, 2012) and the IMF (2004). The country-speci…c characteristics include cultural
characteristics such as the borrowers’ risk aversion or the funding sources of lenders. Countries with
well-developed covered bond markets or mortgage-backed securities markets tend to have a higher share
of FRMs (e.g. Denmark, Germany and the USA). ARMs traditionally make up a higher proportion
of the mortgage market in countries where funding is based on short-term deposits (e.g. UK, Spain,
Portugal).
1 and
2 and
Denmark (not shown in Figure 1)
Ireland, Portugal and Sweden (not shown in Figure 1)
2
The proportion of a mortgage type thus varies around a level that is dependent upon the countryspeci…c institutional features. Whereas the variation around the country-speci…c level is small in most
countries (see Figure 1), the structure of the Belgian mortgage market varied signi…cantly during the
last decade. From 2003 to 2012, the Belgian share of mortgages with an initial rate …xation period of up
to one year fell from 62% in 2004 to 8% in 2006, recovered temporarily to 58% in 2010, and fell again
to 5% in 2012, a remarkably large variation in comparison to other countries. In this paper we seek to
1
.8
.8
.6
.6
Year
Belgium
Germany
2012
2011
2010
2009
2008
2007
2003
2012
2011
2010
2009
2008
2007
0
2006
0
2005
.2
2004
.2
2006
.4
2005
.4
2004
Share
1
2003
Share
understand this variation in the mortgage structure.
Year
Italy
Spain
France
Belgium
(a) Initial …xed rate period <= 1 year
Germany
Italy
Spain
(b) Initial …xed rate period <= 10 years
Figure 1: Share of adjustable rate mortgages
Note: Figure 1a shows the share of mortgages with an initial …xed rate period smaller than or equal to 1 year
from 2003:1 to 2012:12 in Belgium, Germany, Italy, Spain and France. Figure 1b shows the share of mortgages
with an initial …xed rate period smaller than or equal to 10 years. As the available statistics do not allow us to
compute the statistic for France, this country is excluded from Figure 1b. Source: ECB (own calculations).
The choice between …xed and adjustable mortgages has been extensively studied in the United States
(USA) (e.g. Dhillon, Shilling and Sirmans 1987; Brueckner and Follain 1988; Coulibaly and Li 2009)
and the UK (Leece 2000; Bacon and Mo¤att 2012). Within the Eurozone, the FRM-ARM choice has
been examined by Paiella and Pozzolo (2007) and Zocchi (2013) in Italy, a country with a relatively low
developed mortgage market in comparison to the rest of Europe.
One contribution of this paper is to test the external validity of the expected determinants of the
ARM-FRM choice outside the USA. A recent contribution to the literature by Koijen, Van Hemert and
Van Nieuwerburgh (2009) argues that the bond term premium is the key theoretical determinant of time
variation in the ARM share in the USA. The term premium re‡ects the extra yield lenders demand for
holding the riskier long term bonds. We are, however, not aware of any other empirical study outside
the USA that includes this risk premium in the analysis of the ARM choice.
Belgium is, furthermore, of special interest to study mortgage choice as all mortgages are portable by
law. Every borrower has the right to transfer their mortgage to a new residence if the value of the new
house is high enough to serve as security in the mortgage contract. In the USA, portable mortgages are
3
a rarely-used option, as borrowers must have exceptional credit, and interest rates are typically higher
than for standard …xed rate mortgages. American borrowers who expect to move in the near future are
thus more likely to choose an ARM to reduce the cost of the pre-payment option and the term premium
in the FRM. An important contribution of this paper is to test the external validity of the association
between borrower mobility and the mortgage choice when all mortgages are portable by law. If mobile
borrowers have a deeper preference to reduce short term variability in mortgage payments, they may
have a preference for FRMs that was not observed in previous empirical studies.
In the next section we give an overview of the related literature on the determinants of the ARM
choice. We discuss price and macro variables that are expected to explain the large variation over time
in the Belgian ARM share, and borrower and mortgage characteristics that are possible determinants of
cross-sectional variation in the FRM-ARM choice. The data section describes the datasets that we will
use in the empirical study thereafter. We conclude in the …nal section.
Determinants of the ARM choice
Variation in the ARM share over time
Previous empirical studies document the important role of price and macro variables such as the FRMARM spread and the yield curve. It is argued that a larger FRM-ARM spread indicates that the ARM
becomes relatively cheaper and hence more attractive, ceteris paribus. The overview of the empirical
…ndings in the existing literature in Table 1 indeed con…rms the positive relationship between the FRMARM spread and the share of ARMs. The di¤erence between the long,- and short-term government bond
rates (yield curve) is often included as a proxy to control for future interest rate expectations. If the
expectations hypothesis of the term structure of interest rates holds (that long yields are the expected
sum of short yields), a larger yield spread indicates higher future interest-rate expectations. This will,
ceteris paribus, tend to discourage the ARM choice. Table 1 con…rms the negative e¤ect of the yield
spread on the ARM share in the empirical literature, conditional on the control variables.
There is, however, abundant empirical evidence against the expectations hypothesis of the term
structure of interest rates (see Fama and French (1989), Campbell and Shiller (1991) and Cochrane
and Piazzesi (2005) among others). Koijen, Van Hemert and Van Nieuwerburgh (2009) study the link
between the term structure and mortgage choice. A major contribution to the literature is that the
bond term premium is the key theoretical determinant of rational mortgage choice. This risk premium
re‡ects the extra yield lenders demand for holding the riskier long term bonds. The authors show that
the yield spread or FRM rate are poor proxies for the premium, as they introduce an errors-in-variables
problem due to the variation in both interest rate expectations and changes in the risk premium. Koijen
4
et al. argue that American households could be making close-to-optimal mortgage decisions as the risk
premium explains a large part in the time variation in the aggregate ARM share.
Although Koijen et al. (2009) argue that the risk premium is the key determinant of mortgage choice,
the overview of the literature in Table 1 indicates that it is still the only study that explicitly includes
the risk premium in an econometric model of ARM – FRM choice. Even recent studies do not include
the risk premium as a determinant of mortgage choice. An important contribution of our study is to
test the external validity of the role of the risk premium as determinant of mortgage choice outside the
USA.
Nothaft and Wang (1992) and Berkovec, Kogut and Nothaft (2001) also include the National Association of Realtors’a¤ordability index, which incorporates income and mortgage rates in a nonlinear
fashion. If households are borrowing constrained and the FRM becomes una¤ordable, the authors expect the ARM demand to increase when housing becomes less a¤ordable to meet the payment-to-income
quali…cation standards. This is con…rmed by their empirical results.
Cross-sectional variation in the ARM choice
The empirical results with respect to borrower and mortgage characteristics in the second panel of Table
1 are relatively mixed. Many borrower characteristics that are included in empirical studies are not
signi…cantly di¤erent from zero, which is in line with perfect …nancial markets where all information is
re‡ected in the price of the contract.
One important …nding in all previous empirical studies is, however, that households who are more
mobile or expected to move in the near future, are more likely to choose an adjustable rate mortgage.
Most American borrowers with a FRM who decide to move have to prepay their FRM and take out an
entire new mortgage as a large share of the mortgage contracts are not portable in the USA. Because
these borrowers are expected to prepay their mortgage relatively quickly, they are less a¤ected by the
interest-rate risk and therefore less willing to pay the term premium or pre-payment option of the FRM
(Campbell and Cocco 2003). Conditional upon an upward sloping yield curve, the ARM is the cheapest
contract for these households.
In countries where mortgages are portable when the borrower decides to move, the household is able to
lock-in favorable FRM rates and the association between higher mobility and ARM choice does not need
to exist as borrowers do not have to avoid the term premium or the cost of the pre-payment option of the
FRM. In fact, if households who expect to move in the near future prefer to reduce short-term variability
in real mortgage payments, mobile households may be less likely to choose the ARM. A contribution of
our paper is to test the external validity of the relationship between borrowers who expect to move in
5
the near future and the probability of choosing an ARM in a country where all mortgages are portable
by law.
