Applied Economics Quarterly Vol. 55. No 2 (2009)
Duncker & Humblot GmbH, 12165 Berlin
Google Econometrics and Unemployment Forecasting
By Nikolaos Askitas* and Klaus F. Zimmermann**
Abstract
The current economic crisis requires fast information to predict economic behavior early,
which is difficult at times of structural changes. This paper suggests an innovative new
method of using data on internet activity for that purpose. It demonstrates strong correlations
between keyword searches and unemployment rates using monthly German data and exhibits
a strong potential for the method used.
Keywords: Google, internet, keyword search, search engine, unemployment, predictions, timeseries analysis
JEL Classification: C22, C82, E17, E24, E37
1. Introduction
The internet contains an enormous amount of information which, to our knowledge, classical econometrics has yet to appropriately tap into. Such information
comes timely on a continual basis. It is particularly welcome at times of an economic crisis where the traditional flow of information is too slow to provide a
proper basis for sound economic decisions.1 Not only has traditional (and typically
official) statistical data a slow publication scheme, these data also do not reflect
well the structural changes in the economy. While investigating many different
kinds of internet activity, we focus here on Google search data to establish strong
correlations between search activities for certain keywords or keyword groups and
the unemployment rates in Germany. We call the relationship a Google predictor.
Such an application is timely, since we have just experienced a turning-point in the
fall of the unemployment rates after a longer decline caused by labor market reforms and the past economic boom. It is a particular challenge for the new proposed method to capture that turning-point properly.
* IZA.
** Bonn University, IZA, and DIW Berlin. Corresponding author: Klaus F. Zimmermann
IZA, P.O. Box 7240, D-53072 Bonn, Germany. Phone: +49 228 3894 200, Fax: +49 228 3894
210, Zimmermann@iza.org.
1 See Zimmermann (2008, 2009) for an analysis of the current challenges for economic
forecasting.
Applied Economics Quarterly 55 (2009) 2
108
Nikolaos Askitas and Klaus F. Zimmermann
Previous applications of Google search engine query data include Constant and
Zimmermann (2008) measuring economic and political activities, and Ginsberg,
Mohebbi, Patel, Brammer, Smolinski and Brilliant (2009) for studying influenza
epidemics. While the former study purely documents the evolution of particular
keyword searches before the US presidential elections, the latter investigates an
epidemic process using more complex computational methods. The novel feature
here in this paper is to demonstrate that the data can be used to predict economic
behavior measured by traditional statistical sources.
The study is structured as follows. In Section 2, we explain how we use Google
Insights and how we choose our indicator variables from the keyword searches. In
Section 3, we provide the empirical results. Section 4 contains our conclusions and
future plans.
2. Google Econometrics:
Unemployment Rates and Choice of Indicators
In the summer of 2008, Google released a beta version of Google Insights (http: //
www.google.com / insights / search / ). Using the service, search queries can be
compared for keywords across countries and in some cases their regions, in narrow
or wide time frames from 2004 onwards. A Google Insights query may have regional, temporal or keyword specific focus, i.e. you choose the region of interest, the
time frame of interest and the keywords of interest (up to 5). The results are then
delivered scaled and normalized within the query (for the region, the time frame
and the selection of keywords)2. This presents some interesting but not insurmountable challenges in accessing the data. Google Insights has also been modifying the
service since it was started, which caused changes in the way we were able to access the data ourselves. The data access is limited and restricted in many ways.
Ginsberg, Mohebbi, Patel, Brammer, Smolinski and Brilliant (2009) in their
study of influenza epidemics obviously had better access to more data and consequently were able to apply more complex computational methods. They demonstrate how flu epidemics can be predicted using Google Insights as its data source.
When we started to work on the idea to investigate human behavior measured by
traditional statistical sources using internet queries and to apply it to correlate keyword searches and unemployment rates, among other things, we were not aware of
their study. Knowledge of this work, however, encouraged us to proceed with our
paper. Given our restricted access to the data, we decided to attempt a minimalist
approach: theorize the choice of keywords, reduce our investigations to the parsimonious basics and demonstrate the power of the method.
