COMPARING HOUSING BOOMS AND MORTGAGE SUPPLY IN THE MAJOR
OECD COUNTRIES
Angus Armstrong and E Philip Davis
1NIESR and Brunel University
London
Abstract: The house price and lending boom of the 2000s is widely considered to be the main cause of the financial crisis that began in 2007. However, looking to the past, we find a similar boom in the late 1980s which did not lead directly to a global systemic banking crisis – there were widespread banking difficulties in the early 1990s but these were linked mainly to commercial property exposures. This raises the question whether the received wisdom is incorrect, and other factors than the housing boom caused the crisis, while macroprudential policy is overly targeted at the control of house prices and lending per se. Accordingly, in this paper we compare and contrast the cycles in house prices over 1985-94 with 2002-11. There are more similarities than contrasts between the booms. Stylised facts include a similar rise in real house prices where booms took place, and a marked rise in the real mortgage stock along with real incomes. The aftermath periods are also comparable in terms of house price
changes. Econometrically, determinants of house prices are similar in size and sign from the 1980s to date. There remain some contrasts. Leverage rose far more in the later episode and did not contract in the aftermath. Serial correlation of house prices, suggestive of
extrapolative expectations, is greater in the recent period. The earlier boom period showed differences with average house price behaviour which was not mirrored in the most recent boom and inflation was higher. Despite the contrasts, on balance we reject the idea that the recent boom was in some way unique and hence the key cause of the crisis. There is a need for further research to capture structural and conjunctural factors underlying the recent crisis which differ from the earlier boom and some suggestions are made.
Keywords, House price booms, mortgage stock, housing markets
JEL classification: C52, E58, G21
1 Emails, a.armstrong@niesr.ac.uk (Armstrong) and e_philip_davis@msn.com (Davis). We thank John
Muellbauer, anonymous referees and participants at the NIESR/ESRC “Future of housing finance” conference at
the Ban k of England for helpful comments. Errors remain our own responsibility.
Introduction
The house price and lending boom of the 2000s is widely considered to be not only a unique event but also the main cause of the global financial crisis that began in 2007, leading in turn to the biggest losses in financial wealth for generations (IMF (2008a), Kemme and Roy (2012)). Typical of current thinking is a speech earlier this year by Min Zhu, Deputy Managing Director of the IMF who said “housing is an essential sector of the economy but also one that has been the source of vulnerabilities and crises” (my italics). However, looking to the past, we find a similar global housing boom in the late 1980s which did not lead
directly to a global systemic banking crisis – there were widespread banking difficulties in the early 1990s but these were linked mainly to commercial property exposures (Davis 1995).
This raises the question whether the received wisdom is incorrect, and other factors than the housing boom caused the crisis, while macroprudential policy is wrongly targeted at the control of house prices and lending per se.
Accordingly, in this paper we compare the cycles and assess the evolution in house price determination in major OECD countries over the past decades to see whether the current cycle is unique. A key point in this context is that housing differs from other markets in that
informational reasons, transaction costs, credit rationing and supply side factors help explain serial correlation and mean reversion in house prices which may in turn differ across countries and time but may also lead to common patterns in global markets (Capozza et al 2002).
In terms of a comparison, we may ask whether the booms were similar in key features apart from rising house prices, or were there major contrasts? We explore these questions via a statistical comparison of roughly-defined booms periods as well as the “aftermath” of the booms.
2We go on to assess whether there have been changes in the relationship of house prices to their determinants more generally in the two main housing cycles since liberalisation which in most OECD countries happened in the 1980s.
3Furthermore, it is a stylised fact that mortgage debt should not have a direct influence on house prices in a liberalised financial market such as characterised both the recent boom periods (since mortgage debt is then demand-determined). We examine econometrically whether this was the case for the booms in question. Finally we consider other unique factors that may distinguish the recent boom better than house price and lending dynamics per se.
The paper is structured as follows. In the first section, we compare housing booms and assess in particular the changes in real house prices and their main determinants, notably real
personal disposable income (RPDI) and real housing debt in 15 major OECD countries. In the second section we briefly introduce work underlying house price equations before providing a specification for house price determination (similar to Davis, Fic and Karim (2011)) in the third section and results in the fourth. In the fifth section we look specifically at results for the impact of credit supply on house prices, which is omitted by most extant specifications and in the sixth we look at potential structural and conjunctural factors that may distinguish the booms. The final section concludes.
1 Comparing global housing cycles
2 We prefer this word since house prices rarely “crash” in the way that financial asset prices do, not least owing to the dual use of houses for consumption of housing services as well as for investment.
3 For example in the US, portfolio restrictions on banks and non-banks, prohibitions on adjustable rate
mortgages, tax inducements to non-banks and deposit rate ceilings were all abolished in the early 1980s
(Hendershott 1994). In following years, securitisation began to be prominent as a source of mortgage finance
albeit not attaining the importance it did in the 2000s.
We have quarterly data on house prices and other relevant macroeconomic and financial variables covering both boom periods for 15 OECD countries, drawn from the BIS database.
We define the booms roughly as five year periods from 1985q1-1989q4 and 2002q1-2006q4, in line with Dokko et al (2011) of the Fed and incorporating the periods that Igan and
Loungini (2012) of the IMF show for country-by country specific data on house price cycles.
4We also define an “aftermath” period for each boom which is the following five years, namely 1990q1-1994q4 and 2007q1 -2011q4. It is in these periods that output typically remained subdued and banking crises took place in certain countries,
5and falls in house prices tended to occur with tight credit markets.