Among the statistically insigni…cant variables in the existing literature is the presence of children in
the household, which is expected to be associated with a higher risk aversion, larger housing demand,
and higher concern about future consumption and therefore lower probability of ARM choice (Brueckner
and Follain 1988). Also self-employment is not found to be a statistically signi…cant determinant of
mortgage choice, which may be related to multiple theoretical predictions. Self-employment may be
associated with higher risk (Leece 2000) and therefore lower probability of ARM choice or on the other
hand higher expected future income (Zocchi 2013) or cheaper borrowing on their own account than in
their …rm (Dhillon et al. 1987) and thus a higher probability of ARM choice.
Households with higher education are likely to be more …nancially sophisticated (Coulibaly and Li
2009) or have better career prospects such that the probability of ARM choice increases. Empirical
evidence indicates that the …ndings with respect to higher education are mixed. Many existing studies
…nd a non-signi…cant or positive e¤ect of higher education on ARM choice. Zocchi (2013) even reports
a positive e¤ect of low education on ARM choice in Italy, which the author associates with less cautious
behavior.
Theoretical work by Campbell and Cocco (2003) describes that the ARM-FRM choice involves a
basic trade-o¤ between two types of risk: wealth risk and income risk. The FRM is subject to the
former as its real capital value is dependent upon unknown future in‡ation. The prepayment option
protects the borrower against the risk of a decreasing nominal interest rate, but this option does not
come for free. The wealth risk is not present in the ARM. The risk of an ARM is the variability in
necessary mortgage payments due to the variability in interest rates. Variability in mortgage payments
can result in unpleasant consumption variability as most households are risk-averse and hence want a
smooth consumption pattern. Campbell and Cocco argue that the ARM is less attractive for a risk-averse
household with a relatively large mortgage or large houses relative to income. Both determinants will
have an important role in our econometric estimation, even though the existing empirical literature with
respect to income, repayment to income and loan amount found mixed results.
Data
The previous section provided possible determinants of variation in the ARM share over time and crosssectional variation between borrowers. In this section we describe the time series and micro-data that
we will use in the empirical analysis.
6
Time series
To study the time variation in mortgage choice we use aggregate time-series which allows us to study the
whole population of mortgage choices in Belgium. The time series are, furthermore, very suited to study
the role of macro variables as data are available at monthly frequency. Data are from the Professional
Association of Credit Providers (Beroepsvereniging van het krediet - BVK). The series provides business
volumes of new mortgages for di¤erent initial …xed-rate periods. We classify a mortgage as an adjustable
rate mortgage if the initial period of rate …xation is smaller than 10 years. As most mortgages in Belgium
have …xed rates over the whole mortgage term or for one year only, this ARM share is highly correlated
(95%) with a series constructed from ARMs with an initial …xed rate period smaller than 1 year (which
is the legal minimum in Belgium). Mortgage interest rates were taken from the Monetary Financial
Institutions Interest Rates (MIR) survey. The main purpose of the MIR survey is to obtain harmonized
European aggregates of rates intended for households and non-…nancial …rms. The statistics include the
mortgage interest rates for loans that cover house purchases from 2003:01 to 2012:12 for di¤erent periods
of initial rate …xation3 .
Belgian 10-year and 1-year government bond yields are taken from the OECD main economic indicators database. The government bond spread is calculated as the di¤erence between these two series.
The OECD database also provides in‡ation rates for Belgium. Data is available at monthly intervals
spanning 1960:01 to 2012:12.
Micro-data
To examine which borrower characteristics lead the household to prefer one mortgage type over the other,
we use data from the Flemish Housing Survey 2005 (Woonsurvey 2005 )4 and 2013 (Groot Woononderzoek
2013 ). Flanders is the Dutch-speaking northern part of Belgium. The region covers 58% of the Belgian
population. Similar housing surveys from the other regions are not available. The surveys are designed
to assess the needs and preferences for housing in Flanders. The samples are drawn from the National
Register, which resulted in a sample of 5,216 and 10,013 valid interviews for respectively the 2005 and
2013 survey. The surveys include detailed household characteristics: employment status, disposable
income, occupation category, education levels, age, presence of children in the house and the likelihood
of moving. Mortgage information includes the current interest rate, whether or not this rate is …xed,
sale price of the house and the amount that the household chose to borrow. The FRM-ARM spread is
taken from the MIR5 . We limit the sample, as we are only interested in choices by households that are
3 For the FRM rate we use the average monthly rate of mortgages for house purchases with an initial …xed rate period
longer than 10 years. For the ARM rate we use the rate with an initial …xed rate period equal to 1 year.
4 More information on the survey and results can be found in: Heylen K., Le Roy M., Vanden Broucke S., Vandekerckhove
B. & Winters S. (2007). "Wonen in Vlaanderen; De resultaten van de woonsurvey 2005 en de Uitwendige Woningschouwing
2005," Departement Ruimtelijke Ordening, Woonbeleid en Onroerend Erfgoed, Woonbeleid, Brussel.
5 ARM rates for 2001 and 2002 were imputed from a regression of the ARM rate on the short term government bond
rate and government bond spread (including quadratic terms). FRM rates for 2001 and 2002 were imputed from a similar
7
in the early years of mortgage payments. We thus select households who bought a house since 2001 or
2009 for respectively the 2005 and 2013 survey, and examine how their characteristics at the time of the
interview a¤ect their mortgage type in that year. As the households are randomly drawn from the whole
Flemish population, they are not subject to a selection bias, which is an important advantage over data
from single mortgage providers. Data from a single mortgage provider would constrain our study to the
…nancial advice of the lender, while the Flemish Housing Survey allows us to study the mortgage choices
of the whole population. We use weights to adjust for unequal probabilities of selection in the survey
sampling design.
Whereas other studies have to use a proxy for the likelihood to move (Dhillon, Shilling and Sirmans
1987; Brueckner and Follain 1988), the Flemish Housing Survey directly enquires about the borrowers’
moving plans. We code the household as expected to move when the answer to the question "Do you
have plans to move?" is "Yes".
Similar to Nothaft and Wang (1992) and Berkovec, Kogut and Nothaft (2001) we include a measure
of a¤ordability. Our measure is a ratio of the loan amount the household can borrow divided by the house
price, assuming that the household spends 30% of income and the full …scal advantage for homeowners
on a 20-year mortgage. The …scal advantage is the net tax relief the household would receive in the
year of mortgage origination given these mortgage characteristics. Before 2005, the tax deduction for
homeowners was mainly based on capital repayments that were limited to the …rst 60,910 euros of the
borrowed amount per dwelling in 2004. The limit was higher for households with children, and increased
in accordance with the consumer price index over time. The …scal advantage was furthermore limited
in function of income and could not exceed 1,870 EUR in 2004. Since 2005, the …scal advantage for
homeowners is extended to a tax deduction (woonbonus) that includes both interest payments and
capital repayments. The new …scal advantage limits the tax deduction to 2,490 euros per year per person
in the …rst ten years, starting in 2005. For the average household the net tax relief more than doubled.
The tax relief is dependent upon the marginal tax rate of the borrower, which is not observed in the
survey. The marginal tax rate had to be computed based on the net income reported by households.
The a¤ordability index for household j at time t is then given by equation 1. YtF RM is equal to the FRM
rate at the time of the purchase. A higher value indicates a more a¤ordable house relative to income.
(0:3 incomej;t + netT axReliefj;t )
Af f ordabilityj;t =
(1 (1+YtF RM )
YtF RM
housepricej;t
20
)
(1)
Descriptive statistics of all relevant variables are presented in Table 2.
regression of the FRM rate on the long term government bond rate and government bond spread (including quadratic
terms).
8
Empirical results
In this section we empirically examine how the theoretically expected determinants a¤ect mortgage
decisions over time and between households. We …rst analyze the role of the FRM premium that lenders
demand, the bond term premium and the initial mortgage spread in the evolution of the ARM share from
2003:01 to 2012:12. Given a …xed distribution of households, the FRM premium that lenders demand
should be the key determinant of variation in the mortgage structure over time. Thereafter we examine
how households choose their mortgage type using micro-data from the Flemish Housing Survey.