2 We decided to query Google Insights for keywords one at a time. This way we lost the
information of the relative weight of keyword activity but freed ourselves of the problem of
having a large volume variable trivialize a low volume one. The idea is that a smaller group
of people may cause a low volume of keyword activity which contains as much or more
information than a keyword with large volume.
Applied Economics Quarterly 55 (2009) 2
Google Econometrics and Unemployment Forecasting
109
In order to motivate our investigation as well as our use of the data, we need to
set the stage by explaining the challenge we posed to ourselves. In Germany, the
unemployment rates are announced monthly at a press conference by the Federal
Employment Agency. The announcement dates are provided in advance for the next
two years and are almost always at the end of the month, but sometimes early in the
first week of the following month. This means that at the end of a given month M
the unemployment rate “for the month” is made known. We will denote this by UM .
The data used to compute UM is based on administrative data of the unemployment
office between the middle of month M 1 and the middle of month M.
This means that the announced unemployment rates for month M, which are
issued by the end of the month, are based on real unemployment processes occurring in the union of two time intervals:
The first interval denoted by W 34M
of month M 1.
1
is roughly speaking the 3rd and 4th week
The second interval denoted by W 12M is then the 1st and 2nd week of month M.
We should point our that practically in the middle of the two time intervals (i.e.
around the end of month M 1), we have the release of the unemployment rates for
month M 1, which is based on unemployment occurring in the intervals W 34M 2
and W 12M 1 . Figure 1 captures all the relevant information to set the stage for the
real monthly unemployment rate: how it is measured and when is it made known.
month M
month M-1
W34M-1
W12M
15
15
announce U M-2
announce U M-1
announce U M
Note: U is announced unemployment, M month, and Wxy refers to weeks x and y in a particular month.
Hence, UM is unemployment U in month M and W 12M refers to both weeks 1 and 2 in month M. 15 refers
to the 15th day in the particular month. The timing was revealed to us in informal communications with
high-ranked officials of the Federal Employment Agency.
Figure 1: The time structure of announcing unemployment
by the German Federal Employment Agency
Applied Economics Quarterly 55 (2009) 2
110
Nikolaos Askitas and Klaus F. Zimmermann
Google collects, normalizes and scales the number of searches for all kinds of
keywords, provided there is a “sufficient amount” of searches for this keyword.
The exact threshold is not known to us. We will say that the data Google Insights returns for a certain keyword k is the “Google activity along the keyword k”
and denote this by gk;12;M or gk;34;M in weeks 1 and 2, or 3 and 4 of month M.3
As unemployment occurs, people are also using Google for all kinds of keyword
searches. If we had access to the entire recorded Google activity along all keywords, we could attempt a more comprehensive approach, but even so we can
ask whether we can figure out a core set of keywords whose Google activity
would have predictive power for the monthly unemployment rates. Google returns the data in weekly values, and the week boundaries are known to us. They
do not contain the boundaries of our time intervals above, so we needed to resplit the activity proportionally to overcome this issue. We are aware that this
introduces a certain amount of noise, and in fact this is the reason why we
decided to use biweekly rather than weekly time intervals to minimize the noise
we introduce.
Our aim is to investigate the extent to which we can locate keywords whose activity gk;12;M and gk;34;M 1 may be used to predict UM . We expect activities in the
intervals W 34M 1 to have better predictive power than those in the interval W 12M ,
although the latter period is closer to the new announcement than the former. The
reason for this is that the rate UM 1 is announced in between the two intervals and
influences the activity gk;12;M , i.e. people react to the announcement. A similar impact to gk;34;M 1 may only come from UM 2 , which was announced two weeks prior
and is therefore less likely to be remembered.