Our analysis of the booms and aftermath begins with Table 1 below which shows the relevant changes in real house prices over the periods together with real personal disposable income (a key determinant of house prices), the stock of real household sector debt
6, nominal house prices and real gross financial wealth.
Table 1: Changes in house prices, income, debt and wealth during booms Percentage
change
Real house prices
RPDI Real household
debt
Nominal house prices
Real gross financial wealth 1985q1
-89q4
2002q1 -06q4
1985q1 -89q4
2002q1 -06q4
1985q1 -89q4
2002q1 -06q4
198 5q1- 89q 4
2002 q1- 06q4
198 5q1- 89q 4
2002 q1- 06q4
United Kingdom 71 49 23 10 74 50 112 65 61 17
United States 12 29 17 14 40 48 31 44 31 33
Germany 1 -2 18 5 18 -3 6 4 37 9
France 28 64 14 11 51 42 49 78 65 26
Canada 32 25 17 19 53 44 60 35 27 17
Italy 32 20 17 3 88 40 76 36 50 10
Spain 110 62 27 17 23 83 190 90 95 41
Austria 68 -5 21 13 16 26 81 4 35 27
Netherlands 24 11 16 -2 16 42 25 21 46 19
Belgium 32 41 17 3 21 29 46 56 56 1
Denmark -8 56 5 10 21 44 8 67 22 58
Ireland 12 48 16 18 38 145 33 69 76 48
Finland 56 32 24 17 78 83 91 35 57 42
Sweden 35 44 10 12 35 45 78 52 94 52
Japan 27 -17 22 4 59 0 33 -20 80 16
Mean 35 30 18 10 42 48 61 42 55 28
Mean (boom
countries) 40 39 18 11 47 59 72 53 60 28
Correlation 0.74 0.41 0.14 0.58 0.95 0.99 0.47 0.49
Correlation
(boom countries) 0.79 0.42 0.06 0.30 0.97 0.97 0.49 0.35
4 IMF (2008b) date the end of the 2000s cycle in line with us, suggesting a corresponding overvaluation in the
“boom countries” at the end of the upturn of over 10%, with the exceptions being Finland and Canada.
5 Barrell et al (2010) show that the three year lagged difference of house prices is an important predictor of banking crises in OECD countries.
6 We do not have mortgage debt for all countries so use this variable for comparability purposes – and because
it shows the overall vulnerability of the household sector.
Notes: Source:BIS and OECD. Real house prices and real household sector liabilities are deflated by the consumers’ expenditure deflator. Calculations for the “boom countries” excludes Germany, Austria, Denmark and Japan; it includes only UK, US, France, Canada, Italy, Spain, the Netherlands, Belgium, Ireland, Finland and Sweden.
The table shows, first, that not all countries participated in both the first and the second global house price boom. Using a rough benchmark of 10% rise in real house prices to qualify a boom, Japan and Austria only experienced significant rises in house prices in the earlier period, while Denmark saw large rises only in the later period. Germany did not experience sizeable rises in real house prices in either period. The countries that saw rises of 10% or more in both booms are the UK, US, France, Canada, Italy, Spain, the Netherlands, Belgium, Ireland, Finland and Sweden. The average rise in house prices across all 15 countries was somewhat lower in the latest boom than in the earlier one but when calculated only for the boom countries mentioned above, it is almost identical at around a 40% rise in real house prices. So in this fundamental aspect the boom periods are similar. As regards the dispersion of real house price changes, it was lower in the more recent boom suggesting a role for international contagion (the standard deviation of price rises in the boom countries was 17%
in the later boom and 29% in the earlier boom). Agnello and Schuknecht (2011) suggest that global liquidity could have played an important role in occurrence of simultaneous housing booms in the 2000s.
Real personal disposable income was considerably more buoyant in the earlier boom period than in the 2002-6 period. On average incomes rose 18% in the 1980s and only 10-11% in the 2000s. On the other hand, the rise in household debt was higher in the later period, especially for those countries that experienced booms in both periods, where the rise in the later period was 59% compared to 47% in the earlier boom. We decided in the light of this to calculate correlation coefficients for overall changes in each variable with real house prices in the different boom periods. There are marked differences in that the correlation of RPDI with real house prices was much higher in the earlier period, especially when one calculates across the countries experiencing two distinct booms. On the other hand, Table 1 shows that the
correlation with household debt was markedly higher in the later period. This gives a starting indication of differences between the booms that are worthy of further investigation.
Nominal house prices rose more in the earlier boom, corresponding to higher inflation in the 1980s. This in turn had an impact on real mortgage debt, with a greater reduction in value of nominal debt in the earlier period. Real financial wealth grew much more in the earlier period despite the stock market crash of 1987, rising at rates in excess of real house prices whereas in the later boom real house prices rose more than wealth. Of course the series are not directly comparable as real gross financial wealth rises due to accumulation as well as asset price rises.