The bond term premium as key theoretical determinant of rational mortgage
choice
Koijen et al. (2009) derive the optimal mortgage choice within a lifetime utility framework. Within
this framework, households trade o¤ higher expected real consumption against higher variability in
real consumption. With all else equal, households prefer the mortgage that gives them the highest
expected real consumption. When the risk premium on long term bonds is high, the FRM becomes more
expensive, such that the ARM becomes more attractive. In order to derive the risk premium, note that
the government bond spread can be written as
n
Yt (n)
Yt (1) =
1X
Et [Yt+j 1 (1)]
n j=1
|
{z
Yt (1) +
}
Pure EH-spread
t (n)
| {z }
(2)
Risk prem ium
The …rst term is the spread under the pure expectations hypothesis and equals the average expected
increase in short term rates. The second term is the risk or term premium. If we include the spread
and do not di¤erentiate between both components, we thus observe an errors-in-variables problem with
a downward-biased estimate of the risk premium e¤ect.
In order to disentangle both terms in (2) we jointly model the dynamics of government bond yields
and in‡ation in a vector error-correction model (VECM). This model will be able to forecast future
values of the short term rate, such that the risk premium can be calculated as a residual value from
equation (2). The VECM takes the form
Xt = v +
Xt
1+
p 1
X
i
Xt
1
+ "t
i=1
where Xt is a k dimensional column vector of k state variables,
and
i
are k
1 vector of constants,
k matrices of coe¢ cients and "t is a vector white noise process.
information of the data and can be written as
are k
is a k
=
0
contains long-run
under the hypothesis of cointegration.
and
r matrices, where r equals the number of linear combinations of the elements of Xt that are
9
stationary. The matrix
can be interpreted as a matrix of r cointegrating vectors, while matrix
consists of error-correction parameters. The state vector Xt includes the short term government bond
rate and the 10-year government bond spread. Ang and Piazzesi (2003) show that incorporating both
yield-curve factors and macroeconomic factors are important to capture the dynamics of the yields. As
a macroeconomic variable we include the realized one-year in‡ation ( t ) such that the complete state
vector Xt is equal to
0
Xt =
Yt (1)
Yt (10)
Yt (1)
t
We use a VECM instead of a vector autoregressive model (VAR) as the one year government bond rate
and in‡ation rate are integrated of order one, or I(1). Two baseline models are used for estimation, we
refer to them as models A and B. The main di¤erences are summarized in Table 3. Model A uses a
sample of monthly data from 1960:01 to 2012:12, but uses an expanding window of the estimation sample.
Therefore we estimate the VECM using a subsample from 1960:01 to 2002:12 and forecast the one-year
government bond rate over the next 10 years, starting from 2003:01. The dynamic forecasts deliver a
measure of the expected average short rate over the next 10 years. We then increase the subsample to
include 2003:01, re-estimate the model, forecast the short rate and calculate the expected average short
rate over the next 10 years starting from 2003:02. The procedure is repeated until we obtain a series
of the expected average short rate from 2003:01 until 2012:12. Model B estimates the VECM using a
…xed estimation sample from 1999:01 to 2012:12. As monetary policy is conducted independently by the
European Central Bank since 1999, model B can be seen as a robustness check using a recent estimation
sample.
The Johansen test for cointegration indicates that there exists one cointegration equation with cointegration vector
0
=
1
4:28
0:72
7:84
when the VECM is estimated over the whole sample
period. The coe¢ cients refer respectively to the one year government bond rate, the 10-year spread, the
realized one-year in‡ation rate and a constant. The eigenvalue stability condition is also satis…ed given
that all eigenvalues are inside the unit circle except two (which equals the number of I(1) variables). A
graphical representation is given in Figure 5 in the appendix. We therefore assume one cointegration
vector in both model A and B.
Belgium is the only Member State of the European Union that implemented a ceiling on the variability
of the mortgage rate (i¤/ZEW 2010). As required by law, it is now standard practice for adjustable rate
mortgages to have a lifetime cap of 3% according to the Financial Stability Review of the National Bank
of Belgium (2012). Therefore, we limit the forecast of future interest rate expectations such that the
absolute deviation from the initial interest rate cannot exceed 3 percentage points. The adjusted measure
of expected changes in interest rates is then equal to the true expectations that households face.
10
While borrowers are protected from large shifts in interest rates, an increase in the interest rate of
3 percentage points might still have a signi…cant e¤ect on households’…nances. A survey in the United
Kingdom that asked a large sample of households how they would cope with a rise of interest rates
of 2.5%, found that less than half of the borrowers said they would be able to cope with all of their
borrowing commitments without any di¢ culty (FSA 2004 as cited in Miles 2004). Borrowers should
thus attach importance to interest rate expectations.
In Belgium, the lender is unable to change the ARM rate at its own discretion during the contract.
The ARM rates are linked to a reference index, which is computed from the government bond rates.
The expected increase in the ARM rate is thus exactly equal to the expected increase in the government
bond rate such that:
n
1X
ARM
Et [Yt+j
1 (1)]
n j=1
n
YtARM (1) =
We are now able to calculate the risk premium (
1X
Et [Yt+j
n j=1
t (n))
1 (1)]
Yt (1)
as a residual value from equation (2) in
the bond market and the premium that is present in the FRM-ARM di¤erential (
F RM ARM
(n))
t
as a
residual value from equation (3):
F RM
t
ARM
(n)
0
n
X
ARM
@1
Et [Yt+j
1 (1)]
n j=1
= YtF RM (n)
YtARM (1)
= YtF RM (n)
1X
ARM
Et [Yt+j
1 (1)]
n j=1
n
As lenders might demand an additional premium,
of the FRM contract that borrowers face than
t (n)
F RM ARM
(n)
t
1
YtARM (1)A
(3)
(4)
may better capture the true cost
alone. If we rewrite
F RM ARM
(n)
t
as in equation
(4), we can see that it is equal to the FRM-ARM interest spread adjusted for expected changes in the
ARM rate. If borrowers are not myopic and attach value to future ARM rates,
F RM ARM
(n)
t
should
have a higher explanatory power of the variation in the ARM share than the FRM-ARM interest spread.
The estimated measures of
t (n)
and
F RM ARM
(n)
t
are graphically represented over time in Figure 2.
We can see that risk premia fell until the eruption of the recent …nancial crisis, followed by an increase
in risk premia.
Now that we have a measure for the risk premium, we can examine its role upon the variability of
the ARM share in Belgium. The results of a regression with the ARM share as dependent variable and
the risk premium (where n equals 10 years) as regressor are shown in Table 4. A 1 percentage point
increase in the risk premium (
t (n))
increases the ARM share by 15 percentage points, according to
model A. A 1 standard deviation increase in the risk premium increases the ARM share by 8 percentage
11
.05
.03
.04
.02
.03
.01
.02
0
.01
Year
Term Premium (A)
(a)
2013
2012
2011
2010
2009
2008
2007
2006
2005
2003
2013
2012
2011
2010
2009
2008
2007
2006
2005
2004
2003
2004
-.01
0
Year
Term Premium (B)
Term Premium (A)
t (n)
(b)
F RM
t
Term Premium (B)
ARM
(n)
Figure 2: Risk premium
Note: This …gure shows (a) the bond risk premium and (b) the FRM premium that lenders demand. Both
measures are calculated for models A and B.
points. The predictive power of the risk premium is furthermore unsatisfactory. The R-squared of 17%
and 14% teaches us that the explanatory value of the risk premium alone is relatively small. Koijen
et al. (2009) report stronger sensitivities of a 1 standard deviation increase in risk premia and higher
explanatory power in the USA. The FRM premium that lenders demand (
e¤ect according to model A: a 1 standard deviation increase in
F RM ARM
(n))
t
F RM ARM
(n)
t
has a smaller
increases the ARM
share by 6.75 percentage points. According to model B the e¤ect is more important: a 1 standard
F RM ARM
(n)
t
deviation increase in
increases the ARM share by 18.31 percentage points. The R-
squared increases to 39%, such that a larger part of the variation in the ARM share can be attributed
to variation in
F RM ARM
(n).
t
In the next subsection we examine other factors that could explain the
variation in the ARM share.
