Get 20M+ Full-Text Papers For Less Than $1.50/day. Start a 14-Day Trial for You or Your Team.

Learn More →

Cyclicality of labour market search: a new big data approach

Cyclicality of labour market search: a new big data approach This paper exploits big data on online activity from the job exchange of the German Federal Employment Agency and its internal placement-software to generate measures for search activity of employers and job seekers and—as a novel feature—for placement activity of employment agencies. In addition, the average search perimeter in the job seekers’ search profiles can be measured. The data are used to estimate the behaviour of the search and placement activities during the business and labour market cycle and their seasonal patterns. The results show that the search activities of firms and employment agencies are procyclical. By contrast, job seekers’ search intensity shows a counter - cyclical pattern, at least before the COVID-19 crisis. Keywords: Big data, Labour market search intensity, Placement activity, Cyclicality JEL classifications: C55, J61, J64 unemployed and vacancies come together. In many coun- 1 Introduction tries, the public employment service is the central point Standard search and matching theory (e.g. [21]) states of contact for unemployed and firms as it provides sup - that labour market matches are formed using unem- port to form matches between both market sides. How- ployed and vacancies, and an efficiency parameter ever, while there are some studies on search intensity of describing how well unemployed and open positions unemployed [5, 20] and firms [2, 3], there are no empiri- form matches. In this context, the efficiency parameter cal measures of placement intensity of employment covers a range of factors such as mismatch [11, 22] or agencies so far, despite its importance. In Germany, for information and institutions. A key ingredient of match- instance, the relevance of this intermediary showed itself ing efficiency, however, is the behaviour of the relevant during the COVID-19 crisis, when it was impossible for agents: Whether people get into work, for example, employment agencies and job centers to pursue their depends crucially on how intensively unemployed look usual placement activities under corona conditions due for jobs and how much effort employers make when to other priorities such as massively increasing short- trying to fill an open position. This behaviour can be time work. described as search intensity. A further crucial factor of This paper contributes to the literature by measuring matching efficiency is flexibility and the willingness to search intensity using a source of big data that directly make some compromise during the search for a job or a captures online activity: It evaluates how often the job suitable candidate. exchange website of the German Federal Employment Beyond job seekers and employers, a further agent Agency (FEA) and its placement platform have been is present in the labour market: the employment ser- accessed by job seekers and firms for search activities. vice (see e.g. [10]). The activities of the intermediary In this context, “big data” refers to the millions of vis- between the two sides can have an impact on how quickly its per month on the FEA’s online job exchange that are processed and transformed into aggregate search meas- *Correspondence: christian.hutter@iab.de ures in this paper. Furthermore, it takes a closer look Institute for Employment Research (IAB), Nuremberg, Germany © The Author(s) 2021. This article is licensed under a Creative Commons Attribution 4.0 International License, which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if changes were made. The images or other third party material in this article are included in the article’s Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article’s Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http://creat iveco mmons .org/licen ses/by/4.0/. 1 Page 2 of 16 C. Hutter on the average search perimeter from the job seekers’ unemployed with low job opportunities, the agency search profiles which they can submit on the online job might intensify support, especially for difficult cases. exchange platform. The idea is that a higher search radius However, this could then come at the expense of support could signal a higher willingness to make some compro- for “easier” cases, given a limited working time of the mise so that it can serve as a potential proxy for the job employment agents. In the end, the cyclical behaviour of seeker’s flexibility. the search and placement activities is an empirical ques- As a counterpart to search intensity, the notion of tion that will be investigated in this paper. placement intensity of the labour market intermediary is In a comprehensive explorative investigation using introduced. The data at hand allow—for the first time— correlation and regression analysis, the study con- measuring placement activity of employment agencies. trasts search and placement activities with GDP and For this purpose, online activity at VerBIS, the FEA’s labour market tightness. The results show that firms internal placement software, is exploited. With this soft- and employment agencies display pro-cyclical search ware, employment agents perform genuine placement patterns. By contrast, job seekers’ search intensity and activities. For instance, they screen the labour market to search perimeter increases during times of weaker econ- find suitable candidates for job openings (or vice versa) omy and labour market. In general, the cyclical behaviour and generate a placement proposal. To my knowledge, is confirmed also when a broad set of control variables is the placement activities of employment services have not included in the regressions. yet been investigated. Consequently, having a measure of Beyond the cyclical movements, the paper also analy- the central matching process on the labour market is all ses the seasonal patterns of the search and placement the more valuable. activities. They might provide explanations for seasonal A central contribution is to empirically analyse impor- patterns in aggregate unemployment or employment. tant time variation properties of the novel search activity The results show that the seasonal patterns closely align measures. This extends previous literature: While [1], for with the main holiday seasons in winter and (to a less instance, find matching efficiency as a whole to be pro - pronounced extent) in summer, and with the vocational cyclical, this paper aims to investigate the cyclical behav- training cycle. iour of several key factors of matching efficiency: firms’ The paper is structured as follows: The next section and job seekers search intensity, placement intensity, as focuses on how the novel data are obtained. Section  3 well as job seekers’ search perimeter. Several studies ana- discusses theoretical considerations. Section  4 pre- lyse job seekers’ search intensity (e.g. [5, 20]). Moreover, sents the results on cyclicality and seasonality. Section  5 [3] consider recruiting intensity of firms and find procy - focuses on search and placement during the COVID-19 clical recruitment intensity. Employers increase their hir- pandemic. The final section concludes. ing efforts in stronger, and thus tighter, labour markets in order to fill their positions. In contrast, search intensity 2 Measuring search and placement activities of the unemployed is often found to be countercycli- from big data cal. While lower chances to receive a job offer in times Conventional measures of search activity are often based of labour market slackness should discourage search, the on survey data, e.g. from time-use surveys [8, 14, 20]. fact that search must be intensified in downturns to com - With the digitalisation of labour markets, online data pensate for lower job arrival rates works in the opposite such as the presence of individuals in online job search [7, direction [19]. Furthermore, if average characteristics 15] or applications to job postings [4] got into the focus. worsen during economic booms, countercyclicality could However, the job exchange of the FEA and its placement- be reinforced [20]. software VerBIS open up innovative big data sources for While there is at least some ambiguity in the litera- research. This section focuses on how these sources are ture on the cyclical behaviour of search intensity, the exploited for measuring search and placement activities. intermediary’s behaviour in the course of a business The FEA as the central intermediary for the unem - or labour market cycle has not been investigated so far. ployed runs a job exchange website where job seekers One could think of several mechanisms at work here. If (JS) can apply for open positions or offer their workforce, a tighter labour market requires employment agents to and firms (F) can find workers or place job offers. Once more intensively support firms in their search for suit - the job exchange is accessed, server log files are stored able candidates, placement intensity could behave procy- clically. By contrast, the agency might intensify support for unemployed and placement activities in times of economic downturn. Characteristics of the unemployed could play a role, too. If, for example, there are more See https ://jobbo erse.arbei tsage ntur.de. Cyclicality of labour market search: a new big data approach Page 3 of 16 1 in anonymous form. These server log files are then pro - To generate monthly data of online activity, the