Table 2: Changes in house prices, income, debt and wealth during the aftermath of booms
Percentage
change Real house prices RPDI Real household
debt Nominal
house prices
Real gross financial wealth 1990q1
-94q4 2007q1
-11q4 1990q1
-94q4 2007q1
-11q4 1990q1
-94q4 2007q1 -11q4
199 0q1- 94q 4
2007 q1- 11q4
199 0q1- 94q 4
2007 q1- 11q4
United Kingdom -21 -14 12 3 10 -8 -5 0 21 -6
United States -3 -24 12 6 19 -9 11 -17 16 -4
Germany -2 na 11 5 25 -7 16 9 29 2
France -8 -1 7 3 -4 22 1 7 17 3
Canada -18 2 -1 11 13 36 -8 9 18 16
Italy 12 -6 -2 -6 32 10 45 3 15 -17
Spain -7 -23 10 -2 8 -1 22 -16 19 -14
Austria -2 4 12 0 14 5 13 15 21 5
Netherlands 21 -9 8 0 22 18 38 -5 14 9
Belgium 14 7 14 2 9 23 28 18 -3 -2
Denmark 0 -26 8 2 -19 12 9 -18 -1 -4
Ireland 0 na 14 -4 15 8 14 -48 18 6
Finland -42 0 -13 8 -21 20 -32 13 -16 -4
Sweden -26 7 11 8 -18 28 -7 17 -22 5
Japan -9 -8 9 0 19 -2 -2 -13 12 0
Mean -6 -7 7 2 8 10 9 -2 10 0
Mean (boom
countries) -7 -6 6 3 8 13 10 -2 9 -1
Correlation 0.46 0.29 0.62 0.67 0.93 0.97 0.41 0.22
Correlation
(boom countries) 0.46 0.38 0.78 0.88 0.86 0.82 0.19 0.61
Notes: See Table 1
Table 2 shows comparable data and calculations for the post-boom “aftermath” period for each boom. The average change in real house prices was comparable in the earlier
“aftermath” from 1990-1994 with the more recent period covering 2007-11, both being around -6 to -7%, despite the differing levels of general inflation. This masks considerable cross country variation, however, with for example the UK, Sweden and Finland, that experienced banking crises in the early 1990s, showing larger falls in the earlier period, and the US and Spain among others in the later period On average, changes in personal income were larger in the earlier period, at around 7% compared to 2%. On the other hand, real mortgage debt rose more in the aftermath of the 2002-6 boom, at 10% or more compared to 8%. Again, this was not true of all countries, with the UK and US both showing falls in real household debt over the more recent period, as households sought to delever. The correlation of RPDI changes with real house prices is again lower in the later period while that of
household debt with house prices is higher, and is very high for the boom countries (0.88).
Meanwhile, nominal house prices rose in the aftermath of the earlier boom (reflecting general inflation) while they fell in the later one. Similarly to income, real gross financial wealth rose in 1990-4 while it was flat in 2007-11, reflecting the global financial crisis, Canada being the main exception.
7Table 3: Indicators of leverage in booms and aftermath
Debt/personal income
ratio – Debt/house prices –
percentage change Debt/personal
income ratio – Debt/house prices –
7 We focus on the first moment in our presentation. We may add that housing markets are typically
characterised by less volatility than equity, bond or foreign exchange markets, but liberalised credit markets do
give scope for housing to be treated as an asset rather than only a source of housing services. Given the greater
likely weight of such investment demand in a boom we could expect house prices to be more volatile in such
periods. We calculate (not shown in detail) that house price volatility was higher in the earlier boom than the
later one. Also, in the 1985-1994 decade, house price volatility up to 1989 was considerably higher than in 1990-
4, on average, whereas in the 2002-11 period there was a rise in volatility after the onset of the banking crisis, a
pattern which was particularly apparent in the boom countries.
change in percentage
points change in
percentage points
percentage change 1985q1-
89q4 2002q1-
06q4 1985q1-
89q4 2002q1-
06q4 1990q
1- 94q4
2007q 1- 11q4
1990q 1- 94q4
2007q 1- 11q4
United Kingdom 25 30 2 1 -1 -10 39 7
United States 3 6 25 15 1 -3 22 20
Germany -1 -5 17 -1 10 -8 28 -9
France 9 12 18 -13 -4 10 5 23
Canada 14 16 16 15 7 24 39 34
Italy 8 14 42 17 7 9 17 17
Spain -3 35 -42 13 -2 2 17 29
Austria -1 8 -31 34 0 2 16 1
Netherlands 1 43 -6 28 7 24 1 30
Belgium 3 11 -9 -8 -2 10 -5 15
Denmark 8 44 32 -8 -33 16 -19 52
Ireland 13 84 23 65 -2 22 15 100
Finland 17 26 14 39 -9 10 35 20
Sweden 9 21 0 1 -12 24 12 20
Japan 21 -5 25 21 9 -5 31 7
Mean 8 23 9 15 -2 8 17 24
mean (boom
countries) 9 27 8 16 -1 11 18 28
Notes: See Table 1
As a factor possibly underlying these patterns, as well as being of wider relevance to macroprudential policy, we examine the behaviour of two indicators of financial fragility, namely the household debt/personal income ratio (which is of course mainly housing debt), and the household debt/house price ratio, a rough measure of leverage in housing. Note however that the equilibrium level of the debt/income ratio may be rising, as cross country analysis suggests that the income elasticity of credit exceeds 1 (Badev et al 2014). The authors also note the ratio is higher in countries with mortgage bonds as a primary funding source. Table 3 shows the more recent boom period was characterised by greater rises leverage on both measures (and also from a higher base). On average, the debt/income ratio for households rose by around 25 percentage points over 2002-6 as compared to only 8-9% in 1985-89. Obviously underlying this is the greater relative buoyancy of incomes in the earlier period as shown in Table 1. Meanwhile, the rise in debt deflated by house prices was also much higher in the recent boom, being around 15% compared with 8-9%.
These patterns are of interest since the earlier boom is often characterised as an adjustment to desired levels of leverage following liberalisation, when in fact rises were smaller than in recent years. This is an indicator of greater fragility of households in the 2000s. All other things were not of course equal in that interest rates were typically higher in the earlier period, meaning that the rise in the interest burden was less in the later period than if the same rise in debt had occurred in the earlier period. That said, the recent rise in debt and in leverage did leave many households vulnerable to negative equity when nominal house prices fell.