F RM ARM
(n)
t
If
is mean-reverting, the borrower can temporarily choose the ARM when risk premia
are high in order to repay the entire loan and choose an FRM when risk premia have reverted. If the
variation in
F RM
t
ARM
(n) is high enough to cover the cost of taking out an entirely new mortgage, this
can be a pro…table strategy6 . Whether or not households employ this strategy is an empirical question.
If they do, a decrease in
F RM ARM
(n)
t
will, ceteris paribus, cause an increase in re…nance volumes.
Table 8 in the appendix presents regression results with the volume of re…nances as dependent variable
and
F RM ARM
(n)
t
as explanatory variable. We thus expect a signi…cant negative e¤ect if borrowers
exploit this strategy. The estimated coe¢ cient of the FRM-ARM premium is, however, positive in all
speci…cations. Borrowers thus do not temporarily choose the ARM when risk premia are high in order
to re…nance and obtain a FRM when
F RM
t
ARM
(n) decreased.
We provide additional robustness checks in the appendix. Hjalmarsson and Österholm (2007) argue
6 The
authors would like to thank Viggo Nordvik for this suggestion.
12
that there is little a priori reason to believe that interest rates and in‡ation rates have an exact unit root.
Conventional unit root tests have very limited power to distinguish a unit root from a near-integrated
process. The authors argue that the risk of concluding that completely unrelated near-integrated series
are cointegrated cannot be ignored, such that the VECM speci…cation becomes inappropriate. We
therefore frame the state variables in a VAR framework to check the structural validity. Our …ndings
are robust to all di¤erent speci…cations.
The role of initial mortgage payments in the rise and fall of the ARM share
We have shown that changes in the risk premium are insu¢ cient to explain the variation in the ARM
share in Belgium. Theoretical work by Alm-Follain (1987) and Brueckner (1986) suggests that households
should respond to the spread between FRM and ARM rates. A higher spread increases the current cost
of the FRM, but also the expected future ARM rates. Brueckner (1986) indicates that the former e¤ect
is likely to dominate such that more households would prefer ARMs. A higher spread due to an expected
increase in short term interest rates increases the FRM rate and the expected future cost of the ARM,
but allows the household to smooth consumption as he will make larger payments when income has
grown. The overall e¤ect is that we can expect an increase in the ARM share. The FRM-ARM interest
spread is, however, a combination of
F RM ARM
(n)
t
and the expected increase in ARM rates. The
responsiveness of the ARM share to the FRM-ARM spread should thus be smaller than the e¤ect of
F RM ARM
(n)
t
on the ARM share, as the latter only has a cost e¤ect. Column (1) in Table 5 presents
regression results for the share of adjustable rate mortgages as dependent variable and the FRM-ARM
spread as regressor. The FRM-ARM rate spread is highly positively correlated (0.89) to the share of
adjustable rate mortgages over the period 2003:01 to 2012:12. A higher spread thus increases the demand
for adjustable rate mortgages. A 1 percentage point increase in the FRM-ARM spread causes an increase
in the ARM share of 26 percentage points. The predicted ARM share with regression (1) is shown in
Figure 3 [Predicted (1)]. The series tracks the evolution of the actual ARM share remarkably closely.
Regressions (2), (3) and (4) in Table 5 include the government bond spread and our estimated change
in ARM rates from the VECM of models A and B. The coe¢ cient of yield curve in (2) is negative, but
insigni…cant. We are thus unable to reject that the responsiveness of the ARM share to changes in the
yield curve is statistically indi¤erent from zero. A remarkable di¤erence is the e¤ect of a change in the
government bond spread between the USA and our data. Berkovec, Kogut and Nothaft (2001) found
that a 1 percentage point increase causes a 13.33 percentage point decrease in the ARM share. The same
increase in the government bond spread in Belgium would only cause the ARM share to decrease by
2.24 percentage points (although this is not signi…cant). Borrowers in the USA are thus six times more
responsive to the slope of the yield curve. A possible explanation could be that the government bond
13
spread is not a good measure of the interest expectations that households face. The Belgian Mortgage
Credit Act requires the use of a lifetime cap over the whole mortgage duration7 , which often takes the
form of maximum changes of 3 percentage points, according to the National Bank of Belgium (2012).
Mortgages in the USA may also have interest rate caps and life-of-loan rate caps are also obligated
by law, but an increase in interest rate above the cap might carry over to future periods8 and lifeof-loan rate caps are generally higher than 3%. The carry over property and higher life-of-loan rate
caps can explain why borrowers in the USA are more responsive to changes in the government bond
spread. The interest rate cap in Belgium can lead to an errors-in-variables problem in the yield spread
such that the estimated coe¢ cient is biased towards zero. We therefore include the estimated change in
ARM rates from the VECM of models A and B. As the variable is constructed such that the absolute
di¤erence between the ARM rate and every forecast of the ARM rate cannot exceed 3%, it is equal to
the true expectations that households have. The estimated coe¢ cients in (3) and (4) have indeed larger
negative values, but only the interest expectation in (4) is statistically signi…cant. Whereas the estimated
coe¢ cient in (4) is statistically signi…cant, Figure 3 shows that its economic signi…cance is small. An
Pn
ARM
increase in n1 j=1 Et [Yt+j
YtARM (1) of 1 standard deviation decreases the ARM share by only
1 (1)]
2.31 percentage points. Given the important role of the FRM-ARM spread in the choice of mortgage
type, we choose to include it as a main pricing variable in our micro data analysis using the Flemish
Housing Survey in the next subsection.
.8
.6
.4
.2
2013
2012
2011
2010
2009
2008
2007
2006
2005
2004
2003
0
Year
Actual
Predicted (1)
Predicted (4)
Figure 3: Actual and predicted share of ARM mortgages
Note: The …gure shows the actual share of mortgages with an initial period of interest rate …xation smaller than
10 years (black line) from 2003:01 to 2012:12. The broken lines present the …tted values of regression (1) and (4)
in Table 5.
7 See
8 See
article 9, §1, 8 of the Belgian Mortgage Credit Act (Wet Hypothecair Krediet)
the Consumer Handbook on Adjustable-Rate Mortgages (2007), published by the Federal Reserve Board.
14
Given the data and the current time period, all results seem to suggest that Belgian households could
be making mortgage choices that are not optimal, as they attach too much weight to initial mortgage
payments. To summarize: the results show that 1) The ARM share is more responsive to changes in
the FRM-ARM spread than to changes in the risk premium; a remarkable result that contradicts the
theoretical expectations from Koijen et al. (2009). 2) The explanatory power of the FRM-ARM spread
is furthermore much higher than the predictive power of the FRM premium that lenders demand. 3)
The economic signi…cance of interest rate expectations is small.
Who chooses the ARM?
In this section we examine the association between household characteristics and the ARM choice, using
the micro-data from the Flemish Housing Survey (2005 and 2013). We estimate a binary logit model
where the dependent variable equals 1 if the household chooses an ARM. Empirical results of the binary
logit model are presented in Table 6. The results for the full sample are discussed in the …rst subsection.
Thereafter, we divide the sample in a below- and above median income segment, which reveals some
remarkable di¤erences between both segments. Finally, we re‡ect on the possibility of uninformed
borrowers and changes since the …nancial crisis.
Empirical results from the full sample
The …rst column of Table 6 presents the results for the full sample. In line with earlier results, the FRMARM spread is found to be an important determinant of mortgage choice. The variable is statistically
signi…cant at the 0.01 critical level and its economic signi…cance is also strong. A computation of the
marginal e¤ect from an increase of 1 percentage point from the mean spread, keeping all other covariates
…xed at their means, indicates that the probability of choosing an ARM increases by 18 percentage
points. The economic signi…cance of the spread is smaller than the earlier …ndings from the time series.