count- cessed in Netmind, a software that allows accessing the ing days of the FEA’s statistics were used. Hence, the data without publication lag. The data in Netmind pro - way the search activities are measured is consistent with vide valuable information, e.g. about which part of the headline figures of the FEA such as unemployment or job exchange website the user has visited. Thus, they vacancies. A counting day typically is around the mid- allow distinguishing whether the job seekers’ or employ- dle of a month. In order to capture search and place- ers’ area of the job exchange was accessed and hence ment intensities instead of mere accumulated activities, JS F measuring the respective search intensities ( I , I ). the sum of activated visits between two counting days For instance, if the specific URL visited implies that the is divided by either the number of unemployed (in case JS F visitor wanted to look over her job openings or to find of I ), the number of vacancies (in case of I ), or by EA suitable job candidates, the exchange website was most the sum of unemployed and vacancies (in case of I ). likely accessed by an employer. On the other hand, if, for Note that throughout this paper, “search intensity” is instance, the visitor searched for suitable job openings, it used synonymous to activated visits of job seekers per can be assumed that a job seeker accessed the website. unemployed, or activated visits of firms per vacancy. The Throughout the paper, “activated visits” are used, i.e. question whether this search intensity can be explained only online activities where a visitor was active on the by compositional effects rather than actual changes in website beyond merely opening it are counted. Since search behaviour is treated in Sect. 4.3. activated visits involve more than one page view, it can be The three intensity measures are calendar-adjusted, assumed that the visitor is interested in the content and i.e. divided by the number of working days between two took a closer look at it. Thus, activated visits represent counting days, and seasonally adjusted. While Sects.  4.1 the qualified traffic on the online job exchange platform. to  4.4 work with seasonally adjusted data only, Sect.  4.5 Furthermore, this helps exclude unwanted online traffic, treats the seasonal patterns of the novel data in more e.g. by bots, from the data. detail. Occasionally, there are missing data due to In addition to search data of the FEA’s job exchange, changes in the platforms. Potential structural breaks after Netmind also provides access to data of VerBIS, the FEA’s periods of missing data are eliminated by level shift dum- internal placement software. With this software, employ- mies in ARMA models. ment agents (EA) screen the labour supply and demand Throughout this paper, the search and placement sides to identify potential positions for job seekers or measures are calculated counting all relevant activities. suggest candidates for an open position. It is important For instance, all activated visits at the job seekers’ part to know that VerBIS and the FEA’s online job exchange of the job exchange are added up. An alternative would are two different platforms, i.e. the placement procedure be to look at the most important single activities per- is performed by the employment agents using VerBIS formed at the job exchange. So instead of counting all irrespective of how active the job seekers or firms are at online activities of the job seekers at the job exchange, the online job exchange. The placement procedure can for instance, one could count only how often the specific be broken down into single tasks starting with the first URL “search for job offers” or “display job offers” was contact with the unemployed person and ending with accessed. It is possible that such alternative measures are postprocessing after the employment agent has issued subject to different time series dynamics. However, this a placement proposal. All of these single steps must be approach sounds more promising than it eventually turns documented in VerBIS and hence are comprised in out to be: The data show virtually the same dynamics so the aggregate online data at hand. Since administrative no major changes can be expected: The correlation of tasks are also carried out in VerBIS, the information in “search for job offers” and “display job offers” with the job Netmind serves to identify genuine placement activities seekers’ total online activities is r = 0.99 in both cases. (e.g. generating a placement proposal) to measure the This phenomenon is similar when looking at the VerBIS EA placement intensity I . To my knowledge, the place- data. Here, instead of summing up all placement activi- ment activities of employment services have not yet been ties of the employment agents, one could look again only investigated. The counting days of the FEA’s statistics are published here: https ://stati stik. ar b ei t sag e n t ur .de/Nav ig a tion /S t a ti stik/S er v i ce/V er o e ff en t lic h u ng sk a lende r/ It is possible (although not required) for the job seeker to register and log in. Veroeen fft lichu ngska lende r-Nav.html. The log files are stored irrespective of whether a visitor is logged in on the job In addition to job seekers, also firms are supported by employment agen- exchange website or not. cies in their search for suitable candidates. This is why the sum of unem- Multiple visits of a specific URL per day by the same visitor are possible ployed and vacancies is taken for normalisation purposes. The dynamics of EA and counted as such. I do not change substantially if the number of unemployed is used instead. 1 Page 4 of 16 C. Hutter at single URLs (for instance the one connected to the task Table 1 Summary statistics of  search intensities and search perimeter “generating a placement proposal”). However, again the correlation is never less than r = 0.9, which is no surprise JS F EA SP I I I since most of the necessary steps in the placement pro- Mean 1.8089 0.1651 0.8389 44.1953 cess are highly standardized, i.e. regularly performed, and Median 1.7783 0.1664 0.8716 44.2159 must be documented in VerBIS. Maximum 2.5255 0.2490 1.1456 44.7406 On the FEA’s job exchange, a job seeker can also cre- Minimum 0.9472 0.1077 0.4310 43.5653 ate a profile for job searching purposes. Besides informa - Std. dev. 0.3055 0.0261 0.1538 0.3307 tion about the desired job and the applicant’s educational Observations 52 52 45 32 attainments and skills, there is the possibility to enter a search perimeter (in kilometers) or to choose from JS F EA I , I , I : Intensities are measured as activated visits by job seekers / firms / employment agencies per working day per number of unemployed / vacancies default values (e.g. 35 km for apprentices, 50 km for / unemployed+vacancies. SP: Average search perimeter of job seekers in “normal” workers). This way, the job seeker can limit its kilometers search within a certain radius. Data on the search perim- eter reported in the search profiles are available from 2017:5 onwards. Since then, there have been a total of 140 million single entries ranging from 1 to 900 kilome- 2.8 tres. This information is exploited in order to measure the average monthly search perimeter (SP) and to empiri- 2.4 cally check how it changes alongside the economic or labour market cycle. 2.0 From a data quality perspective, the search intensities introduced in this paper have several advantages. They 1.6 are based on big data directly capturing online activity. u Th s, they can build on large samples and do not have 1.2 to rely on survey data or on counting actual applica- tions. Furthermore, they are available without any pub- 0.8 2015 2016 2017 2018 2019 2020 lication lag. This makes them especially valuable during Fig. 1 Search intensity of job seekers. Activated visits on the FEA’s times of big turmoil such as the COVID-19 crisis, while job exchange website by job seekers per working day normalised lagged standard data make it difficult to assess the cur - by number of unemployed. Structural breaks after periods of rent situation of the economy and the labour market. The missing data are eliminated by level shift dummies in ARMA models. novel data open up a new range of possibilities. Policy Seasonally adjusted data makers might use this high frequency measure to adjust specific policies more quickly. It could also help nowcast - ing the economy more accurately, which is important for not be representative for all job seekers since on-the-job researchers in that area but also for the government for searchers could use other search channels, too. How- budget reasons. ever, for an important subgroup of the job seekers (the This being said, the novel data also have their limita - unemployed), the FEA is the (legally required) contact tions. Netmind provides only aggregate data that can- point and a central intermediary. In view of the match- not be merged to individual or administrative data sets. ing function that will be discussed in Sect.  3, the data at It is possible to distinguish between visits and activated hand should be a very good proxy for search intensity of visits and whether the job seekers’ or the firms’ part of the unemployed. Consequently, it can serve as a valuable the online job exchange was accessed. However, Netmind additional ingredient for the matching function (which does not provide access to interesting other information does not model on-the-job search but the part of matches such as the free texts entered by the