An early warning system to predict house price bubbles
Christian Dreger, Konstantin Kholodilin DIW Berlin
Why is the topic relevant?
• Bubbles in housing prices can impede the real economic evolution
–Prices do not indicate scarcity, distorts allocation of resources
–House price bubbles can inflate over long periods, high risk of recession after their burst
• Observed prices include both fundamental and speculative components
–Separation depends on underlying model
• Early detection can improve policy reactions
Early warning system
• Based on two-stage procedure
–Previous bubbles in housing markets need to be identified (bubble chronology)
–Statistical model fitted to explain the chronology
• Results might vary with empirical models –Robustness tests to identify critical price drivers
• Early warning system for panel of countries or regions, as bubbles are rare events
–Monetary policy identical for regions within a country
Chronology of price bubbles
• Identification using structural econometric models and filter methods
–Country-by-country since bubbles are distributed heterogeneously across panel members
• Dominance of structural models
–Prices can be linked to macroeconomic conditions (fundmentals)
• Filter methods to validate structural bubbles –Structural bubbles need to be confirmed, periods
should at least partially overlap
Filter and structural models
• Deviations from HP trend and (smoothed) residuals from structural regression should exceed threshold
–Threshold expressed as a multiple of standard deviations, might differ across panel members
• Thresholds by iteration to obtain maximum concordance
Filter: cycleit = rpit −trendit >θσi(cyclei) Regression: pit =αi′Xit +εit , εˆit > ρσ εi( it)
Set of fundamentals
•
Potential drivers of excessive house price excluded in model fit
–If not, bubbles are wiped out, as OLS provides optimal fit
•
Fundamental variables bear correct sign
–Higher income triggers rising house prices –Increasing real interest rates will worse
financing conditions and dampen house prices –High urbanization implies lower migration
pressure, negative impact on house prices
Chronology house price bubbles
Australien
-0.2-0.10.00.10.20.3
1986q3 1989q3 1992q3 1995q3 1998q3 2001q3 2004q3 2007q3
Kanada
-0.15-0.050.050.15
1970q2 1975q2 1980q2 1985q2 1990q2 1995q2 2000q2 2005q2
Frankreich
-0.6-0.4-0.20.00.2
1970q2 1975q2 1980q2 1985q2 1990q2 1995q2 2000q2 2005q2
Deutschland
-0.10-0.06-0.020.02
1991q2 1994q2 1997q2 2000q2 2003q2 2006q2 2009q2 Italien
-0.4-0.20.00.2
1981q21985q21989q21993q21997q22001q22005q22009q2
Japan
-0.6-0.4-0.20.00.2
1969q4 1974q4 1979q4 1984q4 1989q4 1994q4 1999q4 2004q4 2009q4
Niederlande
-0.8-0.6-0.4-0.20.00.2
1977q2 1982q2 1987q21992q2 1997q22002q2 2007q2
Portugal
-0.040.000.040.08
1995q21997q21999q22001q22003q22005q22007q22009q2 Spanien
-0.8-0.40.00.20.40.6
1971q2 1976q2 1981q2 1986q2 1991q2 1996q2 2001q2 2006q2
Schweden
-0.8-0.6-0.4-0.20.0
1993q2 1996q2 1999q2 2002q2 2005q2 2008q2
Schweiz
-0.3-0.2-0.10.00.10.2
1970q2 1975q2 1980q21985q2 1990q2 1995q22000q2
UK
-1.2-0.8-0.40.0
1971q1 1976q1 1981q1 1986q1 1991q1 1996q1 2001q1 2006q1 USA
-0.6-0.4-0.20.0
1975q21980q21985q21990q21995q22000q22005q2
Distribution of bubbles
Number of bubbles Duration (Quarter)
Australia 3 8.7
Canada 4 8.8
France 2 10.0
Germany 1 8.0
Italy 2 9.0
Japan 1 18.0
Netherlands 1 5.0
Portugal 1 10.0
Spain 4 13.0
Sweden 2 5.5
Switzerland 2 12.0
UK 3 14.3
USA 2 14.0
Prediction of bubbles
• Identification of potential variables driving the emergence and inflation of bubbles
–Calibrated to explain the bubble chronology