A possible explanation can be found in an errors-in-variables problem, as the FRM-ARM spread had to
be approximated using the average spread in the year of house purchase according to the MIR survey. We
are unable to improve the estimation by including the risk premium, which should be the key theoretical
determinant of mortgage choice according to Koijen et al. (2009). If we include the risk premia from
the analysis with the time series data, they are not statistically signi…cant, enter with the wrong sign or
have a weaker economic signi…cance than the FRM-ARM spread (results not reported here). The results
con…rm the earlier …ndings that borrowers do not look beyond initial mortgage payments.
An important result is the negative association between households who plan to move and the ARM
choice. To the best of our knowledge, we are the …rst to document a statistically signi…cant negative
relationship between expected mobility and ARM choice, which is di¤erent from all previous empirical
15
studies as summarized in Table 1. A possible explanation is that borrowers in Belgium do not need to
prepay the mortgage if they decide to move as all mortgages are portable by law. Therefore, households
do not need to avoid the higher initial mortgage rate from the FRM or the cost of the pre-payment option
if they receive higher utility from the FRM. The negative sign does indeed indicate that households who
expect to move in the near future, prefer to reduce short-term variability in real mortgage payments by
means of a FRM. The preference for FRMs from households who expect to move in the near future was,
however, not observed in earlier work as conventional mortgages are not portable in the US.
Age and higher education (both proxies for career and income expectations) are not signi…cant in the
full sample. The partner dummy and presence of children are also not statistically signi…cant.
The a¤ordability index does not appear to be signi…cantly associated with the mortgage choice in
the full sample, but the LTV ratio is statistically signi…cant at the 1 percent level. Conditional on the
other covariates, this indicates that borrowers with relatively large mortgage amounts are more likely
to be associated with an adjustable rate mortgage. Given the possible endogeneity between the chosen
mortgage type and the loan-to-value ratio, the coe¢ cients should be interpreted as descriptive only. An
important source of possible endogeneity is the risk premium. When the risk premium is high, the FRM
is relatively less attractive. If households who still prefer to borrow with a FRM adjust their desired loan
amount downward, a positive association between ARM choice and loan-to-value ratio may be found. In
other speci…cations where we control for di¤erent measures of the risk premia (not reported here), the
estimated coe¢ cient of loan-to-value ratio remains unchanged, indicating that this source of endogeneity
does not a¤ect the results. Another possibility is that borrowers who expect their income to increase in
the future, are able to borrow more with the ARM as the required mortgage payment to initial income
ratio is less likely to be binding. The next subsection, however, provides evidence that ARM borrowers
were not more likely being denied a mortgage, which contradicts the hypothesis that those borrowers
choose the ARM because they were denied a FRM due to a binding mortgage payment to income ratio.
In addition, we do control for higher education and age (expectations of future income), which are both
not signi…cant in the estimation. For these reasons, we conclude that the positive sign contradicts the
theoretical expectations of Campbell and Cocco (2003), who show in a life-cycle model that the ARM is
less attractive for households with relatively larger mortgages.
Unfortunately we only observe income, moving plans and unemployment at the time of the survey
interview instead of the time of the house purchase. Therefore we include an interaction with unemployment, and the moving plans dummy and the a¤ordability index. We do this to control for possible
changes in moving plans or income that is related to changes in employment at the time of the survey that
were not known at the time of the house purchase. Both interaction terms are, however, not signi…cant.
16
Division into two segments
Following the work of Schwartz (2009) and Bergstresser and Beshears (2010) we divide the sample into
two segments: a below median income (column 2) and an above median income segment (column 3).9
In the below median income segment, 52% of the households say that they are unable to save money,
according to the Flemish Housing Survey in 2005. In the above median income segment, only 26% of
the households are unable to save. A comparison reveals that higher educated households in the high
income segment (column 3) are more likely to choose an adjustable rate mortgage. We expect that they
have better long-term career perspectives and should prefer ARMs as they can handle an unexpected
increase in interest rates. The negative signi…cant e¤ect of age on the probability of choosing an ARM is
consistent with the view that younger households are in the early years of their career and hence expect
a more rapid increase in income than older borrowers (Brueckner and Follain 1988). In contrast to the
estimation results from the full sample, the partner dummy is statistically signi…cant. The negative
e¤ect is consistent with the …ndings of Bacon and Mo¤att (2012) that multiple borrowers are more risk
averse and hence more likely to choose a FRM. The interaction of unemployment and moving plans is
omitted in the above median income segment as only one unemployed respondent in the high income
segment expects to move in the near future. We drop this observation for the estimation in column
(3). Relatively higher income households with more a¤ordable houses relative to income have a higher
association with the ARM choice, which is in line with the results from Campbell and Cocco (2003) who
argue that households with smaller houses relative to income should prefer ARMs.
In the below median income segment, the a¤ordability metric enters remarkably with the opposite
negative sign, although it is not signi…cant. The statistically signi…cant coe¢ cient of LTV in the full
sample is, furthermore, mainly driven by low income households. While …nancially constrained households in the lower income segment should not prefer a mortgage whose monthly payments can ‡uctuate,
lower income households with relatively higher loan amounts prefer adjustable rate mortgages against
theoretical expectations of Campbell and Cocco (2003).
It would, however, be possible to rationalize this behavior if these households would prefer to choose
a FRM, but were constrained and had to choose an ARM to lower the initial mortgage payments below a
certain threshold. Campbell (2013) argues that borrowers who are currently borrowing-constrained tend
to prefer ARMs. With an upward-sloping yield curve, the FRM rate may be too expensive during the
initial payments such that the lender is unwilling to provide this mortgage. Johnson and Li (2014) …nd
evidence that ARM borrowers in the USA are currently borrowing-constrained. The authors show that
ARM borrowers are more likely to be turned down for credit than FRM borrowers. The Flemish Housing
9 We use average income per parent to split the sample. We do this to ensure that single borrowers are not necessarily
in the low income group. We split the sample separately in the 2005 and 2013 survey to ensure that 2005 households are
not necessarily in the low income group due to positive income growth over time.
17
Survey in 2005 directly asks whether they were easily granted credit. From the FRM and ARM borrowers,
respectively 4% and 2% borrowers answered that a bank was unwilling to provide a mortgage. Flemish
ARM borrowers thus do not report being denied a mortgage loan more often than FRM borrowers.
Higher educated and younger households (both proxies for good future income expectations), are also
not more likely to choose the ARM in the below median income segment. For these reasons, we attribute
the positive sign of LTV in the below median income segment to suboptimal mortgage choices.
In an additional estimation (not reported here), we also include the mortgage term as control variable
similar to Coulibaly and Li (2009) and Zocchi (2013). The results in the above median income segment
indicate that people who borrow on longer terms are more likely to choose adjustable rate mortgages.
Damen and Buyst (2014) show that an increase in the mortgage term of one year, increases the mortgage
rate by 4.5 basis points in Flanders, ceteris paribus. Households who borrow on longer mortgage terms
are thus more likely to choose an adjustable rate mortgage to avoid the relatively higher term premium
of FRMs. The mortgage term is, however, not signi…cant in the below median income segment.
Uninformed borrowers
An issue that arises is the possibility of the miss-selling of mortgages by …nancial intermediaries. While
borrowers make the complex mortgage choice only a few times in their lives, …nancial intermediaries
have an informational advantage in the …nancial markets10 . The mortgage provider might not inform
the client that a higher FRM-ARM spread may not be the result of a cheaper ARM, but instead due to
higher expected future ARM rates. If the bank does not want to take this interest rate risk, they can
easily pass it on to the uninformed borrower.
Responses to questions in the Flemish Housing Survey in 2005 do indeed suggest that borrowers
are likely to be uninformed: only 40% of the borrowers in our sample were aware that the Flemish
government o¤ers a free insurance that provides an allowance when the borrower loses their income due
to involuntary unemployment. The households in the lowest 25 percent of income per working household
member are even less informed: only 26% knows that the free insurance exists.