As regards the comparable figures for the aftermath periods, households reduced their
debt/income ratios in 1990-4 but they rose over 2007-11, albeit not in the UK or US. The
debt/house price ratio rose in both post-boom periods, with house price rises being lower than
changes in household debt. The run-up is remarkably high on average at around 20% in both cycles.
Concluding this section, we have seen a great deal of commonality between the booms and their aftermath from 1985-94 and 2002-11, notably in real house price rises and in their main determinants. There are also some contrasts. These relate especially to weaker growth in incomes in both the boom and the aftermath in the later period, while on most measures, debt and indebtedness rose to a greater extent, even though house price patterns in both boom and aftermath were on average very comparable. Correlations of house prices with income seem to be lower and those with household debt higher in the later period. We now go on to further investigate of possible changes and similarities to house price determination over the different cycles since liberalisation, which is detailed in the following sections.
2 Specifications for house price determination
8Having looked statistically at the cycles we now seek an econometric approach to house prices to assess differences across cycles more systematically. Typical estimates for determination of house prices are in error correction format. There is first a cointegrating levels equation which forms an inverted demand function for housing but also includes a supply effect such as the stock of housing which determines the long-run price of housing (Meen (2002), Barrell and Kirby (2004 2011) Adams and Fuss (2012), Loungini and Igan (2012), Muellbauer and Murphy (2006, 2008), Capozza et al. (2002)) . The second stage estimation of the dynamics recognises that actual house prices deviate from their fundamental values in the short-run and typically uses an error correction framework to allow for these differences. This allows the examination of factors that drive house price dynamics. The two stages may be combined, as in our work shown below, in a single stage error correction estimation.
In this context, considering housing as an asset among others, Capozza et al (2002) specifically focus on the properties of serial correlation and mean reversion of house prices in such an error correction framework. Informational reasons, transaction costs, credit rationing and supply side factors help explain serial correlation and mean reversion which may in turn differ across countries and time. To test the above proposition, they augment the long-run relationship with dynamic terms according to:
ΔP
t=αΔP
t−1+ β ( P
t¿− P
t−1)+ γ ΔP
t¿(4)
where
α is the serial correlation coefficient
β is the mean reversion coefficient to the gap with the long run value P* determined by the cointegrating equation and the adjustment to disequilibrium 0< β <1
γ is the immediate partial adjustment to the long run value
In general as α increases, the amplitude and persistence of the cycle will increase whilst as
β increases the frequency and the amplitude of the cycle will increase. Note that this structure implies that house prices do not follow a random walk unlike tradable financial assets but rather are predictable. We incorporate this structure into our own work, with the
8 This section draws partly on earlier work for the Swedish Riksbank by Davis, Fic and Karim (2011).
partial adjustment to the long term value being incorporated by dynamic difference terms in each non-stationary variable.
For our long run we follow in the approach in the literature of a log-linear transformation of all the variables, a cointegrating relationship would be identified with those fundamentals that possess a unit root (defining P*). Studies vary in terms of the members of the vector of fundamentals for the inverted demand function. For example, in Capozza et al. (2002) the set of long-run determinants includes population levels, real median income levels, the long-run (5 year) population growth rate, real construction costs and the user cost of housing. In Muellbauer and Murphy (2008) the vector of long-run variables includes real disposable (non- property) income, the sum of mortgage rates and stamp duty rates, the national credit conditions index and a term which interacts the mortgage rate with the credit conditions index. Barrell, Kirby and Whitworth (2011) include the real borrowing rate, the 3-month nominal interest rate, the loan-to-income ratio, the loan-to-value ratio, per capita real disposable income, the ratio of the number of households to the housing stock, and the number of households.
9Adams and Fuss (2010) include economic activity, construction costs and the long term interest rate. Loungini and Igan (2012) model real house price changes as a function of changes in disposable income, working-age population, equity prices, credit, and the level of short- and long-term interest rates. Our previous work (Davis et al 2011) in line with but also broadening the literature, used real personal disposable income, the real long rate, real household liabilities, real gross financial wealth, the unemployment rate, log real housing stock and 20-39 as a share of population (the main house buying cohort).
As regards econometric approaches, the studies cited above among others specify dynamics by using autoregressive distributed lag models in panel error correction form, with a one period lag on the long run to control for endogeneity. The VAR (Hott and Monin, (2008), Calza et al (2013)) and the SVAR (Tsatsaronis and Zhu, 2004) are also commonly used to estimate dynamics since such studies can then focus on the interdependencies of house prices and their determinants such as term spreads, house price inflation, GDP growth and the growth rate of private sector credit,. Other approaches include the VECM (Kemme and Roy (2012), Gattini and Hiebert (2010), Lindner (2014)) and spatio-temporal impulse responses to gauge the degree to which shocks diffuse over time and space (Holly, Pesaran and Yamagata (2010)) Some recent studies have looked at housing booms and busts as individual observations and estimated determinants by probit (Agnello and Schuknecht (2011), Benetrix et al (2012)..
Whereas many studies have focused on house price determination in an individual country (such as Muellbauer and Murphy (1997, 2008) and Barrell et al (2011) for the UK and Lindner (2014) for the US) a number of recent pooled or panel studies are also extant. Besides our own work (Davis et al 2011) for 18 OECD countries, which was focused on the possible use of macroprudential tools in housing, Capozza et al (2002) look at US Metropolitan areas, Adams and Fuss (2010) apply panel cointegration to 15 countries using Dynamic Ordinary Least Squares, while Igan and Loungini (2012) apply pooled OLS to 22 countries.