job seekers in the that stem from an outflow from unemployment). search masks, e.g. on the desired occupation. Further- Figures  1, 2 and  3 depict the development of the more, the job seekers of the online job exchange might resulting monthly search and placement intensi- ties since they are available (2015:11). Table  1 shows the respective summary statistics. All three intensity They are available not via Netmind but through a special data delivery after measures show relevant variation during the sample. a formal request at the FEA’s statistical service. Unfortunately, there is no In the COVID-19 crisis, they experience a dramatic access to data on the free text entered in the search masks with respect to the drop, reflecting the firms’ reluctance to hire and dif- desired occupation or the applicant’s educational attainments and skills due to data protection restrictions. ficulties for employment agencies and job centers in Cyclicality of labour market search: a new big data approach Page 5 of 16 1 0.26 45.5 0.24 45.0 0.22 0.20 44.5 0.18 44.0 0.16 0.14 43.5 0.12 0.10 43.0 2015 2016 2017 2018 2019 2020 2017 2018 2019 2020 Fig. 2 Search intensity of firms. Activated visits on the FEA’s job Fig. 4 Search perimeter of job seekers. Average search perimeter exchange website by firms per working day normalised by number of job seekers’ search profiles on the FEA’s job exchange website. of vacancies. Structural breaks after periods of missing data are Structural breaks after periods of missing data are eliminated by level eliminated by level shift dummies in ARMA models. Seasonally shift dummies in ARMA models. No seasonal pattern was found in adjusted data the data 3 Theoretical considerations Search and matching theory (e.g. [19, 21]) provides 1.2 guidance for what can be expected regarding the cycli- cal behaviour of search activities. It states that vacancies 1.1 (V) and unemployed (U) form matches (H for hirings) 1.0 through a Cobb-Douglas production function. After log- 0.9 linearisation, the matching function reads 0.8 ln(H ) = µ + αln(V ) + (1 − α)ln(U ), t t t−1 t−1 (1) 0.7 where α and (1 − α) are the elasticities of new matches 0.6 with respect to vacancies and unemployed, respec- 0.5 tively, under the assumption of constant returns to scale. 0.4 Matching efficiency µ represents the productivity meas- 2015 2016 2017 2018 2019 2020 ure of this function. It depends on determinants such as Fig. 3 Placement intensity of employment agencies. Activated visits the institutional quality of employment services, search (performing genuine placement activities) on the VerBIS platform by employment agents per working day normalised by the sum intensity, willingness to take up work, or mismatch of unemployed and vacancies. Structural breaks after periods of (compare [3, 12, 16]). Since time variation in matching missing data are eliminated by level shift dummies in ARMA models. efficiency can be substantial (e.g. [12, 23]), matching effi - Seasonally adjusted data ciency is allowed to vary over time. Subtracting ln(U t−1 from both sides of the equation yields ln(jfr ) = µ + αln(θ ), t t t−1 (2) where jfr denotes the job finding rate and θ = V /U pursuing their placement tasks under corona condi- labour market tightness. tions and high priority attached to short-time work. This theoretical framework has several implications Job seekers’ search intensity decreased, too, both due for the cyclical behaviour of search intensity. From the to a decreasing nominator (activated visits) but also firms’ perspective, an upswing is—ceteris paribus—con - due to an increasing denominator (unemployment). nected to decreasing unemployment and hence a lower Figure  4 and the last column of Table  1 show the level of hirings (Eq.  1). However, the firms can react by development and summary statistics of the job seek- posting more vacancies and increasing search effort ers’ search perimeter. It moves within a rather limited (Eq. 2) in order to obtain the same level of hirings. In fact, range of 1.2 km. This is mainly due to the measure there is evidence in the literature supporting this reason- being a monthly average among all entries and needs ing. Davis et al. [3], for instance, find procyclical recruit - to be kept in mind when interpreting the regression ment intensity. Employers increase their hiring efforts in results in Sect. 4.2 with respect to economic relevance. 1 Page 6 of 16 C. Hutter stronger - and thus tighter - labour markets in order to 0.04 -0.9 fill their positions. 0.02 -1.0 From the job seekers’ perspective, applying the same reasoning leads to counter-cyclical search behaviour. A 0.00 -1.1 stronger labour market with lower unemployment levels -0.02 -1.2 leads—ceteris paribus—to higher job offer arrival rates -0.04 -1.3 (see, e.g. [19]), requiring less search effort from the job -0.06 -1.4 seeker to obtain the same level of hirings. In a similar reasoning, DeLoach and Kurt [5] argue that search effort -0.08 -1.5 can be countercyclical because it is intensified in down - -0.10 -1.6 turns to prevent declines in household wealth. A procy- 2015 2016 2017 2018 2019 2020 clical search effort on the firms’ side and countercyclical logofGDP,linearlydet rended (leftscale) search intensity on the job seekers’ side is in line with the logoflabour market tightness(right scale) compensation argument: Both market sides can com- Fig. 5 GDP and labour market tightness. GDP: Imputed data [6] using pensate for a tighter (firms) or worse (job seekers) labour quarterly GDP as anchor variable and monthly industrial production market via increased search efforts. This would also apply as auxiliary variable. Source: Federal Statistical Office. Labour market tightness: Number of vacancies divided by number of unemployed. for the search perimeter since job seekers’s willingness Source: Federal Employment Agency. Seasonally adjusted data to compromise probably increases as the labour market situation worsens. However, one could also think of a mechanism where lower chances to receive a job offer in the fact that the latter usually lags the former, Klinger times of labour market slackness can discourage search and Weber [13], for instance, document a sizeable [5]. Then, job seekers’ search intensity (and probably also decoupling between business cycle and the labour mar- search perimeter) would be procyclical if the discourage- ket in Germany, especially so during the last decade. To ment mechanism dominates. In this context, it is possible cover both, gross domestic product (GDP) and labour that the perceived severity and permanence of a down- market tightness ( θ ), defined as number of vacancies turn could influence the relative importance of the two divided by the number of unemployed, are used. Cal- competing channels. If a crisis is not being perceived as endar- and seasonally adjusted GDP (index: 2015=100) transitory, it could well be that the discouragement effect was taken from the Federal Statistical Office (FSO). It is dominates the compensation effect. only available at a quarterly frequency, which is why the The behaviour of the third actor, the employment imputation algorithm by Denton [6] was implemented agency, can also vary over time, although it is not clear a using industrial production (also from the FSO) as aux- priori whether it follows the cycle of the demand or sup- iliary variable in order to generate a monthly GDP time ply side, or none at all. One could think of procyclical series. Figure  5 shows that there was a rather stable placement behaviour if a tighter labour market requires upswing in GDP until the end of 2017 before the down- the agency to more intensively support firms in their turn in 2018 and 2019. Labour market tightness grew search for suitable candidates. By contrast, the agency strongly until the end of 2017, followed by a period of might intensify support for unemployed if the economic slower growth until mid of 2019, after which it steadily conditions worsen. It is conceivable, for example, that the decreased. Both variables dropped severely due to the job search could be made more comprehensive beyond COVID-19 crisis. All in all, both variables experienced the standard or that the placement officer to job seeker upswing and downturn during the sample even before ratio could be improved. Hence, the cyclical behaviour of COVID-19, which allows to investigate the cyclical pat- the search and placement activities is an empirical ques- terns of the search and placement intensities. tion that will be answered in the following. A visual impression of the behaviour of the search and placement intensities in the course of a business 4 Cyclicality and seasonality of search or labour market cycle can be obtained by contrast- and placement activities ing them with GDP and tightness in scatter plots. This 4.1 S catter plots on cyclicality provides a first idea about the pro- or countercycli - To get a first impression, this subsection analyses the cal nature of the search activities on the labour mar- cyclical behaviour of search and placement intensity via ket. Figures  6 and  7 show the pairwise relationships scatter plots. for GDP and labour market tightness, respectively. First, it needs to be addressed whether “cycle” means All scatter plots contain data until 2020:3 and hence the business cycle or the labour market cycle. Beyond exclude the extreme months of the COVID-19 crisis. Cyclicality of labour market search: a new big data approach Page 7 of 16 1 2.6 0.26 2.4 0.24 2.2 0.22 2.0 0.20 1.8 0.18 1.6 0.16 1.4 