In the analysis, we assume that borrowers knew their mortgage terms and did not misreport them
to the Flemish Housing Survey. If they misreported whether or not the interest rate could change over
time, this could result in a high amount of misclassi…cation that is not necessarily due to inappropriate
choices. Bucks and Pence (2008), however, provide evidence that most borrowers (in the USA) know
their basic mortgage terms (ARM vs. FRM, maturity). The main problem according to them is that
ARM borrowers do not know other characteristics of their contract, such as the degree that their interest
rate can change. The participants of the Flemish Housing Survey were furthermore questioned by trained
1 0 Bolton,
Freixas and Shapiro (2007) studied similar situations of con‡icts of interest with information provision.
18
professionals who could always use the "don’t know" option in cases where the households did not know
their contract type. We therefore expect that misclassi…cation does not in‡uence the estimation.
Financial crisis
To explore the role of the …nancial crisis on the choice of mortgage contract we run two additional
estimations where we include an interaction with a time dummy, and the loan-to-value ratio and mortgage
term. In the …rst estimation, the time dummy equals one if the respondent purchased a house in 2009
or thereafter. This dummy variable is thus equal to one for all observations from the 2013 survey. The
variable equals zero for all purchases from 2001 until 2005, and therefore allows a robustness check for
mortgage choices pre- and post-crisis. In a second estimation we create a time dummy that equals one
if the respondent purchased a house after 2012. Data from the Bank Lending Survey indeed indicates
that banks were restricting mortgage admission criteria since 2012 based on loan-to-value ratios and
mortgage terms. For this reason, we interact the time dummies (in two separate regressions) with the
loan-to-value ratio and mortgage term. Both interaction terms are not signi…cant (results not reported
here), such that we cannot reject that the recent …nancial crisis had changed the association between
these variables and the mortgage choice.
Figure 4: House prices and apartment prices in Flanders (in euro)
240.000
220.000
200.000
180.000
160.000
140.000
2005Q1
2005Q3
2006Q1
2006Q3
2007Q1
2007Q3
2008Q1
2008Q3
2009Q1
2009Q3
2010Q1
2010Q3
2011Q1
2011Q3
2012Q1
2012Q3
2013Q1
2013Q3
2014Q1
120.000
Houses
Apartments
Note: The solid line presents the average nom inal sale price of houses on the secondary m arket in Flanders (in euro). The dotted
line presents the average sale price of apartm ents, ‡ats and studios on the secondary m arket in Flanders (in euro). Source:
Statb el
Although Belgian banks were largely a¤ected by the Global Financial Crisis, the impact on the
Belgian and Flemish housing market was indeed relatively limited (see Figure 4). Belgium did not
experience a housing market correction due to an increase in the …scal advantage for homeowners in
2005 and decreasing long-term interest rates since 2008 (Damen and Buyst 2014).
19
Conclusion
Recent contributions to the theoretical literature by Koijen et al. (2009) argue that variation in the bond
term premium should be the key theoretical determinant of variation in the ARM share over time. The
term premium re‡ects the extra yield risk averse lenders demand for holding the riskier long term bonds.
In order to examine whether or not households make close-to-optimal decisions, we thus di¤erentiate
between this risk premium and the expected change in ARM rates. The results indicate that the risk
premium explains almost no variation in the mortgage structure.
As the risk premium does not appear to be the main driver in the rise and fall of the ARM, we
elaborate on the …nancial decision households actually make. The results indicate that the FRM-ARM
spread explains almost 80 percent of the variation in the ARM share from 2003:01 to 2012:12. The
economic importance of expectations of future interest rates is small. Belgian households thus base their
decision on the level of initial mortgage payments, and give only limited importance to future increases
in interest rates.
Furthermore, we study the determinants of households that would lead them to prefer the adjustable
over the …xed rate mortgage in the Flemish region of Belgium. We use data from the Flemish Housing
Survey (2005 and 2013), as it contains detailed micro-data on the household level. Remarkably, we …nd
evidence that low income and high income households choose the ARM for di¤erent reasons. Low income
households choose the ARM if they have relative high loan amounts, against theoretical expectations
by Campbell and Cocco (2003). We do not …nd evidence that these households are constrained in
their mortgage choice, as 98% of ARM borrowers report they were easily granted credit. Therefore, we
attribute the empirical …ndings of the low income segment to suboptimal mortgage choices. High income
households choose the ARM if they have higher education, are younger or can a¤ord the risk of the ARM
due to cheaper houses relative to income.
An important …nding is that households who expect to move in the near future are more likely to
choose the FRM. As all mortgages are portable by law, borrowers are able to lock-in a favorable FRM
rate and do not need to prepay if they decide to move. Expected movers thus have a preference to reduce
variability in mortgage payments, which was not observed in the previous literature.
An issue that arises is the possibility of miss-selling mortgages by …nancial intermediaries to low
income households. As households make the complex mortgage choice only a few times in their lives,
they are likely not well informed. Financial intermediaries, on the other hand, have an informational
advantage in the …nancial markets and can easily pass the risk of higher future interest rates to the
uninformed myopic borrower who only looks at the initial FRM-ARM spread.
20
Acknowledgements
The authors would like to thank all participants of the ENHR housing economics work group in Tarragona (June, 2013), Frank Vastmans and Roel Helgers for their valuable suggestions. Funding from the
University of Leuven is gratefully acknowledged.
References
[1] Alm, James and James R. Follain. (1987) "Consumer demand for adjustable rate mortgages." Housing Finance Review 6(1): 1-16.
[2] Ang, Andrew and Monika Piazzesi (2003). "A no-arbitrage vector autoregression of term structure
dynamics with macroeconomic and latent variables." Journal of Monetary Economics 50(4):
745-787.
[3] Bacon, Philomena M. and Peter G. Mo¤att. (2012). "Mortgage Choice as a Natural Field Experiment
on Choice under Risk." Journal of Money, Credit and Banking 44(7), 1401-1424.
[4] Bergstresser, Daniel and John Beshears. (2010). "Who Selected Adjustable-Rate Mortgages? Evidence from the 1989-2007 Surveys of consumer Finances." Harvard Business School Working
Paper, 10-083.
[5] Berkovec, James A., David J. Kogut and Frank E. Nothaft. (2001). "Determinants of the ARM
Share of FHA and Conventional Lending," Journal of Real Estate Finance and Economics 22(1),
23-41.
[6] Bolton, Patrick, Xavier Freixas and Joel Shapiro. (2007). "Con‡icts of interest, information provision, and competition in the …nancial services industry." Journal of Financial Economics 85,
297-330.
[7] Brueckner, Jan K. (1986). "The Pricing of Interest Rate Caps and Consumer Choice in the Market
for Adjustable-Rate Mortgages" Housing Finance Review 5: 119-136.
[8] Brueckner, Jan K. and James R. Follain. (1988). "The Rise and Fall of the Arm: An Econometric
Analysis of Mortgage Choice," The Review of Economics and Statistics 70(1), 93-102.
[9] Bucks, Brian and Karen Pence. (2008). "Do borrowers know their mortgage terms?" Journal of
Urban Economics 64, 218-233.
[10] Campbell, John Y. (2013). "Mortgage Market Design." Review of Finance 17(1), 1-33.
[11] Campbell, John Y. and Joao F. Cocco. (2003). "Household Risk Management and Optimal Mortgage
Choice." The Quarterly Journal of Economics 118(4), 1449-1494.
21
[12] Campbell, John Y. and Robert J. Shiller. (1991). "Yield spreads and interest rate movements: A
bird’s eye view." The Review of Economic Studies 58(3): 495-514.
[13] Cochrane, John and Monika Piazzesi. (2005). "Bond risk premia." American Economic Review
95(1): 138-160.
[14] Coulibaly, Brahima and Geng Li. (2009). "Choice of Mortgage Contracts: Evidence from the Survey
of Consumer Finances." Real Estate Economics 37(4), 659-673.