All of these approaches are fraught with identification problems, which make it difficult to separate supply and demand factors, and exogenous and endogenous determinants of house prices. All work on house prices faces this challenge and there is no definitive solution.
Concerning identification in error correction models,
10there are several hard to observe variables in a house price model: the risk premium and expected appreciation. Identifying
9 Estimating solely for the UK, there is scope for a much wider range of variables than in panel studies such as Adams and Fuss (2010), Loungini and Igan (2012) and our own work
10 We thank John Muellbauer for these insights.
these would be a problem inside or outside the single equation framework. So it will always be hard to give strict structural interpretations to an error correction model in the absence of very good survey data that tried to measure these concepts. However, it can still be argued that on reasonably plausible assumptions, one can still identify structural parameters such as the implied income elasticity of demand for housing and the implied price elasticity by estimating an inverse demand model, as do the authors above. Muellbauer argues that if the risk premium is determined by the same variables as house prices, then one can still identify the income or price elasticity of demand. Meanwhile expected appreciation may be captured by a lagged difference as in most extant work. We follow his approach in our work.
Meanwhile SVARs can impose appropriate identifying restrictions, while in VARs and VECMs shocks can be identified using the Chelsi decomposition.
Some variables have typically been omitted from house price equations, although economic reasons for their inclusion can be suggested. For example, unemployment may impact on house prices via demand and also if it entails widespread defaults and consequent “fire sales”
but is typically not included in house price equations. Indeed in Andrews (2010) the unemployment rate is used as part of the identification framework as a form of demand shock.
Financial liberalisation distinguishes periods when there is or is not credit rationing and is also used by Andrews (ibid) as showing demand shocks. Banking crises give rise to uncertainty and credit rationing that other variables may not adequately capture and is a third form of demand shock. We add all three of these variables to our work.
Mortgage spreads (loan less deposit rates) are also typically not included in house price equations, whereas these could be relevant to the impact of capital requirements on interest rates, as in Barrell et al (2009) and Davis and Liadze (2012) and have important
consequences for household incomes as well as for house price dynamics.
Furthermore, although housing is part of the asset portfolio of the household sector, most studies do not include household gross financial wealth, as a substitute asset, a rise in whose value would lead to rising demand for housing for portfolio balance reasons. Another
portfolio effect could be included via the long term interest rate, which is both a proxy for the user cost (especially influencing mortgage rates) but also the opportunity cost of investing in housing when the bond yield changes (Adams and Fuss 2010).
3 Specification and data
In the light of the data and the above brief literature survey, we sought to estimate panel equations for house prices in OECD countries. Given the extensive availability of cross- country data from the BIS, UN and OECD databases,
11we have scope to investigate the common patterns of property price movements, while at the same time controlling for heterogeneity across time in housing dynamics as well as between countries. From an
econometric perspective, a panel approach gives more informative data, more variability, less collinearity among variables, more degrees of freedom and more efficiency (Baltagi, 2005, p.
5). Following Capozza et al (2002) we allow for serial correlation and mean reversion as well as sensible long run variables in an inverse demand function estimated as an error correction model
The data sample we are able to use for most countries is back to the 1970s. We hence include periods when there has been liberalisation as well as structural regulation in the housing market. This can be justified by the need for cointegration equations to have as long a data period as possible, but will also enable us to capture the differences in behaviour between
11 Note that the population data that we use are interpolated annual data from the UN Demographic database.
liberalised and non-liberalised periods as well as between the cycles incorporating boom periods outlined in the tables of Section 1. We accordingly estimate for three sub-periods namely the pre-liberalisation period before 1982, the first post liberalisation cycle over 1982- 1997 and the second broad cycle over 1998-2013. Note that we use quarterly data for the cross country panel work and focus on the boom countries, namely UK, US, France, Canada, Italy, Spain, the Netherlands, Belgium, Ireland, Finland and Sweden.
Our variables are as follows: log real house prices, log real personal disposable income, the real long rate, log real household liabilities, log real gross financial wealth, unemployment rate, log real housing stock and 20-39 as a share of population (the main house buying cohort – which in countries such as the UK is also strongly driven by immigration in recent years, in turn affecting house prices). The Im-Pesaran-Shin panel unit root tests for the main variables (not illustrated) show most variables, being trended, are I(1) thus justifying an error
correction model based approach to estimation, while the share of 20-39s is stationary (I(0)).
Changes in real house prices were regressed on contemporaneous changes in explanatory variables, and lagged dependent and explanatory variables (both in levels) as well. This error- correction specification is able to deal with non-stationarity in the data (as mentioned above), and at the same time distinguishing short- and long-run influences, and differences between cycles. The significance of the coefficients for lagged non-stationary variables (in levels) and their magnitude reveal the long-term relationship among those variables.
Our modelling started from the approach of Capozza et al (2002) set out above with variables as in Davis, Fic and Karim (2011) with a basic set of variables including real house prices, real personal disposable income and the long term real interest rate (proxying the user cost), to which we add extra variables in difference and level; rate of unemployment, real gross financial wealth (as a portfolio balance effect), housing stock (lag only), the population of 20- 39 as a proportion of the total (lag only) and dummies for financial crises and the onset of liberalisation. We undertake panel regression that treats all countries as equally important, while the fixed effects take account of heterogeneity, and we impose cross section weights.
The breakdown over sub periods offer deeper insights by allowing for richer heterogeneity, e.g. distinctive economic determinants in each sub-sample (compared to the full sample regression). The combination of the full period regression and the sub-sample panel
regressions reveal elements of both commonality and uniqueness in cycles in those countries.
To confirm the existence of the long-term relationship, we also implement the panel
cointegration test proposed by Kao (1999) among those variables with significant lagged level terms in a simple levels equation (i.e. the first step of an Engle and Granger (1987) two-step estimation).