We use measurements of Google activity along the disjunction of four groups of
keywords (Google Insights supports queries for disjunctions of keywords):
k1 = Arbeitsamt OR Arbeitsagentur (“unemployment office or agency”)
k 2 = Arbeitslosenquote (“unemployment rate”)
k 3 = Personalberater OR Personalberatung (“Personnel Consultant”)
k 4 = Stepstone OR Jobworld OR Jobscout OR Meinestadt OR meine Stadt OR Monster
Jobs OR Monster de OR Jobboerse (“most popular job search engines in Germany”)
We expect Google activity along k1 (Arbeitsamt or Arbeitsagentur) to be connected with people having contacted or being in the process of contacting the unemployment office. As such it should have something to do with the “flow into
unemployment”. The keyword k 2 (Arbeitslosenquote) is just the easiest and most
natural keyword to think of when dealing with unemployment. The activity around
the disjunction k 3 (Personalberater or Personalberatung) is expected to correlate
3 In our presentation and the resulting tables and graphs the variable convention we use is
as follows: the variable containing the values gn;34;M is called w34kn and the one containing
the values gn;12;M is called w12kn. Here w and k stand for week segment and keyword.
Applied Economics Quarterly 55 (2009) 2
Google Econometrics and Unemployment Forecasting
111
with high-skilled workers reacting to fears of layoffs and companies preparing for
layoffs or personnel restructuring. The keyword k 4 (Stepstone or Jobworld or Jobscout or Meinestadt or meine Stadt or Monster Jobs or Monster de or Jobboerse) is
expected to be related to job searching activities, and hence should be associated
with the “flow out of unemployment.”
Figure 2 shows plots of Google activity along the keyword sets above in the first
monthly halves, while Figure 3 exhibits Google activity along the keyword in the
second monthly halves. All indicators seem to follow a somewhat similar seasonal
pattern, while activities measured through variables k1 – k3 move down and activity k 4 moves up. Data collected in weeks 1 and 2 (Figure 2) provide somewhat
different signals than those collected in weeks 3 and 4 (Figure 3). As discussed
above, the special sequence of announcements makes it likely that the more recent
information before an announcement is clouded by the previous announcement.
Furthermore, in terms of using the data for predictions, it is useful to rely on earlier
information because this may allow the analyst to obtain forecasts much faster.
Below we investigate further whether the older search activity data predicts unemployment rates better than the more recent data.
Note: w12kx is variable kx collected in weeks 1 and 2; the variables are k1: Arbeitsamt or Arbeitsagentur, k 2: Arbeitslosenquote, k 3: Personalberater or Personalberatung, and k 4: Stepstone or Jobworld or
Jobscout or Meinestadt or meine Stadt or Monster Jobs or Monster de or Jobboerse.
Figure 2
Applied Economics Quarterly 55 (2009) 2
112
Nikolaos Askitas and Klaus F. Zimmermann
Note: w34kx is variable kx collected in weeks 1 and 2; the variables are k1: Arbeitsamt or Arbeitsagentur, k 2: Arbeitslosenquote, k 3: Personalberater or Personalberatung, and k 4: Stepstone or Jobworld or
Jobscout or Meinestadt or meine Stadt or Monster Jobs or Monster de or Jobboerse.
Figure 3
We close this section with some comments on the Google access data and its
usefulness. Expecting that search engine data on keyword searching contains information which correlates with people’s lives is a natural and, we believe, commonly
accepted expectation. In fact, provided we are able to weed out the noisy activity
and get to the signal in any kind of effective way, this approach may be thought
of as an indirect form of anonymous interviewing resulting in a noisy aggregate
time-series data set. It is not surprising that the study of Google search activity
contains a large portion of the general search engine activity. As of December
2008, Google’s share of the search engine market was close to 63 %, with Yahoo
being a distant second at about 17 %, MSN third at about 10 %, followed by AOL
search at 4 % and ask.com at 2 %.4
Most people use Google not just as a search engine but also as a directory of
their sites of interest. It is quite common for someone to first google a familiar
website and click on the appropriate URL, rather than enter the required site in
the address bar. Consequently not only does search activity contain residual infor4 See for example the Nielsen Online December 2008 Search Engine Share Rankings,
http: // searchenginewatch.com / 3632382.