0.14 1.2 0.12 -2 -1 0 1 2 3 -2 -1 0 1 2 3 Gross domestic product Gross domestic product 1.2 44.8 1.1 44.6 1.0 44.4 0.9 44.2 0.8 44.0 0.7 43.8 0.6 43.6 0.5 43.4 -2 -1 0 1 2 3 -2 -1 0 1 2 3 Gross domestic product Gross domestic product Fig. 6 Scatter plots of search intensities and GDP. Gross domestic product: Linearly detrended data (i.e. a linear trend was subtracted from the original data). Search activity variables: The lags of GDP yielding the strongest correlation in the regressions of Table 2 were used in the scatter plots: JS F EA I and GDP , I and GDP , I and GDP , SP and GDP . The red lines visualise the fitted linear relationship stemming from OLS regressions t−1 t−3 t−2 t t−3 t t t 2.6 0.26 0.24 2.4 0.22 2.2 0.20 2.0 0.18 1.8 0.16 1.6 0.14 1.4 0.12 1.2 0.10 0.20 0.24 0.28 0.32 0.36 0.20 0.24 0.28 0.32 0.36 0.40 Labourmarkettightness Labourmarkettightness 44.8 1.2 44.6 44.4 1.0 44.2 0.8 44.0 0.6 43.8 43.6 0.4 43.4 0.20 0.24 0.28 0.32 0.36 0.40 0.20 0.24 0.28 0.32 0.36 Labourmarkettightness Labourmarkettightness Fig. 7 Scatter plots of search intensities and labour market tightness. Labour market tightness: Number of vacancies divided by number of JS F EA unemployed. The lags of θ yielding the strongest correlation in the regressions of Table 3 were used in the scatter plots: I and θ , I and θ , I and t t−1 t t t θ , SP and θ t−2 t t−1 Placement intens ity Job seekers’ search intensity Placement intensity Jobseekers’ search intensity Search perimeter Firms’ search intensity Search perimeter Firms’ search intensity 1 Page 8 of 16 C. Hutter 0.4 0.12 0.2 0.08 0.0 0.04 -0.2 0.00 -0.4 -0.04 -0.6 -0.08 -0.8 -0.12 -0.02 0.00 0.02 0.04 0.06 -0.02 0.00 0.02 0.04 0.06 Grossdomesticproduct Grossdomesticproduct 0.3 1.00 0.2 0.75 0.1 0.50 0.0 0.25 -0.1 0.00 -0.2 -0.25 -0.3 -0.50 -0.4 -0.75 -0.5 -0.6 -1.00 -0.02 0.00 0.02 0.04 0.06 -0.02 0.00 0.02 0.04 0.06 Grossdomesticproduct Grossdomesticproduct JS F EA Fig. 8 Scatter plots of search intensities and GDP—annual differences. The notes of Table 6 apply, except that annual differences of GDP, I , I , I , and SP have been used Both the placement intensity of employment agencies cycle variables lets the data speak about whether search and the firms’ search intensity seem to move pro-cycli - activity reacts immediately or after a delay to cyclical cally. The opposite holds for the labour supply side: Job movements. For each possible combination of search seekers’ search intensity and their search perimeter show activity and cycle variable, the following regression signs of counter-cyclical movements. equation is estimated: To the extent the variables are subject to persistence, there is a risk that the respective scatter plots might dis- ln(y ) = α + β ln(x ) + γ t + ǫ , t i t−i t (3) play pro- or countercyclicality due to spurious correla- i=0 tions. However, the following scatter plots based on the JS F EA where y ∈ (I , I , I , SP) , x ∈ (GDP, θ) , p is the lag annual differences of the variables show that this risk does length, and ǫ is a normally-distributed error term. Equa- not materialize here. Differencing, i.e. subtracting lagged tion (3) controls for a linear trend (γ t ). This way, β to β values, removes potential stochastic as well as deter- 0 p capture the cyclical effects and do not pick up correlation ministic trends from the time series and hence is a suit- stemming from linear trends in the variables. As in the able method to deal with persistent time series. Figures 8 scatter plots, the estimation period does not cover the and 9 show that in all cases, the scatter plots look similar COVID-19 crisis months 2020:4 and 2020:5 due to their to those of Figures 6 and 7. Hence, the visual impressions extreme values. regarding pro- or countercyclicality are confirmed for A further matter of interest are data properties such both the level and the differenced variables. as the persistence of variables in the given sample period. The scatter plots already gave a first impres - 4.2 Regression analysis sion that this might not be a decisive issue here. Indeed, The scatter plots gave a first impression. The follow - Augmented Dickey-Fuller (ADF) tests confirm that ing steps involve a more formal analysis in which also GDP does not show the persistence usually found in information about significance and lag structure of the longer samples: They find that GDP behaved as trend - cyclical relationships can be obtained. While so far stationary variable during the sample at hand. The it remained unclear how long it takes for the cycle to null hypothesis of GDP having a unit root is rejected materialise in the search activities on the labour mar- at the 1% significance level in an ADF test with a con - ket, regression analysis is suited to address this ques- stant and a deterministic trend as exogenous variables. tion. Including contemporaneous as well as lagged Placementintensity Jobseekers’ search intensity Search perimeter Firms’ search intensity Cyclicality of labour market search: a new big data approach Page 9 of 16 1 0.4 0.12 0.2 0.08 0.0 0.04 -0.2 0.00 -0.4 -0.04 -0.6 -0.08 -0.8 -0.12 -0.06 -0.02 0.02 0.04 0.06 0.08 -0.06 -0.02 0.02 0.04 0.06 0.08 Labourmarkettightness Labour market tightness 0.3 1.00 0.2 0.75 0.1 0.50 0.0 0.25 -0.1 0.00 -0.2 -0.25 -0.3 -0.50 -0.4 -0.75 -0.5 -0.6 -1.00 -0.06 -0.02 0.02 0.04 0.06 0.08 -0.06 -0.02 0.02 0.04 0.06 0.08 Labourmarkettightness Labour market tightness Fig. 9 Scatter plots of search intensities and labour market tightness—annual differences. The notes of Table 7 apply, except that annual differences JS F EA of θ , I , I , I , and SP have been used Table 2 Regression results: search intensities and GDP Dependent variable JS F EA SP I I I Constant 39.23*** (14.05) − 27.31∗ (14.39) − 77.84∗∗ ∗ (13.53) 7.18∗∗ ∗ (0.75) GDP 0.39 (0.81) 0.27 (1.14) 0.15 (1.29) 0.09** (0.04) GDP − 5.55*** (1.08) − 0.63 (2.35) 3.46** (1.74) − 0.13 (0.12) t−1 GDP − 1.65 (1.76) 0.07 (1.53) 7.22* (4.03) − 0.26** (0.11) t−2 GDP − 1.52 (0.99) 5.82*** (0.74) 6.03** (3.03) − 0.43*** (0.14) t−3 trend 0.45 (0.42) − 0.81** (0.33) − 2.69*** (0.45) 0.04** (0.02) R-squared 0.6684 0.1505 0.5104 0.4409 Estimated effects from OLS regressions with heteroscedasticity- and autocorrelation-robust standard errors (in parentheses). *, **, *** denote significance at the 10, JS F EA 5, 1% level, respectively. I , I , I : Search intensity of job seekers/firms/employment agencies. SP: Average search perimeter of job seekers in kilometers. All variables enter the regressions in logarithms Nonetheless, regressions on the annual growth rate of relationships prove to be significant in any of the regres - GDP were conducted as robustness check in Sect. 4.4. sions for at least one lag, which is remarkable given the Equation (3) is estimated using ordinary least squares limited number of observations. (OLS) with heteroscedasticity- and autocorrelation- Although I would recommend not to over-interpret robust standard errors. All variables appear in logs. As the lag structure at this early stage, one result seems evi- baseline, a lag length of p = 3 is chosen to allow delayed dent: On average, the reactions to labour market cycle effects up to a quarter (i.e. three months). A robustness movements are quicker than to business cycle move- check on the lag length is presented in Sect. 4.4. Tables 2 ments. For instance, a tighter labour market materialises and 3 show the results. instantaneously in higher search efforts of firms (second In general, they confirm the visual impression obtained column of Table 3) while it needs a while in case of GDP- F EA by the scatter plots: I and I move pro-cyclically while changes (second column of Table  2). While surprising at JS I and SP move counter-cyclically. Furthermore, these first glance, it could indeed be rational for firms not to Placementintensity Jobseekers’ search intensity Search perimeter Firms’ search intensity 1 Page 10 of 16 C. Hutter Table 3 Regression results: Search activities and tightness Dependent variable JS F EA SP I I I Constant 0.74 (0.51) − 1.41∗∗ ∗ (0.46) 1.38∗∗ ∗ (0.24) 3.76∗∗ ∗ (0.06) θ − 0.67 (1.33) 2.75∗∗ (1.38) 5.29∗∗ ∗ (1.36) 0.22 (0.29) θ − 2.28∗ (1.28) − 0.74 (2.01) − 6.81∗ (3.84) − 0.32∗∗ (0.12) t−1 θ 0.50 (3.07) − 2.19 (3.26) 5.79∗ (2.98) − 0.04 (0.25) t−2 θ 2.32 (3.67) 0.50 (2.74) − 3.20∗∗ (1.58) 0.11 (0.13) t−3 trend 0.94∗∗ (0.41) − 0.08 (0.38) − 0.96∗∗ ∗ (0.16) − 0.01 (0.06) R-squared 0.6050 0.1214 0.6863 0.2784 Estimated effects from OLS regressions with heteroscedasticity- and autocorrelation-robust standard errors (in parentheses). *, **, *** denote significance at the 10, JS F EA 5, 1% level, respectively. I , I , I : Search intensity of job seekers/firms/employment agencies. SP: Average search perimeter of job seekers in kilometers. θ : labour market tightness. All variables enter the regressions in logarithms Table 4 Regression results including both GDP and tightness Dependent variable JS F EA SP I I I Constant 63.08∗∗ ∗ (16.89) 6.22 (25.23) − 5.45 (20.72) 6.66∗∗ ∗ (1.82) GDP − 0.26 (0.66) − 0.48 (0.98) − 1.50 (1.24) 0.10 (0.08) GDP − 6.78∗∗ ∗ (2.10) − 4.47 (3.16) − 0.23 (2.11) − 0.06 (0.20) t−1 GDP − 3.26∗ (1.69) − 1.96 (2.38) 0.80 (3.64) − 0.24∗ (0.13) t−2 GDP − 2.98∗∗ (1.42) 5.21∗∗ ∗ (1.94) 2.44 (2.54) − 0.42∗∗ ∗ (0.14) t−3 θ 0.06 (1.35) 2.66∗ (1.34) 5.49∗∗ ∗ (1.30) 0.10 (0.22) θ − 0.53 (1.51) 1.52 (1.53) − 7.38∗∗ (3.49) − 0.22∗ (0.11) t−1 θ 1.31 (3.49) − 3.21 (3.14) 5.22∗ (3.03) 0.16 (0.26) t−2 θ − 0.19 (3.35) − 0.74 (2.89) − 2.19 (1.85) − 0.05 (0.29) t−3 trend 0.51 (0.58) 0.41 (0.69) − 1.35∗∗ (0.60) 