[15] Damen, Sven and Erik Buyst (2014). "The long-run relationship between house prices and income
reexamined: The role of mortgage interest deduction and mortgage product innovation." CES
Discussion Paper 14.09.
[16] Damen, Sven and Erik Buyst. (2014). "Mortgage shoppers: how much do they save?" paper presented at the AREUEA international conference, Reading, 9-12 July 2014.
[17] Dhillon, Upinder S., James D. Shilling and C. F. Sirmans. (1987). "Choosing Between Fixed and
Adjustable Rate Mortgages: Note," Journal of Money, Credit and Banking 19(2), 260-267.
[18] ECB. (2009). "Housing Finance in the Euro Area," ECB Occasional Paper Series, 101.
[19] European Mortgage Federation. (2006). "Study on Interest Rate Variability in Europe." European
Mortgage Federation Publication.
[20] European Mortgage Federation. (2012). "Study on Mortgage Interest Rates in the EU." European
Mortgage Federation Publication.
[21] Fama, Eugene F. and Kenneth R. French. (1989). "Business conditions and expected returns on
stocks and bonds." Journal of Financial Economics 25(1): 23-49.
[22] Hjalmarsson, Erik and Pär Österholm. (2007). "Testing for Cointegration Using the Johansen
Methodology when Variables are Near-Integrated." IMF Working Paper, WP/07/141.
[23] HM Treasury. (2003a). "Housing, Consumption and EMU."
[24] HM Treasury. (2003b). "UK Membership of the Single Currency. An Assessment of the Five Economic Tests. "
[25] i¤/ZEW. (2010). "Study on interest rate restrictions in the EU, Final Report for the EU
Commission DG Internal Market and Services," Project No. ETD/2009/IM/H3/87, Brussels/Hamburg/Mannheim.
[26] International Monetary Fund. (2004). "World Economic Outlook," Chapter 2, September.
[27] Johnson, Kathleen and Geng Li. (2014). "Are Adjustable-Rate Mortgage Borrowers Borrowing Constrained?" Real Estate Economics, 42(2), 457-471.
[28] Koijen, Ralph S. J., Otto Van Hemert and Stijn Van Nieuwerburgh. (2009). "Mortgage Timing,"
Journal of Financial Economics 93, 292-324.
22
[29] Leece, David. (2000). "Household choice of …xed versus ‡oating rate debt: a binomial probit model
with correction for classi…cation error," Oxford Bulletin of Economics and Statistics 62(1), 61-82.
[30] Maclennan, Duncan, John Muellbauer and Mark Stephens. (1998). "Asymmetries in housing and
…nancial market institutions and EMU," Oxford Review of Economic Policy 14(3), 54-80.
[31] Miles, David. (2004). "The UK Mortgage Market: Taking a Longer-Term View. Final Report and
Recommendations" London: UK Treasury.
[32] NBB. (2012). "Financial Stability Review 2012," Brussels: National Bank of Belgium.
[33] Nothaft, Frank E. and George H.K. Wang. (1992). "Determinants of the ARM Share of National
and Regional Lending." Journal of Real Estate Finance and Economics 5: 219-234.
[34] Paiella, Monica and Alberto F. Pozzolo. (2007). "Choosing between …xed and adjustable rate mortgages." Unpublished.
[35] Rubio, Margarita. (2011). "Fixed- and Variable-Rate Mortgages, Business Cycles, and Monetary
Policy," Journal of Money, Credit and Banking 43(4), 657-688.
[36] Sa-Aadu, Jarjisu and C. F. Sirmans. (1995). "Di¤erentiated contracts, heterogeneous borrowers,
and the mortgage choice decision." Journal of Money, Credit and Banking 27(2): 498-510.
[37] Schwartz, Allie. (2009). "Who takes out adjustable rate mortgages?" Unpublished.
[38] Tucker, Michael. (1989). "Adjustable-rate and …xed-rate mortgage choice: A logit analysis." Journal
of Real Estate Research 4(2): 81-91.
[39] Woodward, Susan E. and Robert E. Hall. (2012). "Diagnosing Consumer Confusion and Sub-Optimal
Shopping E¤ort: Theory and Mortgage-Market Evidence," The American Economic Review
102(7), 3249-3276.
[40] Zocchi, Paola. (2013). "Why do Italian households prefer adjustable rate mortgages?" Journal of
European Real Estate Research 6(1): 90-110.
23
Appendix
Robustness checks for risk premia
See Table 7.
Roots of the companion matrix
1
Imaginary
.5
0
-.5
-1
-1
-.5
0
.5
1
Real
The VECM specification imposes 2 unit moduli
Figure 5: Roots of the companion matrix
ARM as option for the FRM
Table 8 presents regression results with the volume of re…nances as dependent variable and
F RM ARM
(n)
t
as explanatory variable. The results teach us that the amount of re…nances does not increase when
F RM ARM
(n)
t
decreases. Borrowers thus do not seem to temporarily choose the ARM when risk pre-
mia are high in order to repay the entire loan and choose an FRM when risk premia have reverted.
24
25
- or n.s.
Conclusions
+
+
+
+
+
+
n.s.
+
+
Yield
-
-
-
-
-
-
-
Curve
+
+
n.s.
+
+
rate
FRM
n.s.
n.s.
n.s.
Children
m ixed
n.s.
+
+
-
+
-
n.s.
n.s.
Age
n.s. or +
+
n.s.
n.s.
n.s.
+ (but m o dest)
Incom e
Higher
n.s. or +
- (but m odest)
+
n.s.
n.s. or +
n.s.
education
-
-
-
index
A¤ordability
Borrower and mortgage characteristics
+
+
spread
+
FRM -ARM
Risk
prem ium
n.s.
- (but m odest)
n.s.
n.s.
n.s.
em ploym ent
Self
+
+
+
M aturity
n.s.
n.s.
n.s.
to incom e
Repaym ent
ARM rate - base rate (+); base rate (+ )
ARM rate (+)
ARM rate - base rate (+)
ARM rate (-)
base rate (-); CPI (+)
Other
Loan
m ixed
+
-
n.s.
am ount
Note: A plus or minus sign indicate respectively a positive or negative e¤ect. "n.s." indicates that the variable is not signi…cant in the estimation.
- (but m odest)
-
n.s.
- or n.s.
n.s.
+
Zocchi (2013)
Bacon and M o¤att (2012)
+
+
Coulibaly and Li (2009)
Koijen, et al. (2009)
n.s.
Paiella and Pozzolo (2007)
Berkovec, et al. (2001)
Leece (2000)
Sa-Aadu and Sirm ans (1995)
Nothaft and Wang (1992)
+
+
Brueckner and Follain (1988)
Tucker (1989)
+
Dhillon, et al. (1987)
# b orrowers
or # workers or m arried
High
m obility
Italy
Zocchi (2013)
Conclusions
UK
USA (and UK)
Bacon and M o¤att (2012)
Koijen, Van Hem ert and Van Nieuwerburgh (2009)
USA
Leece (2000)
Coulibaly and Li (2009)
UK
Sa-Aadu and Sirm ans (1995)
USA
USA
Nothaft and Wang (1992)
Italy
USA
Tucker (1989)
Paiella and Pozzolo (2007)
USA
Brueckner and Follain (1988)
Berkovec, Kogut and Nothaft (2001)
USA
USA
Dhillon, Shilling and Sirm ans (1987)
Country
Price and macro variables
Table 1: Overview of the empirical literature
+
+
+
to incom e
Loan am ount
Table 2: Summary statistics
Variables
FRM-ARM spread
Moving plans (D)
Higher education (D)
Partner (D)
Age
Children in the house (D)
A¤ordability index
LTV
Full sample
Below median income
Above median income
0.69
0.11
0.52
0.72
32.56
0.47
0.96
0.85
0.69
0.11
0.35
0.78
32.02
0.43
0.93
0.87
0.70
0.12
0.67
0.67
33.06
0.50
0.99
0.82
Note: The table presents descriptive statistics for all variables. Column 2 presents the mean for the full sample.