Applied Economics Quarterly 55 (2009) 2
Google Econometrics and Unemployment Forecasting
113
mation on the Google user but it also contains information on the sites the Google
user intends to visit.
Lastly, we need to discuss issues of keyword choice. In constructing the “search
for a job” keyword set k 4, for example, we had to ascertain what kind of online job
directory services there were. The keyword set which defines k 4 is not constant
over time: sites may come in and out of existence etc. The concept under study
captured by the choice of keywords may depend on linguistic developments, generational parameters, social and economic levels and a host of other factors. It is
therefore important to use keywords which remain constant during the period
observed. We tried a wide range of other keyword families capturing such concepts as consumption, retail activity and online dating, but we restricted ourselves
to k1, k 2, k 3 and k4, as they seem to be sufficient in order to model the process of
unemployment we aim to investigate.
3. Empirical Results
To investigate the usefulness of the Google search activity data for predicting
real economic behavior, we employ a time-series causality approach using the wellknown error-correction model specification (Engle and Granger, 1987; Greene,
2008). This approach implies that the change of the variable of interest is regressed
on its past level, the change of the explanatory variables of interest, and their past
levels. The real data variable to explain used here is the seasonally unadjusted
monthly unemployment rate of Germany5 from January 2004 to April 2009. This
particular time-frame has been enforced by the availability of the Google query
data and the latest available data point at the time when this investigation had to be
carried out. To calculate the change of the variables used, a 12 month lag operator
is used; consequently, the past stock variables are of lag 12. This has the advantage
that we do not need to model seasonality explicitly.
Given the severe economic crisis and the sudden strong decline in economic
activity, the unemployment variable is currently of particular interest to the general
public and for scientists. A surprisingly long continual decline in unemployment
rates in the first quarters of the German recession until December 2004 were observed, which was mainly driven by a long period of economic boom in connection
5 Collecting the information from the Federal Employment Agency on the internet was a
bit cumbersome. The current monthly report is posted in PDF format under: http: // www.pub.
arbeitsagentur.de / hst / services / statistik / 000000 / html / start / monat / aktuell.pdf. In order to
collect the monthly data we needed to download and parse all pdf documents. The authors
believe that the data posted by the Federal Employment Agency would be more complete if it
included “machine actionable” data streams (in SDMX standard for example) in addition to
PDF reports for historical data. The work done at the European Central Bank in that direction
(http: // www.ecb.eu / stats / services / sdmx / html / index.en.html) is a good example of such a
service.
Applied Economics Quarterly 55 (2009) 2
114
Nikolaos Askitas and Klaus F. Zimmermann
with the significant and effective labor market reforms undertaken in the previous
years. The economic decline, however, became suddenly very pronounced in the
fourth quarter of 2008, and in specific economic sectors: namely the export oriented high-quality investment goods industries. It resulted in a labor policy measure which sought to encourage government supported short-time working and was
accompanied by a strong PR campaign by the Federal Employment Agency. The
period of short-time working was increased from previously 6 months to first 18
and finally 24 months. The short-time working allowance increased: Employers do
only have to pay half of the normal social security contributions for short-time
workers, and even nothing if short-time workers engage in further education. Also,
access to short-time working has been improved. This all resulted in strong incentives to retain staff, encouraged further education, and lead to a reduction of a possible loss of income by employees. Companies adopted the policy at unprecedented
levels, contributing to the only moderate increase in unemployment in early 2009.
In this environment, unemployment predictions are very difficult even in the shortterm, and a soft approach using internet activity data might be even more warranted. We want to evaluate its potential here.