0.04 (0.07) R-squared 0.6893 0.2330 0.7051 0.4636 Estimated effects from OLS regressions with heteroscedasticity- and autocorrelation-robust standard errors (in parentheses). *, **, *** denote significance at the 10, JS F EA 5, 1% level, respectively. I , I , I : Search intensity of job seekers/firms/employment agencies. SP: Average search perimeter of job seekers in kilometers. θ : labour market tightness. All variables enter the regressions in logarithms react immediately to (potentially short-lived) fluctuations placement intensity. Indeed, when estimating the third but instead to wait until an economic upswing or down- column of Table 3 with differenced log of theta instead of turn proves to be stable before making any decisions the level, the signs do not “jump” anymore, and β and β 0 2 with respect to their search behaviour. Furthermore, the are highly significant with estimated effects of 6 to 7%. labour market itself lags the real economy. At first glance, the effects seem to be less relevant in The significant effects are also relevant in size, although case of the search perimeter. They are much smaller, with to different extents: For instance, a positive 1% GDP the sum of the effects ranging between − 0.3 and − 0.5% F JS change increases I and decreases I by 5 to 6%, while it after 3 months (fourth column). However, the standard EA increases I by almost 17% after 3 months. In case of θ , deviation of SP amounts to only 0.33 km (or 0.75% in the effects are about half the size, which is compensated terms of its average). Hence, put into context, the esti- by the fact that during the sample, the variation of tight- mated effects are relevant after all. ness is much higher than that of GDP. Overall, the explanatory power of the trend and cycle With respect to placement intensity, the changing variables varies from 12 to 69%. While the search inten- signs (positive for lags of 0 and 2, negative for lags 1 and sity of firms is obviously influenced a lot by factors 3) indicate that the growth rate of tightness matters for beyond the aggregate business or labour market cycle, This can be seen when putting the absolute effects outside the brackets. Cyclicality of labour market search: a new big data approach Page 11 of 16 1 Table 5 Regression results: controlling for compositional effects Dependent variable JS F EA SP I I I Constant 78.01∗∗ ∗ (27.87) 50.75∗ (28.18) − 8.69 (24.00) 3.55∗∗ ∗ (1.06) GDP − 0.40 (1.01) − 1.16 (1.27) − 1.66 (0.97) 0.12 (0.08) GDP − 7.16∗∗ (2.90) − 5.54 (3.95) 0.47 (2.62) 0.02 (0.12) t−1 GDP − 3.53 (2.25) − 3.48 (2.62) 3.05 (3.65) − 0.02 (0.12) t−2 GDP − 3.31∗ (1.64) 4.09∗ (2.33) 1.46 (2.48) − 0.24∗∗ (0.12) t−3 θ 0.33 (1.54) − 0.99 (1.63) 7.29∗∗ ∗ (2.35) 0.36∗∗ ∗ (0.11) θ 0.02 (1.74) 5.06∗∗ (2.20) − 9.74∗ (4.01) − 0.34∗ (0.18) t−1 θ 0.60 (3.87) − 2.55 (3.59) 5.24 (3.72) 0.00 (0.17) t−2 θ − 0.70 (3.28) − 0.82 (2.86) − 1.63 (3.00) − 0.02 (0.14) t−3 trend − 0.38 (1.02) 0.06 (1.70) − 2.65∗∗ (0.97) 0.14 (0.12) R-squared 0.7668 0.3986 0.7968 0.6271 The notes of Table 4 apply. The control variables comprise the shares of high-education (college degree), low-education (neither vocational training nor high school degree), older (> 55 years of age), younger (≤ 25 years of age), female, and of foreign people among total inflow into unemployment two thirds of the variation in job seekers’ search intensity and hence influence the observed search intensity. For can be explained by the business cycle (the trend is irrel- instance, the composition of the monthly inflow into evant here). unemployment could (partly or fully) explain the cyclical In a more general setting, regression analysis allows patterns of the search activities, not only of job seekers both GDP and θ to play a role. By including both variables but also of firms and placement agents. To account for in the same equation, it is possible to know whether there such potential compositional effects, this subsection adds is a dominant cycle that influences search and placement several control variables to all regression equations. activities. Table 4 shows the results. The set of monthly available variables capturing rele - In case of the job seekers’ search intensity, the results vant characteristics of job seekers comprises the shares of show that the business cycle dominates the labour market high-education (college degree), low-education (neither cycle. The dependence of the firms’ search intensity on vocational training nor high school degree), older (> 55 the cycle variables remains rather unchanged when both years of age), younger (< = 25 years of age), female, and GDP and θ are included in the regression. It is still con- of foreign people among total inflow into unemployment. temporaneous labour market tightness and the 3-month The shares are taken from the FEO’s statistics. Table  5 lag of GDP that exert the procyclical effects. In addition, shows the results. also the size of the estimated effects barely change. In With respect to the job seekers’ search intensity, the case of placement intensity, it is the labour market cycle control variables are able to explain an additional share of that dominates the business cycle. The coefficients of θ the variation in search intensity (The R-squared increases do not change much, which means also the more com- by 8% points). Including them does not lead to decreas- plex lag structure found above remains. Both GDP and θ ing cyclical effects, though. The effects of job seekers’ remain significant factors for the search perimeter of the search intensity with respect to GDP remain basically job seekers even when both are added. Also the lag struc- unchanged. However, statistical significance is slightly ture remains unchanged compared to Tables 2 and 3. weaker when control variables are included. To sum up, for job seekers and firms, the economic In case of the firms’ search intensity, adding the control cycle seems to be the more relevant factor, while employ- variables increases the R-squared by a substantial amount ment agencies are mainly driven by the labour market (almost 17% points). However, the procyclical movement cycle. found so far does not disappear. Both GDP and θ remain significant factors, although the contemporaneous effect 4.3 C ontrolling for compositional effects of θ becomes insignificant and is replaced by the first lag While the results so far reveal cyclical patterns of the instead. search activities, one has to be careful in interpreting Also for the relationship between placement inten- these findings in terms of changing search behaviour. sity and cycle, the control variables do not play a game- Beyond the search behaviour, also the composition of the changing role. Although adding them increases the searchers could change during an upswing or downturn R-squared by approximately 9% points, the estimated 1 Page 12 of 16 C. Hutter cyclical coefficients do not change much. Again, their sta - market sides with increased search on the labour market. tistical significance is slightly weaker when control vari - Job seekers search more and widen their search perim- ables are included. eter in times of weaker economy and labour market while In case of the search perimeter, adding the control firms search more in times of stronger economy and variables increases the R-squared by 16% points. The labour market. Since the employment agencies are the countercyclical behaviour with respect to labour mar- intermediary between both market sides, their cyclical ket tightness disappears although it remains intact with behaviour is ambiguous from a theoretical point of view. respect to the business cycle. The results on the search The empirical results indicate that—like firms’ search perimeter, however, are based on 30 observations only, intensity—placement intensity increases when the labour so estimating 16 parameters might be a challenge that is market becomes tighter. A potential explanation could just a bit too high. Thus, probably a bit more time is war - be that in times of tighter labour markets there are shifts ranted in order to collect longer time series before even- within the FEA away from the placement of unemployed tually answering this question for the search perimeter. towards the employer service (“Arbeitgeberservice”, To sum up, there is some evidence that observable a department of the FEA supporting firms in finding compositional effects are able to explain a part of the employees). development of search and placement activities. How- ever, they do not substantially weaken the findings 4.4 Robustness checks regarding the cyclical behaviour. Of course, it is still pos- To check robustness of the results, richer specifications sible that there are unobservable compositional effects are estimated using 6 instead of 3 lags of GDP or θ to at play. For instance, the relation of unemployed search- allow for delayed effects up to half a year. The choice of ers versus on-the-job searchers could change during an the lag length is supported by statistical tests and infor- upswing or downturn. Similarly, it is possible that job mation criteria. For instance, in no case do the Schwarz seekers and firms change their preferred search channel criterion and the Akaike information criterion recom- during a business cycle, which could potentially change mend using a lag length of more than 6 months. Tests the composition of job seekers or firms using the FEA’s based on the Ljung-Box Q-statistics reveal that the null online job exchange. Although the aggregate search activ- hypothesis