4 Results
Further to the discussion above, we present the results for an extended equation including house prices, RPDI and real long rates but also including the log of real gross financial wealth, the unemployment rate, the log of the real housing stock and the 20-39 age group as a share of the population (see Davis Fic and Karim 2011 for earlier estimates of such a wider specification using annual data).
Table 4: Panel results for the log difference of house prices – boom countries
All Pre 1982 1982-1997 1998-2013
Constant
0.001(0.1)
-0.77**
(2.4)
0.25**
(2.1)
0.093 (1.2)
Log
difference of
RPDI
0.17**(6.7) 0.25**(3.3) 0.15**
(4.0) 0.19**
(5.5)
Difference
real long rate
-0.00011(0.2)0.00099 (0.5)
-0.00094 (1.0)
-7.13E-05 (0.1)
Log
difference of house prices
(-1)
0.56**(28.1)0.41**
(7.1)
0.53**
(15.7)
0.54**
(16.6)
Log of house
prices (-1)
-0.0097**(4.7) -0.045**(2.4) -0.034**
(5.5) -0.013**
(3.0)
Log of
RPDI(-1)
-9.26E-05(0.0)0.078 (1.5)
-0.0073 (0.5)
0.043**
(3.6)
Real long
rate (-1)
-0.0008**(4.0)-0.00071 (0.7)
-0.00073*
(1.7)
-0.00062 (0.7) Population 20-
39 as share of total (-1)
0.032 (1.4)
-0.61*
(1.9)
0.12*
(1.8)
-0.081*
(1.8) Log stock of
housing (-1)
-0.0054 (1.1)
0.027 (0.5)
-0.037**
(2.7)
-0.031**
(3.1) Difference of
unemployment rate
-0.0041**
(3.5)
-0.0077*
(1.7)
-0.0048**
(2.9)
-0.0027**
(2.1) Unemploymen
t rate (-1)
-1.47E-05 (0.1)
-0.00052 (0.3)
-0.00074**
(2.3)
-0.00054*
(1.8) Log difference
of real gross financial
wealth
0.053**
(4.6)
0.052*
(1.6)
0.07**
(4.1)
0.038**
(2.5) Log of real
gross financial wealth (-1)
0.008**
(3.5)
-0.00089 (1.1)
0.028**
(5.2)
-0.014**
(2.2) Dummy for
banking crises
-0.0032**
(2.7)
-0.0034**
(2.2)
-0.0038**
(2.5) Dummy for
financial
liberalisation 0.00026 (0.2)
Countries 11 10 11 11
Obs 1612 275 687 650
Adjusted R2 0.5 0.38 0.53 0.6
SE of regression
0.16 0.02 0.16 0.011
Durbin
Watson 2.13 2.09 2.09 2.11
Kao -1.58
(0.06)*
-1.85 (0.03)**
-2.37 (0.01)**
-2.54 (0.01)**
Notes:
(-1) indicates a first lag. Boom countries for both recent cycles are the UK, US, France, Canada, Italy, Spain, the Netherlands, Belgium, Ireland, Finland and Sweden. Estimated using fixed effects and cross-section weights.
Coefficients marked ** are significant at the 95% level and * are significant at the 90% level (t values are in brackets under each coefficient).We find a consistent short run income effect, albeit it is lower after liberalisation. On the other
hand, the short run effect of interest rates is insignificant. The serial correlation effect is very
strong (i.e. the lagged first difference of real house prices) and rising over the sample. As
noted above this implies a higher amplitude and persistence of the cycle and a growing role
for extrapolative expectations in most recent cycles. The lagged house price variable is
generally significant. The implied speed of adjustment to the long run is lower since
liberalisation, suggesting longer cycles. Adams and Fuss (2010) find a similar long
adjustment period of 14 years in a cross country panel on a recent sample. The long run
income effect is positive and significant but only in the most recent period. The long run interest rate effect is significant at the 10% level in the 1982-97 period only.
For the population distribution, signs change between periods. The share of 25-39’s in the total population who are the main house buyers may be overwhelmed by the ageing of the large baby boom generation that has the resources to buy houses at any age. The long run effect of the housing stock is significant post liberalisation with an expected negative sign whereby a higher stock (indicating greater supply) leads to lower house prices. The change in unemployment is generally significant, albeit lower post liberalisation. The long run effect of unemployment is significant post liberalisation. The short run financial wealth effect is generally significant and positive, suggesting a portfolio balance effect (higher financial wealth is distributed to housing as an additional asset). On the other hand, whereas the long run financial wealth effect is significant post liberalisation its sign changes (this may reflect stock market patterns). The banking crisis dummy is consistently significant, while the liberalisation dummy is not. The Kao (1999) tests show consistent cointegration in the first stage levels variables. On balance we suggest that these results do not suggest radical differences between the two cycles since liberalisation.
Table 5: Leveraged coefficients for 1982-1997 (in regression 1982-2013)
Coefficien
t T-
value
Log difference of RPDI 0.022 (0.4)
Difference real long rate 0.0015 (0.8) Log difference of house prices (-1) -0.029 (0.6) Log of house prices (-1) 0.0087 (2.3)**
Log of RPDI(-1) -0.0024 (0.8)
Real long rate (-1) 0.0011 (1.2)
Population 20-39 as share of total (-1) -0.051 (1.3) Log stock of housing (-1) 0.0031 (1.0) Difference of unemployment rate -0.0048 (2.2)**
Unemployment rate (-1) 0.00041 (1.1)
Log difference of real gross financial
wealth 0.034 (1.4)
Log of real gross financial wealth (-1) -0.00066 (0.3)
Notes: (-1) indicates a first lag. Boom countries for both recent cycles are the UK, US, France, Canada, Italy, Spain, the Netherlands, Belgium, Ireland, Finland and Sweden. Estimated using fixed effects and cross-section weights.