Tables 1 to 4 contain the estimated error correction models for two and more regressors capturing the effects of weeks 1 and 2 of the current month (see Tables 1
and 2) and weeks 3 and 4 of the previous month (see Tables 3 and 4). The regressors used are the four variables k1 (Arbeitsamt or Arbeitsagentur), k 2 (Arbeitslosenquote), k 3 (Personalberater or Personalberatung), and k4 (Stepstone or Jobworld or Jobscout or Meinestadt or meine Stadt or Monster Jobs or Monster de or
Jobboerse). The estimates are created in a systematic way and presented in the
tables together with coefficients, their t-ratios, and information criteria (R2 , loglikelihood values, AIC, and BIC). We will base our judgement of the statistical
performance of a model on the BIC; the other measures are for comparison only.
The correct choice of model has to be seen in the context of parsimony, prediction success, usefulness, and sound economic basis. The economic variables included should have short- and long-run effects in line with economic intuition, and
there should be a long-run stationary solution of the model. The statistical model
should be parsimonious, and therefore we want to use as few explanatory variables
as possible. This is typically investigated with an information criteria like the BIC
or the AIC. The approach is useful if it employs regressors that are available early,
and hence enables early forecasts. Finally, prediction success can only be judged in
practice after the model has been used a number of times ex ante.
Our findings based on the BIC suggest that using the earlier data of weeks 3 and
4 of the previous month is statistically acceptable. This makes the Google activity
data even more useful, since one gains in practice two weeks for prediction purposes due to their earlier availability. We also find that a more parsimonious specification is justified, since using the BIC the models including k1 and k 4 only are
doing best in comparison to other or more complex specifications; the BIC also
chooses the model using data from weeks 3 and 4 of the previous month against
Applied Economics Quarterly 55 (2009) 2
Google Econometrics and Unemployment Forecasting
115
the data from weeks 1 and 2 of the current month. Therefore, the model of the third
column of Table 3 is the best, based on statistical grounds. The lagged level variable of unemployment has a negative sign and is significant, and hence there is a
stable long-run solution. k1 measuring the process of contacting the unemployment
office have a positive and statistically significant impact on unemployment in the
short- and long-run. Jobsearch activities measured by k4 predict a strong and significant decline in unemployment in the short-term, but somewhat less strong and
significant in the long-run.
Table 1
Models with two variables involving activity in weeks 1, 2
L12.srates
L12.w12k1
S12.w12k1
L12.w12k 2
S12.w12k 2
w12k1_2
b=t
w12k1_3
b=t
w12k1_4
b=t
w12k 2_3
b=t
w12k 2_4
b=t
w12k 3_4
b=t
–0.387***
(–4.41)
0.037***
(4.50)
–0.010
(–0.92)
0.014***
(3.58)
0.011*
(2.07)
–0.475***
(–5.68)
0.011
(1.09)
–0.019
(–1.65)
–0.310***
(–4.60)
0.021***
(4.25)
–0.000
(–0.06)
–0.611***
(–7.30)
–0.256**
(-3.00)
–0.350***
(–4.51)
0.007
(1.62)
0.007
(1.24)
0.036***
(7.71)
0.013
(1.76)
0.006
(1.79)
0.003
(0.65)
L12.w12k 3
0.028***
(4.86)
0.016*
(2.12)
S12.w12k 3
L12.w12k 4
–2.968***
(–3.69)
–0.686
(–0.74)
–0.028***
(–8.09)
–0.013**
(–2.84)
4.121***
(4.57)
52
94.191
105.899
–41.096
0.756
52
84.961
96.669
–36.481
0.795
52
45.895
57.603
–16.948
0.904
S12.w12k 4
_cons
N
AIC
BIC
Log Lik.