of no serial correlation in the residuals is ity data at hand do not allow to look into the details of not rejected at the 5 percent significance level in case of who accesses the websites but only how often the websites p = 3, and even at the 10 percent level in case of p =  6. are accessed, there is little evidence that changing search Table 6 shows the results for p = 6. channel preferences on the firms’ side play a decisive role They confirm the general results of Tables  2 and  3 on during the sample period. According to the job vacancy the cyclical behaviour of the search and placement activi- survey of the Institute for Employment Research (see, e.g. ties. In many cases, also the same lags of the cycle vari- [17]), the FEA’s online job exchange was mentioned as ables are significant—and they are similar in size. For relevant search channel in one third of the cases in every I , the 6th lag of GDP is significant, too, substantially single year since 2015, which does not leave much room increasing its procyclical response. Thus, the smaller for major compositional effects. model with 3 lags can be considered conservative. For EA Furthermore, one could think of other potential con- I , the bigger model prefers more delayed reactions to trol variables. For instance, the reservation wage could GDP-changes, and the sum of the effects is a bit higher be a relevant factor for the job seekers’ search behav- than in the specification with 3 lags. In case of SP , the iour. However, the reservation wage cannot be observed significant effects of the 4th and 6th lag of GDP balance directly but at most be modelled as a function of labour each other out so that the total effect does not change market tightness, the structure of the unemployed (both much. However, the countercyclical reaction of SP to of which are accounted for in the regressions), or of the θ disappears in the specification with 6 lags, a phenom - net replacement rate of the unemployed. With regards to enon already found when control variables were added. the latter, annual OECD data show that the net replace- In a further robustness check, the regressions are per- ment rate basically remained unchanged since 2015 at formed on the annual growth rate of GDP. The pro- or 59%, so again no major effects can be expected here. countercyclical patterns as well as the lag structure do JS Against the background of the theoretical reason- not change substantially, though. In case of I , the big- ing in Sect.  3, the results indicate that job seekers and gest effect is still stemming from the first lag of GDP firms seem to compensate difficulties on their respective growth with an estimated effect of −  4.79. For I , the highest effect of 3.48 is estimated to occur at the third The share never drops below 32% and never exceeds 34%. 9 10 See https ://data.oecd.org/benwa ge/bene fits-in-unemp loyme nt-share -of- However, linking the VerBIS data to other data sets to control for the com- previ ous-incom e.htm. position of the placement agents using VerBIS is not (yet) possible. Cyclicality of labour market search: a new big data approach Page 13 of 16 1 Table 6 Robustness checks on lag length Dependent variable JS F EA SP I I I GDP Constant 26.23∗∗ (12.69) − 46.00∗∗ ∗ (12.98) − 95.16∗∗ ∗ (5.04) 8.00∗∗ ∗ (0.93) GDP 0.01 (0.61) − 0.45 (0.98) − 1.12 (1.23) 0.07∗∗ (0.03) GDP − 6.80∗∗ ∗ (1.69) − 2.52 (1.89) − 0.85 (3.19) − 0.16 (0.15) t−1 GDP − 2.51∗∗ (1.02) − 1.18 (1.43) 3.51 (3.31) − 0.34∗∗ (0.13) t−2 GDP − 1.85 (1.72) 5.40∗∗ ∗ (1.16) 5.19∗ (3.07) − 0.48∗∗ ∗ (0.11) t−3 GDP 2.40 (2.63) 0.86 (1.01) 4.53∗∗ (2.19) − 0.24∗ (0.13) t−4 GDP 0.18 (1.35) 1.64 (3.26) 3.39∗∗ (1.42) − 0.03 (0.16) t−5 GDP 3.05 (2.08) 5.83∗∗ ∗ (1.80) 5.99∗∗ (2.83) 0.27∗∗ ∗ (0.08) t−6 Trend 0.04 (0.40) − 1.42∗∗ ∗ (0.32) − 3.42∗∗ ∗ (0.17) 0.04∗ (0.02) Tightness Constant 1.74∗ (0.89) − 0.53 (1.11) 1.99∗∗ ∗ (0.64) 3.87∗∗ ∗ (0.05) θ 0.37 (1.29) 3.64∗∗ (1.43) 6.33∗∗ ∗ (0.84) 0.42∗∗ (0.20) θ − 2.42∗∗ (1.17) − 0.68 (1.60) − 8.04∗∗ (3.09) − 0.39∗ (0.22) t−1 θ − 0.51 (3.51) − 2.91 (3.27) 5.24∗ (2.85) − 0.09 (0.26) t−2 θ − 0.52 (3.50) − 0.38 (3.98) − 6.44∗∗ ∗ (2.24) − 0.25 (0.18) t−3 θ 2.11 (2.30) -1.27 (4.81) 6.59∗∗ ∗ (1.85) 0.01 (0.15) t−4 θ 0.33 (3.33) 0.71 (2.69) − 3.48∗∗ (1.49) 0.13 (0.27) t−5 θ 1.12 (1.09) 1.76 (2.69) 1.23 (2.61) 0.21 (0.20) t−6 Trend − 1.73∗∗ (0.68) − 0.69 (0.82) − 1.49∗∗ (0.61) − 0.11∗∗ ∗ (0.04) Estimated effects from OLS regressions with heteroscedasticity- and autocorrelation-robust standard errors (in parentheses). *, **, *** denote significance at the 10, JS F EA 5, 1% level, respectively. I , I , I : Search intensity of job seekers/firms/employment agencies. SP: Average search perimeter of job seekers in kilometers. θ : labour market tightness EA 0.4 activated visits per unemployed per working day lag of GDP growth. In case of I , the second lag of GDP below the trend-cycle level) and highest in February growth exerts the strongest effect with an estimated (around 0.5 activated visits above the trend-cycle). effect of 4.24. And SP reacts strongest to the third lag of Note that the depicted months are not calendar GDP growth with an effect of − 0.19. All reported effects months but months between the counting days (see are found to be statistically significant. Sect. 2). Thus, a typical January covers the period from Finally, the monthly available index of industrial pro- mid of December to mid of January, a typical February duction was used instead of imputed GDP. However, the F EA goes from mid of January to mid of February, and so results do not change substantially. I and I still move JS forth. Consequently, the period of lowest search inten pro-cyclically while I and SP move counter-cyclically. sity on the job seekers’ side usually covers the holiday season of Christmas and New Year’s Eve. Another, less 4.5 The seasonal patterns of search behaviour pronounced, local minimum in the seasonal pattern is The data on search and placement activities used until visible in August and September, most probably due to now were seasonally adjusted. Beyond the cyclical move- summer vacation. ments, also the usual seasonal behaviour in the search The minima of the firms’ search intensity occur in and placement intensities could be of interest as it might January and September, too. However, the highest provide explanations for seasonal patterns in aggregate efforts can be detected from June to August, most likely unemployment or employment. Figure 10 shows the sea- due to additional efforts to duly recruit apprentices sonal patterns in the novel data. They are obtained by before the vocational training cycle starts. The firms applying the additive X12-ARIMA seasonal adjustment seem to be supported by the intermediary since July is procedure. also the period in which employment agencies under The graphs show how many additional activated take the highest placement efforts. The other maxi - visits usually occur in a specific month, beyond the mum of placement intensity is in February, matching trend-cycle level. For instance, job seekers’ search the maximum of job seekers’ search intensity. And also intensity usually is lowest in January (approximately 1 Page 14 of 16 C. Hutter 0.6 140 0.4 0.2 0.0 -0.2 -0.4 Jobseekers’searchintensity -0.6 2015 2016 2017 2018 2019 2020 M2 M3 M4 M5 M6 M7 0.03 0.02 0.01 Job seekers’ search intensity 0.00 Firms’ search intensity Placement intensity -0.01 -0.02 Fig. 11 Search and placement during the COVID-19 pandemic. Daily -0.03 data using a 7-day moving average to eliminate day-of-the-week effects; holiday-adjusted using ARMA models with dummies; index -0.04 Firms’ search intensity (January 6 2020 = 100) -0.05 2015 2016 2017 2018 2019 2020 0.08 the labour market. Consequently, January typically is the month with the highest unemployment rate in the 0.04 course of a year. 0.00 No seasonal pattern could be detected for the search perimeter (neither significant nor relevant in size). This -0.04 is no surprise given the development shown in Fig. 4. The maxima and minima do not occur at the same months. -0.08 Placementintensity -0.12 2015 2016 2017 2018 2019 2020 5 Search and placement during the COVID‑19 Fig. 10 Seasonal patterns of search and placement intensities. The pandemic seasonal patterns for search and placement intensities stem from So far, the COVID-19 months were excluded in the additive X12-ARIMA seasonal adjustment analysis. This is justified by the extreme disruption this pandemic has caused, as can be seen in Figs.  1, 2, 3, 4 and  5. Including such extreme values in scatter plots or regression analysis would dominate the results and make the minima of placement intensity closely match those it difficult to investigate reactions as they would occur in of the labour demand and supply sides: They occur in normal times. At the same time, the long-term conse- January, September, and November. quences of this crisis are not yet clear. It is possible that Measuring the extent of the seasonal pattern relative long-lasting shifts in the economy and the labour market to the mean search intensity, the seasonality on the job will remain even when the virus is under control. seekers’ side is found to be a bit higher (ranging from However, the data at hand allow for a descriptive analy- − 25 to +  30% of its mean search intensity) than that sis of search and placement in the COVID-19 pandemic. on the firms’ side, ranging from − 25 to +  15% (com- Since during the first lockdown in March 2020 the news pare also [4] who find that seasonality is much bigger situation changed