The Economist - UK (2019-06-29)

100

300

200

70

400

Log

scale

1990 95 2000 05 10 15 Q2

2020

Q1

2019

Australia

Canada

Germany

France

Britain

Ireland

Italy

NewZealand

Spain

United

States

←Actual Forecast→

90%

confidence

interval

Mean

forecast

House-priceforecast,Q2 2020

%changeona yearearlier,realterms

Confidenceintervals,%

-10 -5 0 5 10 15

50 75 90 95

Median

Australia

N.Zealand

Canada

Britain

France

Germany

Ireland

Italy

US

Spain

Adecadeafterthefinancialcrisis,housepricesareatnewhighs

Sources:OECD;BIS;IMF;nationalstatistics;TheEconomist

Realhouseprices

Q11990=100

TheEconomistJune 29th 2019 85

I

nvestors focuson shares and bonds,

but one asset class is bigger than the two

combined. Put together, the world’s homes

are worth over $200trn. House prices are

crucial harbingers of economic trends: the

last time they fell across the rich world, it

set off the deepest downturn in decades.

Ten years have passed since the Great

Recession, and home values have made

back most of their losses. In Canada and

New Zealand they are 40% above the pre-

crisis peak. Does another crash loom?

None of the main international institu-

tions, such as the imfor oecd, includes

residential property in its standard battery

of economic forecasts. That may be be-

cause home values depend on local factors.

However, The Economisthas kept a database

of house prices for decades, using figures

from the oecdand national agencies. And

even an inexact forecast provides more in-

sight than no forecast at all. As a result, we

have designed a model to predict changes

in real home values at the national level.

Our system relies on three types of data.

First come economic figures such as gdp

growth and interest rates. Next are market

fundamentals, like the ratios of home

prices to rents and incomes. Last come his-

torical prices, to take into account momen-

tum and mean reversion.

The impact of each of these variables of-

ten depends on the others. To combine

them, we used a machine-learning algo-

rithm called a random forest. This method

creates a “forest” of “decision trees”, each

containing a series of yes/no choices such

as “Has gdpbeen rising?” or “Are price-to-

rent ratios below the long-run average?”,

and averages the output of each tree.

The model fares well in back-testing. On

average, its forecasts with 18 months’ lead

time came within three percentage points

of actual yearly price changes. These errors

are larger during booms or busts—but still

small enough for the model to be useful.

For example, in the year to March 2006

American house prices rose by 8%. Our

model expected growth would slow to 0.3%

in the year to September 2007. That was too

sanguine: prices actually fell by 4.7%. But it

still would have served as an early warning.

According to our model, conditions to-

day are not similar to those of 2006. Across

ten countries, the average of its median es-

timates for the year to June 2020 is an ap-

preciation of 2.3%. The model does not rule

out a downturn: there is a one-in-seven

chance that Italian prices will fall by at least

5%. But the most likely scenario is that the

rally has room left to run. 7

Our model finds that prices are likely

to keep rising in the short run

As safe as houses

Graphic detailResidential property

Globalhouseprices,forecastv actual*

%changeona yearearlier,realterms

*Averageoftenrich-worldcountriesweightedbyGDP

-5

0

5

10

-10

1990 2000 2010 2019

Actual

Model’s forecast

18 months before