R2
0.522
(0.77)
–0.036***
(–10.92)
–0.020***
(–3.91)
6.942***
(8.00)
0.016***
(3.82)
0.004
(0.86)
–0.029***
(–7.91)
–0.018***
(–3.91)
5.435***
(6.17)
52
95.708
107.415
–41.854
0.749
52
66.811
78.518
–27.405
0.856
52
53.245
64.952
–20.622
0.889
Ln and Sn are the nth monthly lag and difference operators respectively. The variable naming convention is as follows: w12 = first monthly half, w34 = second monthly half; k1, k 2, k 3, k 4 are the keywords
defined in Section 2. A model which is denoted by eg w12ki_j is one involving the two activity variables
in the first monthly halves i and j whereas w34ki_j_l is a model with 3 keywords i, j and l in the second
monthly halves. The variable srates is the seasonal unemployment rates. Finally the significance stars
mean: * p < 0.05; ** p < 0.01; *** p < 0.001.
Applied Economics Quarterly 55 (2009) 2
116
Nikolaos Askitas and Klaus F. Zimmermann
Table 2
Models with more than two variables involving activity in weeks 1, 2
L12.srates
L12.w12k1
S12.w12k1
L12.w12k 2
S12.w12k 2
L12.w12k 3
S12.w12k 3
w12k1_2_3
b=t
w12k1_2_4
b=t
w12k 2_3_4
b=t
w12k1_3_4
b=t
w12k1_2_3_4
b=t
–0.490***
(–5.72)
0.014
(1.32)
–0.016
(–1.38)
0.007
(1.82)
0.007
(1.45)
0.022**
(3.42)
0.011
(1.29)
–0.371***
(–5.17)
0.023***
(4.45)
0.004
(0.54)
0.005
(1.87)
–0.000
(–0.15)
–0.359***
(–4.46)
–0.380***
(–5.38)
0.023**
(3.19)
0.008
(0.95)
L12.w12k 4
_cons
–0.964
(–1.03)
–0.026***
(–7.52)
–0.012**
(–2.75)
3.723***
(4.09)
N
AIC
BIC
Log Lik.
R2
52
84.662
100.272
–34.331
0.812
52
44.280
59.890
–14.140
0.913
S12.w12k4
0.003
(0.93)
0.002
(0.57)
0.015**
(3.36)
0.003
(0.62)
–0.028***
(–7.50)
–0.018***
(–3.80)
5.251***
(5.72)
0.003
(0.59)
–0.007
(–1.13)
–0.027***
(–7.33)
–0.013**
(–2.88)
4.068***
(4.64)
–0.411***
(–5.64)
0.026**
(3.51)
0.011
(1.28)
0.005
(1.70)
0.001
(0.32)
–0.000
(–0.01)
–0.009
(–1.38)
–0.026***
(–7.13)
–0.012**
(–2.74)
3.750***
(4.21)
52
56.208
71.818
–20.104
0.891
52
43.190
58.800
–13.595
0.915
52
43.395
62.908
–11.698
0.921
Ln and Sn are the nth monthly lag and difference operators respectively. The variable naming convention is as follows: w12 = first monthly half, w34 = second monthly half; k1, k 2, k 3, k 4 are the keywords
defined in Section 2. A model which is denoted by eg w12ki_j is one involving the two activity variables
in the first monthly halves i and j whereas w34ki_j_l is a model with 3 keywords i, j and l in the second
monthly halves. The variable srates is the seasonal unemployment rates. Finally the significance stars
mean: * p < 0.05; ** p < 0.01; *** p < 0.001.