almost on a daily basis, the fact that for applications than postings). the data at hand allow for a daily analysis becomes all the All in all, the seasonal patterns seem to be largely more valuable. Figure 11 shows the daily development of influenced by holiday seasons and the vocational train - the search and placement intensities of all three agents ing cycle. Since the seasonal patterns on both market on the labour market. For scaling purposes, the lines are sides and the employment agencies largely coincide, indexed so that they all start at a value of 100. they most probably reinforce each others’ effects on Cyclicality of labour market search: a new big data approach Page 15 of 16 1 Between March 5th and March 25th, the firms’ search activities bounce back more quickly after the COVID-19 intensity dropped substantially by 42%. Until the end of crisis. the sample (July 13th), it recovered again, reaching 92.5% In the future, further data from the BA job exchange of the pre-crisis level. There was also a sharp (− 35%) and could be exploited for scientific use, especially data on even faster (until March 17th) slump of placement inten- search behaviour. Beyond the search perimeter, other sity, whereas the subsequent recovery proceeded much entries on the job exchange could be made accessible more slowly. Until the end of the observation period, for research. For instance, the FEA plans a change from only 71% of the pre-crisis level had been reached again. Netmind to Matomo, after which information from the Placement intensity declined strongly since employment search masks such as the desired profession will be saved agents had to prioritize the processing of short-time work and could then be accessed for research purposes. Future requests, in the course of which the placement activity projects may make it possible to merge the search activity that would be usual in normal times could be carried out data to administrative data. Then, it could be investigated only on a considerably reduced scale. how characteristics such as qualification, occupation or The job seekers’ search intensity displays a more com - region affect the search duration. It would also be con - plex development which was hidden in the monthly fig - ceivable to analyze the relationship between the duration ures. Within ten days, it initially declined sharply before of unemployment or the period of time a job vacancy strongly recovering and even overshooting pre-crisis-lev- exists and the search intensity. els. However, it quickly declined afterwards to very low Acknowledgements levels and has not recovered since. A possible explana- I thank Hermann Gartner, Tobias Hartl, Ortwin Herbst, Maren Müller, Clemens tion for this development is that an intensified search in Usbeck, Enzo Weber, and participants of the annual meeting of the FEA’s sta- tistics department 2018, the workshop “Genesis of information from online job the first days after the lockdown was not seen as a pri - advertisements” 2019, and the IWH/IAB workshop 2020 for valuable support ority due to extraordinary challenges to cope with the and helpful suggestions. situation, and the search intensity therefore initially sank. Authors’ contributions Subsequently, a compensation effect responsible for the I am the sole contributor to the writing of this article. The author read and usual counter-cyclical search pattern (see Sect. 3) appears approved the final manuscript. to have been dominant until the end of March, while the Funding discouragement effect could have been the dominant Not applicable. driver relatively quickly from April onwards. Availability of data and materials The data on online visits of the job exchange platform of the Federal Employ- 6 Conclusion ment Agency and its internal placement software VerBIS are initially stored This article introduces innovative online data allowing in server log files. In an aggregated form, they can be downloaded via the the instantaneous measurement of search and—for the software Netmind. Access is provided by the Federal Employment Agency. The other aggregate data used in this article stem from the Federal Statistical first time—placement intensity in the labour market in Office (in case of GDP and production index), and from the Statistics Depart - form of online activity. These data are used to estimate ment of the Federal Employment Agency (in case of unemployment, vacan- their behaviour during the economic and labour mar- cies, and the control variables). They can be downloaded from the respective websites. Alternatively, I can provide the data upon request. ket cycle, as well as their usual seasonal patterns. The results show that firms’ and employment agencies’ search Competing interests and placement intensity displays a pro-cyclical pattern I declare that there are no competing interests, neither financial nor non-financial. while—at least before the COVID-19 pandemic—job seekers’ search intensity is counter-cyclical. Received: 2 July 2020 Accepted: 15 December 2020 In the COVID-19 crisis so far, the data reveal that the placement intensity of employment agencies and the firms’ search intensity dropped substantially. Looking at the daily data, the job seekers’ search intensity displays a References 1. Barnichon, R., Figura, A.: Labor market heterogeneity and the aggregate more complex development. Here, the procyclical driv- matching function. Am. Econ. J. 7, 222–249 (2015) ers seem to eventually dominate during the COVID-19 2. Carrillo-Tudela, C., Gartner, H., Kaas, L.: Recruitment policies, job-filling pandemic, contrary to the counter-cyclical pattern found rates and matching efficiency. IAB-Discussion Paper 15/2020, 51 pages (2020) for normal times. The overall drop in search and place - 3. Davis, S.J., Faberman, R.J., Haltiwanger, J.C.: The establishment-level ment activities naturally will have adverse impacts on behavior of vacancies and hiring. Q. J. Econ. 128(2), 581–622 (2013) the labour market, especially on the job finding rate [9]. 4. Davis, S. J., de la Parra, B. S., 2018, Application Flows (online presenta- tion) https ://stati c1.squar espac e.com/stati c/5e2ea 3a809 7ed30 c779b The most critical labour market effects of the crisis may d707/t/5e315 71592 58077 0732d ac3b/15802 91866 530/Appli catio arise not via the separation but via the hiring margin [18]. n+Flows +Febru ary+2018.pdf It is therefore important that the search and placement 1 Page 16 of 16 C. Hutter 5. DeLoach, S., Kurt, M.: Discouraging workers: estimating the impacts of 15. Kuhn, P., Skuterud, M.: Internet job search and unemployment durations. macroeconomic shocks on the search intensity of the unemployed. J. Am. Econ. Rev. 94(1), 218–232 (2004) Labor Res. 34, 433–454 (2013) 16. Launov, A., Wälde, K.: The employment effect of reforming a public 6. Denton, F.T.: Adjustment of monthly or quarterly series to annual totals: employment agency. Eur. Econ. Rev. 84, 140–164 (2016) an approach based on quadratic minimization. J. Am. Stat. Soc. 66(333), 17. Mario, B., Nicole, G.A.: The IAB job vacancy survey: design and research 99–102 (1971) potential. J. Labour Market Res. 12, 54 (2020) 7. Faberman, R.J., Kudlyak, M.: The intensity of job search and search dura- 18. Merkl, C., Weber, E.: Rescuing the labour market in times of COVID-19: tion. Am. Econ. J. 11, 327–357 (2019) don’t forget new hires. VoxEU 07 April (2020) 8. Gomme, P., Lkhagvasuren, D.: Worker search effort as an amplification 19. Mortensen, D.T.: Job search and labor market analysis. In: Ashenfelter, mechanism. J. Monet. Econ. 75, 106–122 (2015) O., Layard, R. (eds.) Handbook of Labor Economics, vol. 2, pp. 849–919. 9. Hartl, T., Hutter, C., Weber, E.: Neueinstellungen in der Krise. Makronom, Elsevier, New York (1987) https ://makro nom.de/coron a-arbei tsmar kt-auswi rkung en-neuei nstel 20. Mukoyama, T., Patterson, C., Sahin, A.: Job search behavior over the busi- lunge n-in-der-krise -36315 (2020) ness cycle. Staff Reports 689, Federal Reserve Bank of New York (2014) 10. Holzner, C., Watanabe, M.: Do intermediaries reduce search frictions? 21. Pissarides, C.: Equilibrium unemployment theory, 2nd edn. The MIT Press, https ://www.pe.econ.uni-muenc hen.de/perso nen/postd ocs/holzn er/ New York (2000) forsc hung/pea_searc hfric tions 01.pdf (2019) 22. Sahin, A., Song, J., Topa, G., Violante, G.L.: Mismatch unemployment. Am. 11. Hutter, C., Weber, E.: Mismatch and the forecasting performance of Econ. Rev. 104(11), 3529–3564 (2014) matching functions. Oxford Bull. Econ. Stat. 79(1), 101–123 (2017) 23. Sedlacek, P.: Match efficiency and firms’ hiring standards. J. Monet. Econ. 12. Klinger, S., Weber, E.: Decomposing beveridge curve dynamics by cor- 62, 123–133 (2014) related unobserved components. Oxford Bull. Econ. Stat. 78, 877–894 (2016) Publisher’s Note 13. Klinger, S., Weber, E.: GDP-employment decoupling in Germany. Struct. Springer Nature remains neutral with regard to jurisdictional claims in pub- Change Econ. Dynam. 52, 82–98 (2020) lished maps and institutional affiliations. 14. Krueger, A.B., Mueller, A.: Job search, emotional well-being and job find- ing in a period of mass unemployment: evidence from high-frequency longitudinal data. Brook. Papers Econ. Activ. 42, 1–81 (2011) http://www.deepdyve.com/assets/images/DeepDyve-Logo-lg.png Journal for Labour Market Research Springer Journals

Cyclicality of labour market search: a new big data approach

Hutter, Christian
Journal for Labour Market Research , Volume 55 (1) – Jan 23, 2021
Download PDF
16 pages

Loading next page...