Coefficients marked ** are significant at the 95% level and * are significant at the 90% level (t values are in brackets under each coefficient).Coefficients shown in Table 4 are also included but not reported.
Complementing Table 4, in Table 5 we show leveraged coefficients for the earlier cycle 1982- 97 in a regression for 1982-2013. This shows that the only significant differences are mean reversion being lower in the 1997-2013 period, while the impact of unemployment was higher in the 1980s. Serial correlation is the same. Overall, this is strong evidence that the cycles are similar.
In a further exercise we looked at leveraged effects during the booms, testing whether there is a differential effect of the determinants in such periods, as shown in Table 6.
Table 6: Panel results for the log difference of house prices – boom countries – leveraged coefficient for booms
Estimation period, 1982q1 to 2013q4 Leveraged
coefficient for Leveraged
coefficient Leveraged
coefficien
period 1985q1- 1989q4 and 2002q1- 2006q4
for period 1985q1- 1989q4
t for period 2002q1- 2006q4 Log difference of RPDI
0.10*(1.9) 0.22**
(3.3) -0.083 (1.0)
Difference of real long
rate
0.0036**(2.4)0.0034**
(2.0)
0.00099 (0.3)
Log difference of house
prices (-1)
0.099**(2.2)0.076 (1.5)
0.015 (0.2)
Log of house prices (-1)
0.0016(0.6) 0.0047
(1.5) -0.015*
(1.9)
Log of RPDI(-1)
0.002(0.8) -0.00089
(0.2) 0.0041 (1.1)
Real long rate (-1)
-0.00012(0.3)
-0.00048 (0.6)
-0.0023 (1.0) Population 20-39 as share
of total (-1)
0.048*
(1.7)
0.047 (1.1)
0.064 (1.5) Log stock of housing (-1)
-0.0025 (1.4)
0.00068 (0.2)
-0.0036 (0.9) Difference of
unemployment rate 0.0032
(1.0) -0.0016
(0.3) 0.0032 (0.7) Unemployment rate (-1) 0.00033
(1.1) 0.00047
(1.3) -0.00041 (0.5) Log difference of real
gross financial wealth
0.014
(0.6) 0.023
(0.9) -0.0078 (0.2) Log of real gross financial
wealth (-1)
-0.0018 (0.6)
-0.0072*
(1.7)
0.00091 (0.1)
Notes: (-1) indicates a first lag. Boom countries for both recent cycles are the UK, US, France, Canada, Italy, Spain, the Netherlands, Belgium, Ireland, Finland and Sweden. Estimated using fixed effects and cross-section weights.
Coefficients marked ** are significant at the 95% level and * are significant at the 90% level (t values are in brackets under each coefficient).Coefficients shown in Table 4 are also included but not reported.
Leveraged coefficients show a higher effect for the rise in RPDI and a lower (negative) effect for RR. There is shown to be more serial correlation with a larger coefficient on the lagged difference of house prices, consistent with the suggestion in Dokko et al (2011) and Shiller (2007) that expectations of future house price growth among borrowers, lenders and investors plays a key role in bubbles. The demographic effect of a higher number of 25-39 year olds has a higher effect in booms also consistent with Muellbauer and Murphy (1997) on the 1980s boom in the UK. In the extended equation, it is again in the 1985-89 case that there are larger effects of rising income and lesser effects of rising interest rates. The earlier boom also saw a lower long run effect of gross financial wealth and a higher effect of debt, suggesting
households were leveraging themselves into real assets and partly substituting out of financial assets. The only difference for the later boom in the leveraged coefficients is in the long run adjustment coefficient, with a significant negative sign suggestive of more rapid adjustment to long run equilibrium. All of these leveraged results are of potential relevance for
macroprudential policy, suggesting normal house price behaviour in respect of determinants is not always maintained in booms. On the other hand they should not be exaggerated, for the most part the equations are stable.
5 House prices and mortgage supply
Mortgage market innovations that have greatly altered the terms and availability of credit
have emerged in OECD financial markets over the past 30 years (OECD, 2005). Financial
deregulation in the 1980s not only increased competition, it has also led to the creation of new
products such as buy to let mortgages, interest only loans and offset mortgages which allow
borrowers to offset their savings against the mortgage balance. Meanwhile, the widespread
development of the securitisation markets in the 2000s, following their earlier evolution in the
US (Hendershott 1994) eased access to mortgage credit further since it is no longer limited by the capital of the originating institution.
As a result of such innovations, the availability of mortgage credit has risen dramatically in Europe and the US. Miles and Pillonca (2008) note that although the mortgage debt to GDP ratio varies across Europe (exceeding 70% in countries like the UK and Denmark), the stock of mortgage debt has risen in all cases. Consequently house buyers have seen a relaxation in their borrowing constraints and they contend that this has fed back positively to house prices.
Few house price models have taken these fundamental changes into account. Indeed, a key question raised by financial liberalisation is whether the stock of mortgages is appropriately included in house price equations. This was traditionally the case in pre liberalisation estimates in countries such as the UK (e.g. Hendry 1984) but was judged by authors such as Muellbauer and Murphy (1997) to be inappropriate in a post liberalisation sample, since the stock of lending is endogenous to the determination of house prices. On the other hand, if there remains a degree of rationing for some participants in the housing market, then the mortgage stock could have a role to play, and all the more if macroprudential policies have an effect of reintroducing forms of credit rationing.