Forecasts and realizations of the unemployment rate are shown in Figure 4, and
move together quite well. In a few events the forecasts indicate much earlier that
there is a change in trend; for instance, the predictions for October to December
2008 were conservative, and they anticipated the turning point to the rise in unemployment early on. However, after a perfect fit in January, the two curves split increasingly in the sequel. Our understanding is that this is a result of a change in labor
policy which was announced only during December 2008 and came into effect in
January 2009 concerning the role of government supported short-time working
Applied Economics Quarterly 55 (2009) 2
Google Econometrics and Unemployment Forecasting
117
Table 3
Models with two variables involving activity in weeks 3, 4
w34k1_2
b=t
–0.156
(–1.95)
L13.w34k1
0.000
(0.04)
LS12.w34k1 –0.045***
(–4.80)
L13.w34k 2
0.014**
(2.95)
LS12.w34k 2 0.032***
(4.71)
L13.w34k 3
L12.srates
LS12.w34k 3
w34k1_3
b=t
w34k1_4
b=t
w34k 2_3
b=t
w34k 2_4
b=t
w34k 3_4
b=t
–0.378***
(–4.01)
–0.021*
(–2.11)
–0.039***
(-4.08)
–0.406***
(–6.65)
0.024***
(6.65)
0.016*
(2.34)
–0.414***
(–4.51)
–0.429***
(-4.95)
–0.455***
(-6.81)
0.004
(0.75)
0.021*
(2.48)
0.020***
(3.58)
–0.010
(–0.83)
0.012***
(3.66)
0.013*
(2.45)
0.031***
(4.32)
0.017
(1.58)
L13.w34k 4
_cons
–0.918
(–1.06)
1.600
(1.41)
–0.031***
(–11.41)
–0.014**
(–3.46)
5.503***
(6.98)
N
AIC
BIC
Log Lik.
R2
51
102.860
114.451
–45.430
0.692
51
103.445
115.036
–45.722
0.688
51
40.470
52.061
–14.235
0.909
LS12.w34k4
0.482
(0.58)
–0.028***
(–10.72)
–0.010*
(–2.19)
6.911***
(6.61)
0.017***
(6.41)
0.009
(1.50)
–0.025***
(–11.79)
–0.013**
(–3.40)
5.977***
(7.41)
51
112.394
123.985
–50.197
0.628
51
72.194
83.785
–30.097
0.831
51
46.150
57.741
–17.075
0.899
Ln and Sn are the nth monthly lag and difference operators respectively. The variable naming convention is as follows: w12 = first monthly half, w34 = second monthly half; k1, k 2, k 3, k 4 are the keywords
defined in Section 2. A model which is denoted by eg w12ki_j is one involving the two activity variables
in the first monthly halves i and j whereas w34ki_j_l is a model with 3 keywords i, j and l in the second
monthly halves. The variable srates is the seasonal unemployment rates. Finally the significance stars
mean: * p < 0.05; ** p < 0.01; *** p < 0.001.
already discussed above. The increased interest in short-time work unmeasured in
our regression models have likely contributed to the predicted decline in unemployment. To examine this hypothesis in an informal way, we have replaced k1 in our
final regression model by the search activity on “Kurzarbeit” (short-time work) and
obtained Figure 5 for evaluation. This graph demonstrates that through this variable
most of the differences between forecasts and realizations disappear. However, the
actual prediction for a decline in future unemployment remains. Please also note that
the policy change has been quite recent, and in May, the German labor minister
Applied Economics Quarterly 55 (2009) 2
118
Nikolaos Askitas and Klaus F. Zimmermann
Table 4
Models with more than two variables involving activity in weeks 3, 4
L12.srates
L13.w34k1
LS12.w34k1
L13.w34k 2
LS12.w34k 2
L13.w34k 3
LS12.w34k 3
w34k1_2_3
b=t
w34k1_2_4
b=t
w34k 2_3_4
b=t
w34k1_3_4
b=t
w34k1_2_3_4
b=t
–0.321***
(–3.81)
–0.017
(–1.91)
–0.043***
(–5.11)
0.008
(1.76)
0.028***
(3.94)
0.018*
(2.38)
–0.004
(–0.37)
–0.391***
(–6.01)
0.022***
(4.84)
0.013
(1.58)
0.002
(0.74)
0.004
(0.93)
–0.432***
(–6.39)
–0.432***
(–6.82)
0.018**
(2.82)
0.014
(1.91)
1.185
(1.19)
–0.030***
(–9.54)
-0.014**
(–3.40)
5.199***
(6.04)
0.003
(0.89)
0.008
(1.65)
0.015***
(4.77)
0.004
(0.63)
–0.024***
(–11.10)
–0.014***
(–3.60)
5.567***
(6.63)
0.005
(0.92)
–0.001
(–0.20)
–0.029***
(–10.08)
–0.014**
(–3.36)
5.589***
(7.10)
–0.412***
(–6.34)
0.016*
(2.46)
0.009
(1.14)
0.002
(0.60)
0.007
(1.32)
0.003
(0.66)
–0.005
(–0.69)
–0.027***
(–8.23)
–0.013**
(–3.10)
5.188***
(6.12)
51
91.379
106.834
–37.690
0.772
51
43.401
58.855
–13.700
0.911
51
46.997
62.452
–15.499
0.905
51
41.494
56.949
–12.747
0.914
51
43.288
62.606
–11.644
0.918
L13.w34k 4
LS12.w34k4
_cons
N
AIC
BIC
Log Lik.