• Signal approach
–Signal for bubble extracted, if particular variable exceeds critical value
• Probability for presence of bubbles captured by logit and probit models
• Panel with country fixed effects, institutional conditions
Signal approach
• Evolution of variables in periods of previous bubbles
–Standardization of variables to eliminate country specific impact on overall volatility
• Thresholds ensure optimal bubble forecast for individual variable
–Bubbles should be detected, false alarms avoided
• Signal extracted, if variable above threshold
• Overall indicator results from the aggregation of individual signals
Thresholds for signal extraction
• Ratio of correctly predicted bubbles non bubbles) to all bubbles (non bubbles)
• Variables with more accurate forecasts receive higher weight in overall indicator
Event Bubble No bubble
Signal a b
No signal c d
i max
a d
z = a c +b d →
+ +
Logit and probit models
• Binary response model (bubble, no bubble)
• Probability for the presence of bubble traced to set of variables
–Logistic or standard normal distribution
–Instead of thresholds, regressors are evaluated by statistical significance
–Ordering of contribution of regressors by their log-odds ratio
• Goodness-of fit measures to rank different approaches
Empirical results
• Signal-, logit- and probit model based on the same set of variables
• Excessive liquidity and credit conditions are highly relevant to predict speculative housing bubbles
–Confirms previous findings
• But other variables are also important
–Investment ratios, public finances, interest rates
• Logit and probit models outperform the signal approach
Variables to predict bubbles
Weights Signal Logit
Short term interest rates 15.0
Real exchange rate 6.9 7.1
Rent 5.0
House prices / income 7.6 21.7
House prices / rent 8.1 7.2
Investment ratio 7.9 7.2
Credit growth 19.5 20.9
GDP growth 6.5 7.2
Liquidity (based on M3) 19.7 14.4
Public finances 3.8 7.0
Financial regulation 7.3
Ranking of models
• Quadratic probability score: yis (0, 1) variable ppredicted probability for a bubble
• One-step recursive forecasts, perfect foresight for exogeneous variables
2 1
1 ( ) min
T
j
it it
t
QPS y p
T =
=
∑
− →In sample Out of sample
Signal approach 0.278 0.292
Logit model 0.081 0.134
Probit model 0.081 0.139
Logit detection of bubbles
Australien
0.00.20.40.60.81.0
1986Q3 1993Q3 2000Q3 2007Q3
Kanada
0.00.20.40.60.81.0
1970Q41981Q4 1992Q42003Q4
Frankreich
0.00.20.40.60.81.0
1980Q31988Q31996Q32004Q3
Deutschland
0.00.20.40.60.81.0
1991Q21996Q22001Q22006Q2 Italien
0.00.20.40.60.81.0
1997Q2 2001Q2 2005Q2 2009Q2
Japan
0.00.20.40.60.81.0
1991Q41996Q42001Q42006Q4
Niederlande
0.00.20.40.60.81.0
1991Q11996Q12001Q12006Q1
Spanien
0.00.20.40.60.81.0
1998Q2 2002Q2 2006Q2
Schweden
0.00.20.40.60.81.0
Schweiz
0.00.20.40.60.81.0
UK
0.00.20.40.60.81.0
USA
0.00.20.40.60.81.0
Conclusions
• Development of a tool for early detection of speculative bubbles in housing markets
–Robustness analysis of results
• Logit and probit outperform signal approach –Probably not in general, since different weights for
threshold components can affect results
• Liquidity and credit variables not sufficient to predict bubbles
–Represent 35 (logit) and 55 (signal) percent of forecast of overall indicator
Further information
• Email: [email protected]
• Early warning systems to detect price bubbles in the stock and housing market
–Sponsored by the German Ministry of Finance
• An early warning system to predict speculative house price bubbles, Economics: Open Access, Open Assessment E-Journal 7, 2013-8 (Dreger and Kholodilin).