Columns 3 and 4 present, respectively, mean values for the subsample of households with below and above median
income per parent.
26
Table 3: Comparison of models A and B
Estimation sample
Expanding / …xed window?
Lags included
Model A
1960:01- 2002:12
Expanding to 2012:12
24
27
Model B
1999:01-2012:12
Fixed
12
Table 4: The ARM share and the risk premium
Variables
Constant
FRM-ARM Risk premium (
Risk premium (
t (n)
F RM
t
ARM
(n))
)
Model A
-0.02
0.20 ***
(0.04)
(0.04)
8.51 **
(3.33)
15.27 *
(9.21)
0.17
0.12
Model B
0.12
0.03
(0.09)
(0.04)
14.33 ***
(2.92)
5.76
(3.65)
0.14
0.39
R2
State vector (Xt )
Yt (1) Yt (10) Yt (1)
t
VAR / VECM
VECM
Lags
24
12
Estimation sample of risk premia
1960:1 - 2012:12 (E) 1999:1-2012:12 (F)
Note: This table represents the estimation results from 2003:01 to 2012:12 of the ARM share regressed on the
risk premium (for n = 10 years) estimated from two models (A and B). Model A is estimated with a sample from
1960:01 to 2002:12 that expands to 2012:12. Model B uses a …xed estimation sample from 1999:01 to 2012:12.
(E) and (F) indicate respectively whether an expanding or …xed sample was used for the estimation. The share
of ARMs is de…ned as the mortgages with an initial …xed rate period smaller than 10 years as a percentage of all
new mortgages. The state vector of the VECM includes the short term government bond, 10-year government
bond spread and in‡ation. The VECM is then able to dynamically forecast the short rate, which is used to
calculate the risk premium as a residual value from (4) and (5). Newey-West standard errors are reported in
parentheses (12 lags). *, **, *** refer respectively to p-values lower than 0.10, 0.05 and 0.01
28
Table 5: ARM share predicted by FRM-ARM spread, FRM rate and government bond spread
Variables
Constant
FRM-ARM spread
(1)
0.13 ***
(0.02)
26.06 ***
(3.27)
Yield spread
Pn
ARM
Et [Y ARM
(1)
t+j 1 (1)] Y t
(2)
0.15 ***
(0.03)
28.38 ***
(3.98)
-2.24
(1.65)
(3)
0.11 ***
(0.02)
27.77 ***
(3.79)
(4)
0.08 ***
(0.02)
26.84 ***
(3.04)
-2.53
-3.28 *
(1.82)
(1.79)
R2
0.80
0.81
0.81
0.81
Risk premium model?
A
B
Note: This table presents the estimation results from 2003:01 to 2012:12 of the ARM share regressed on the
FRM-ARM interest spread in combination with interest expectations from the yield spread or ARM forecasts
estimated from two models (A and B). Model A is estimated with a sample from 1960:01 to 2002:12 that expands
to 2012:12. Model B uses a …xed estimation sample from 1999:01 to 2012:12. The state vector of the VECM
includes the short term government bond, 10-year government bond spread and in‡ation. The VECM is then
able to dynamically forecast the short rate. The share of ARMs is de…ned as the mortgages with an initial …xed
rate period smaller than 10 years as a percentage of all new mortgages. Newey-West standard errors are reported
in parentheses (12 lags). *, **, *** refer respectively to p-values lower than 0.10, 0.05 and 0.01
1
n
j=1
29
Table 6: Logit estimation results
Variables
(1)
(2)
(3)
Full sample Below median income Above median income
FRM-ARM spread
0.73***
0.42*
1.21***
(0.17)
(0.25)
(0.27)
Moving plans (D)
-0.83**
-0.94*
-0.95*
(0.34)
(0.53)
(0.50)
Higher education (D)
0.30
-0.05
0.79**
(0.19)
(0.29)
(0.33)
Partner (D)
-0.32
-0.05
-0.77**
(0.23)
(0.38)
(0.34)
Age
-0.02
-0.01
-0.03*
(0.01)
(0.02)
(0.02)
Children in the house (D)
-0.07
-0.19
0.20
(0.19)
(0.27)
(0.29)
A¤ordability index
0.23
-0.43
1.33***
(0.32)
(0.45)
(0.49)
LTV
1.07***
1.69***
0.36
(0.41)
(0.63)
(0.55)
Moving plans (D) Unemployed (D)
1.17
0.63
(omitted)
(0.95)
(1.11)
A¤ordability index Unemployed (D)
-0.00
-0.05
-0.19
(0.30)
(0.37)
(0.72)
Constant
-1.07
-1.12
-1.70
(0.68)
(0.90)
(1.14)
Observations
700
350
349
Note: This table presents logit estimation results where the dependent variable equals 1 if the household has an
ARM. The a¤ordability index is de…ned in the data section. Standard errors are reported in parentheses. *, **,
*** refer respectively to p-values lower than 0.10, 0.05 and 0.01
30
Table 7: The ARM share and the risk premium: robustness checks
Variables
Constant
C.1
0.26 ***
(0.06)
FRM-ARM risk premium (
Risk premium (
t (n)
)
F RM
t
ARM
(n))
16.68*
(10.87)
0.11
C.2
0.38 ***
(0.10)
11.22 **
(5.09)
D.1
0.10
(0.13)
6.52
(5.46)
0.10
D.2
-0.10
(0.07)
22.40 ***
(4.16)
R2
0.10
0.42
State vector (Xt )
Yt (1) Yt (10) Yt (1)
Yt (1) Yt (10)
t
t
VAR / VECM
VAR
VECM
Lags
Xt 12 ;X t 24
6
Estimation sample of risk premia
1960:01-2012:12 (E)
1999:01-2012:12 (F)
Note: This table represents the estimation results from 2003:01 to 2012:12 of the ARM share regressed on the
risk premium (for n = 10 years) estimated from two models (C and D). Model C is estimated with a sample from
1960:01 to 2002:12 that expands to 2012:12. Model D uses a …xed estimation sample from 1999:01 to 2012:12.
(E) and (F) indicate respectively whether an expanding or …xed sample was used for the estimation. The share
of ARMs is de…ned as the mortgages with an initial …xed rate period smaller than 10 years as percentage of all
new mortgages. The state vector of the VAR in model C includes the short term rate, 10-year government bond
rate and in‡ation. The state vector of the VECM in model D includes the short term government bond, 10-year
government bond spread and in‡ation. The models are then able to dynamically forecast the short rate, which is
used to calculate the risk premium as a residual value. Newey-West standard errors are reported in parentheses
(12 lags). *, **, *** refer respectively to p-values lower than 0.10, 0.05 and 0.01
31
Table 8: Re…nance volume explained by risk premium and FRM rate
Variables
Constant
(1a)
254.79
(50.61)
-785.49
(3107.66)
(1b)
(2a)
(2b)
964.76 ***
174.41 ***
876.66 ***
(148.62)
(42.51)
(130.74)
FRM-ARM risk premium ( Ft RM ARM (n))
1591.94
2925.60 **
1633.92
(2676.80)
(1372.58)
(1235.32)
FRM rate (YtF RM (n))
-15252.96 ***
-13780.74 ***
(2676.80)
(2684)
2005-2012 dummy
-67.05
-144.29 ***
-40.53
-132.35 ***
(46.06)
(33.02)
(42.53)
(35.40)
R2
0.07
0.45
0.13
0.46
Risk premium model?
A
B
Note: This table represents the estimation results from 2003:01 to 2012:12 of the re…nance volumes as dependent
variable on the FRM premium that lenders demand ( tF RM ARM (n)) and the FRM rate. We set n equal to 10
years. We also include a dummy for the period since 2005 as the re…nance volumes only include external re…nances
since 2005. Newey-West standard errors are reported in parentheses (12 lags). *, **, *** refer respectively to
p-values lower than 0.10, 0.05 and 0.01
32