R2
Ln and Sn are the nth monthly lag and difference operators respectively. The variable naming convention is as follows: w12 = first monthly half, w34 = second monthly half; k1, k 2, k 3, k 4 are the keywords
defined in Section 2. A model which is denoted by eg w12ki_j is one involving the two activity variables
in the first monthly halves i and j whereas w34ki_j_l is a model with 3 keywords i, j and l in the second
monthly halves. The variable srates is the seasonal unemployment rates. Finally the significance stars
mean: * p < 0.05; ** p < 0.01; *** p < 0.001.
announced an even larger increase in the duration of short-time work.6 Hence, it is
more difficult to adjust the modeling to a realistic approach at this time; we would
like to wait for more data points to make a realistic effort to do so. What remains
important for the purpose of this paper is that we can demonstrate that the internet
activity data is useful to help predict under complex and fast changing conditions.
6 Originally, workers were only able to receive the program for 6 months. The increase in
the duration of the program in January 2009 was for 18 months, and a further increase to 24
months was decided at the end of April and put into practice on May 1, 2009.
Applied Economics Quarterly 55 (2009) 2
Google Econometrics and Unemployment Forecasting
Figure 4
Figure 5
Applied Economics Quarterly 55 (2009) 2
119
120
Nikolaos Askitas and Klaus F. Zimmermann
4. Conclusions
The internet contains an enormous amount of information which, to our knowledge, classical econometrics has yet to appropriately tap into. Such information
comes timely on a continual basis. It is particularly welcome at times of an economic crisis where the traditional flow of information is too slow to provide a
proper basis for sound economic decisions. To examine this potential, this paper
has examined the use of internet activity data to predict economic behavior under
complex and changing circumstances. Of much interest is when and how, and at
what magnitude unemployment is affected after a long period of strong recovery.
Therefore, we have suggested an innovative new method of using data on internet
activity for that purpose and have demonstrated strong correlations between keyword searches and unemployment rates using monthly German data on a simple
and parsimonious level. This suggests that there is a strong potential for the method used, which needs to be further explored.
References
Constant, A. / Zimmermann, K. F. (2008): Im Angesicht der Krise: US-Präsidentschaftswahlen in transnationaler Sicht, DIW Wochenbericht 44, 688 – 701.
Engle, R. F. / Granger, C. W. J. (1987): Co-Integration and Error Correction: Representation,
Estimation, and Testing, Econometrica 55, 251 – 276.
Ginsberg, J. / Mohebbi, M.H. / Patel, R. S. / Brammer, L. / Smolinski, M. S. / Brilliant, L. (2009):
Detecting Influenza Epidemics using Search Engine Query Data, Nature 457, 1012 – 1014.
Greene, W. H. (2008): Econometric Analysis, 6th Edition, Upper Saddle River: Wharton
School Publishing.
Zimmermann, K. F. (2008): Schadensbegrenzung oder Kapriolen wie im Finanzsektor?, Wirtschaftsdienst 12, 18 – 20.
– (2009): Prognosekrise: Warum weniger manchmal mehr ist, Wirtschaftsdienst 2, 86 – 90.
Applied Economics Quarterly 55 (2009) 2