An alternative way of considering this question is set out in Lindner (2014), who notes there are two alternative views of the link from asset prices (such as those of housing) to credit. The first is the Bernanke and Gertler (1989) and Kyotaki and Moore (1987) view that it is asset prices that drive credit availability via changes in the net worth of borrowers that in turn eases borrowing constraints in the presence of asymmetric information. This is consistent with the exclusion of credit from house price equations. On the other hand, Allen and Gale (2000) suggest that the availability of credit is the more exogenous factor, with the key influence being risk shifting by lenders and borrowers in the presence of asymmetric information and limited liability, with consequent moral hazard. These may in turn be facilitated by financial deregulation. Lindner (2014) suggests that the net worth argument is most relevant to credit availability in general whereas risk shifting is appropriate for the financing of a particular asset such as housing by credit. Consistent with this, empirical studies using total credit (such as Davis and Zhu 2011) tend to be more consistent with one-way causality from asset prices to credit than those focused on housing (such as Gimeno and Martinez-Carrascal 2010) which find two way causality. Lindner (2014) finds mortgage credit does drive house prices in the US although there is also Granger causality in the other direction.
Calza et al (2013) show that the structure of housing finance has an impact on the transmission of interest rates to both house prices and consumption. Igan and Loungini (2012) find a significant effect of the difference of credit but add that due to potential endogeneity they comment that “we refrain from interpreting the positive correlation between credit growth and house price appreciation as causation and leave establishment of such a causal link for further research” (ibid p16) We proxy credit to attempt to overcome this problem.
Meanwhile, Muellbauer and Murphy (2008) include a credit conditions index which they
introduce both alone and as an interaction term with the mortgage rate. The credit conditions
index is constructed using 10 consumer credit and mortgage market indicators as described in
Fernandez-Corugedo and Muellbauer (2006). It is included so as to capture shifts in the credit
supply function faced by households in the post-1980s era. The authors note that by omitting
this variable, previous house price models in the literature (which typically utilise pre-1980s
data) suffer from omitted variable bias. Meanwhile, Claessens et al (2011) contend that credit
spreads and credit conditions may be more relevant to macroeconomic trends than the volume
of credit.
In our work we use the simpler measure of the real stock of mortgages as a credit variable, to provide some suggestive results on the potential effects of credit and liberalisation thereof in the different booms.
Table 7: Panel results for the log difference of house prices – boom countries – adding debt variables
All Pre 1982 1982-1997 1998-2013 1982-2013
Proxy for log difference of real
household debt
0.092**(11.0) 0.11**(4.9) 0.07**
(5.5) 0.1**
(6.4)
0.088**
(9.1)
Log of real
household debt(-
1)
-0.0022(0.9)0.0082 (0.2)
-0.0047 (0.8)
-0.0046 (0.8)
-0.004*
(1.7)
Notes: (-1) indicates a first lag. Boom countries for both recent cycles are the UK, US, France, Canada, Italy, Spain, the Netherlands, Belgium, Ireland, Finland and Sweden. Estimated using fixed effects and cross-section weights.
Coefficients marked ** are significant at the 95% level and * are significant at the 90% level. (t values are in brackets under each coefficient)Coefficients shown in Table 4 are also included but not reported.
We went on to test within the panel error correction framework by adding the level and difference of the real mortgage debt stock to the extended equation. As regards mortgages, no long run effect of the debt stock on house prices is detectable, even pre liberalisation; on the other hand, the short run effect is consistently significant (proxied by lags to avoid
simultaneity). Credit is shown to have a short run but not a long run impact on house prices during boom periods, justifying a focus of macroprudential policy on credit for this reason as well as due to risk,but with no major distinction for the latest cycle.
Table 8: Panel results extended equation – boom countries – leveraged coefficient for booms and aftermaths
Notes: (-1) indicates a first lag. Boom countries for both recent cycles are the UK, US, France, Canada, Italy, Spain, the Netherlands, Belgium, Ireland, Finland and Sweden. Estimated using fixed effects and cross-section weights.
Coefficients marked ** are significant at the 95% level and * are significant at the 90%
level (t values are in brackets under each coefficient).
Coefficients shown in Table 4 are also included but not reported.
Using leveraged coefficients, we see that both the difference and the level effect of credit is significantly more positive in booms than in other periods while there is no corresponding effect in the aftermath except in 1990-4when effects were again more sizeable. In other words, both a rise in credit and a higher level have a significant effect on house prices. This is consistent with the suggestion that financial liberalisation had a significant effect on the booms, again offering grounds for caution in macroprudential policy. This effect was most strongly present in the earlier boom
Estimation period, 1982q1 to 2013q4
Leveraged coefficient for period 1985q1- 1989q4 and 2002q1- 2006q4
Leveraged coefficient for period 1985q1- 1989q4
Leveraged coefficient for period 2002q1- 2006q4 Log difference of real
liabilities (proxy)
0.034*(1.7) 0.028(1.3) -0.0014 (0.1)
Log real liabilities (-
1)
0.00053**(3.6)0.00071**
(3.6)
-0.00012 (0.7)
Leveraged
coefficient for periods 1990q1- 1994q4 and 2007q1- 2011q4
Leveraged coefficient for period 1990q1- 1994q4
Leveraged coefficient for period 2007q1- 2011q4
Log difference of real
liabilities (proxy)
0.016(0.7)0.061*
(1.8)
0.045 (1.3)
Log real liabilities (-
1)
-9.8E-05(0.7) 0.00033*(1.6) -0.00016 (0.8)