 
/lp/springer-journals/cyclicality-of-labour-market-search-a-new-big-data-approach-8PwhLQcs69

References (24)

Publisher
Springer Journals
Copyright
Copyright © The Author(s) 2021
ISSN
1614-3485
eISSN
2510-5027
DOI
10.1186/s12651-020-00283-9
Publisher site
See Article on Publisher Site

Abstract

This paper exploits big data on online activity from the job exchange of the German Federal Employment Agency and its internal placement-software to generate measures for search activity of employers and job seekers and—as a novel feature—for placement activity of employment agencies. In addition, the average search perimeter in the job seekers’ search profiles can be measured. The data are used to estimate the behaviour of the search and placement activities during the business and labour market cycle and their seasonal patterns. The results show that the search activities of firms and employment agencies are procyclical. By contrast, job seekers’ search intensity shows a counter - cyclical pattern, at least before the COVID-19 crisis. Keywords: Big data, Labour market search intensity, Placement activity, Cyclicality JEL classifications: C55, J61, J64 unemployed and vacancies come together. In many coun- 1 Introduction tries, the public employment service is the central point Standard search and matching theory (e.g. [21]) states of contact for unemployed and firms as it provides sup - that labour market matches are formed using unem- port to form matches between both market sides. How- ployed and vacancies, and an efficiency parameter ever, while there are some studies on search intensity of describing how well unemployed and open positions unemployed [5, 20] and firms [2, 3], there are no empiri- form matches. In this context, the efficiency parameter cal measures of placement intensity of employment covers a range of factors such as mismatch [11, 22] or agencies so far, despite its importance. In Germany, for information and institutions. A key ingredient of match- instance, the relevance of this intermediary showed itself ing efficiency, however, is the behaviour of the relevant during the COVID-19 crisis, when it was impossible for agents: Whether people get into work, for example, employment agencies and job centers to pursue their depends crucially on how intensively unemployed look usual placement activities under corona conditions due for jobs and how much effort employers make when to other priorities such as massively increasing short- trying to fill an open position. This behaviour can be time work. described as search intensity. A further crucial factor of This paper contributes to the literature by measuring matching efficiency is flexibility and the willingness to search intensity using a source of big data that directly make some compromise during the search for a job or a captures online activity: It evaluates how often the job suitable candidate. exchange website of the German Federal Employment Beyond job seekers and employers, a further agent Agency (FEA) and its placement platform have been is present in the labour market: the employment ser- accessed by job seekers and firms for search activities. vice (see e.g. [10]). The activities of the intermediary In this context, “big data” refers to the millions of vis- between the two sides can have an impact on how quickly its per month on the FEA’s online job exchange that are processed and transformed into aggregate search meas- *Correspondence: christian.hutter@iab.de ures in this paper. Furthermore, it takes a closer look Institute for Employment Research (IAB), Nuremberg, Germany © The Author(s) 2021. This article is licensed under a Creative Commons Attribution 4.0 International License, which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if changes were made. The images or other third party material in this article are included in the article’s Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article’s Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http://creat iveco mmons .org/licen ses/by/4.0/. 1 Page 2 of 16 C. Hutter on the average search perimeter from the job seekers’ unemployed with low job opportunities, the agency search profiles which they can submit on the online job might intensify support, especially for difficult cases. exchange platform. The idea is that a higher search radius However, this could then come at the expense of support could signal a higher willingness to make some compro- for “easier” cases, given a limited working time of the mise so that it can serve as a potential proxy for the job employment agents. In the end, the cyclical behaviour of seeker’s flexibility. the search and placement activities is an empirical ques- As a counterpart to search intensity, the notion of tion that will be investigated in this paper. placement intensity of the labour market intermediary is In a comprehensive explorative investigation using introduced. The data at hand allow—for the first time— correlation and regression analysis, the study con- measuring placement activity of employment agencies. trasts search and placement activities with GDP and For this purpose, online activity at VerBIS, the FEA’s labour market tightness. The results show that firms internal placement software, is exploited. With this soft- and employment agencies display pro-cyclical search ware, employment agents perform genuine placement patterns. By contrast, job seekers’ search intensity and activities. For instance, they screen the labour market to search perimeter increases during times of weaker econ- find suitable candidates for job openings (or vice versa) omy and labour market. In general, the cyclical behaviour and generate a placement proposal. To my knowledge, is confirmed also when a broad set of control variables is the placement activities of employment services have not included in the regressions. yet been investigated. Consequently, having a measure of Beyond the cyclical movements, the paper also analy- the central matching process on the labour market is all ses the seasonal patterns of the search and placement the more valuable. activities. They might provide explanations for seasonal A central contribution is to empirically analyse impor- patterns in aggregate unemployment or employment. tant time variation properties of the novel search activity The results show that the seasonal patterns closely align measures. This extends previous literature: While [1], for with the main holiday seasons in winter and (to a less instance, find matching efficiency as a whole to be pro - pronounced extent) in summer, and with the vocational cyclical, this paper aims to investigate the cyclical behav- training cycle. iour of several key factors of matching efficiency: firms’ The paper is structured as follows: The next section and job seekers search intensity, placement intensity, as focuses on how the novel data are obtained. Section  3 well as job seekers’ search perimeter. Several studies ana- discusses theoretical considerations. Section  4 pre- lyse job seekers’ search intensity (e.g. [5, 20]). Moreover, sents the results on cyclicality and seasonality. Section  5 [3] consider recruiting intensity of firms and find procy - focuses on search and placement during the COVID-19 clical recruitment intensity. Employers increase their hir- pandemic. The final section concludes. ing efforts in stronger, and thus tighter, labour markets in order to fill their positions. In contrast, search intensity 2 Measuring search and placement activities of the unemployed is often found to be countercycli- from big data cal. While lower chances to receive a job offer in times Conventional measures of search activity are often based of labour market slackness should discourage search, the on survey data, e.g. from time-use surveys [8, 14, 20]. fact that search must be intensified in downturns to com - With the digitalisation of labour markets, online data pensate for lower job arrival rates works in the opposite such as the presence of individuals in online job search [7, direction [19]. Furthermore, if average characteristics 15] or applications to job postings [4] got into the focus. worsen during economic booms, countercyclicality could However, the job exchange of the FEA and its placement- be reinforced [20]. software VerBIS open up innovative big data sources for While there is at least some ambiguity in the litera- research. This section focuses on how these sources are ture on the cyclical behaviour of search intensity, the exploited for measuring search and placement activities. intermediary’s behaviour in the course of a business The FEA as the central intermediary for the unem - or labour market cycle has not been investigated so far. ployed runs a job exchange website where job seekers One could think of several mechanisms at work here. If (JS) can apply for open positions or offer their workforce, a tighter labour market requires employment agents to and firms (F) can find workers or place job offers. Once more intensively support firms in their search for suit - the job exchange is accessed, server log files are stored able candidates, placement intensity could behave procy- clically. By contrast, the agency might intensify support for unemployed and placement activities in times of economic downturn. Characteristics of the unemployed could play a role, too. If, for example, there are more See https ://jobbo erse.arbei tsage ntur.de. Cyclicality of labour market search: a new big data approach Page 3 of 16 1 in anonymous form. These server log files are then pro - To generate monthly data of online activity, the count- cessed in Netmind, a software that allows accessing the ing days of the FEA’s statistics were used. Hence, the data without publication lag. The data in Netmind pro - way the search activities are measured is consistent with vide valuable information, e.g. about which part of the headline figures of the FEA such as unemployment or job exchange website the user has visited. Thus, they vacancies. A counting day typically is around the mid- allow distinguishing whether the job seekers’ or employ- dle of a month. In order to capture search and place- ers’ area of the job exchange was accessed and hence ment intensities instead of mere accumulated activities, JS F measuring the respective search intensities ( I , I ). the sum of activated visits between two counting days For instance, if the specific URL visited implies that the is divided by either the number of unemployed (in case JS F visitor wanted to look over her job openings or to find of I ), the number of vacancies (in case of I ), or by EA suitable job candidates, the exchange website was most the sum of unemployed and vacancies (in case of I ). likely accessed by an employer. On the other hand, if, for Note that throughout this paper, “search intensity” is instance, the visitor searched for suitable job openings, it used synonymous to activated visits of job seekers per can be assumed that a job seeker accessed the website. unemployed, or activated visits of firms per vacancy. The Throughout the paper, “activated visits” are used, i.e. question whether this search intensity can be explained only online activities where a visitor was active on the by compositional effects rather than actual changes in website beyond merely opening it are counted. Since search behaviour is treated in Sect. 4.3. activated visits involve more than one page view, it can be The three intensity measures are calendar-adjusted, assumed that the visitor is interested in the content and i.e. divided by the number of working days between two took a closer look at it. Thus, activated visits represent counting days, and seasonally adjusted. While Sects.  4.1 the qualified traffic on the online job exchange platform. to  4.4 work with seasonally adjusted data only, Sect.  4.5 Furthermore, this helps exclude unwanted online traffic, treats the seasonal patterns of the novel data in more e.g. by bots, from the data. detail. Occasionally, there are missing data due to In addition to search data of the FEA’s job exchange, changes in the platforms. Potential structural breaks after Netmind also provides access to data of VerBIS, the FEA’s periods of missing data are eliminated by level shift dum- internal placement software. With this software, employ- mies in ARMA models. ment agents (EA) screen the labour supply and demand Throughout this paper, the search and placement sides to identify potential positions for job seekers or measures are calculated counting all relevant activities. suggest candidates for an open position. It is important For instance, all activated visits at the job seekers’ part to know that VerBIS and the FEA’s online job exchange of the job exchange are added up. An alternative would are two different platforms, i.e. the placement procedure be to look at the most important single activities per- is performed by the employment agents using VerBIS formed at the job exchange. So instead of counting all irrespective of how active the job seekers or firms are at online activities of the job seekers at the job exchange, the online job exchange. The placement procedure can for instance, one could count only how often the specific be broken down into single tasks starting with the first URL “search for job offers” or “display job offers” was contact with the unemployed person and ending with accessed. It is possible that such alternative measures are postprocessing after the employment agent has issued subject to different time series dynamics. However, this a placement proposal. All of these single steps must be approach sounds more promising than it eventually turns documented in VerBIS and hence are comprised in out to be: The data show virtually the same dynamics so the aggregate online data at hand. Since administrative no major changes can be expected: The correlation of tasks are also carried out in VerBIS, the information in “search for job offers” and “display job offers” with the job Netmind serves to identify genuine placement activities seekers’ total online activities is r = 0.99 in both cases. (e.g. generating a placement proposal) to measure the This phenomenon is similar when looking at the VerBIS EA placement intensity I . To my knowledge, the place- data. Here, instead of summing up all placement activi- ment activities of employment services have not yet been ties of the employment agents, one could look again only investigated. The counting days of the FEA’s statistics are published here: https ://stati stik. ar b ei t sag e n t ur .de/Nav ig a tion /S t a ti stik/S er v i ce/V er o e ff en t lic h u ng sk a lende r/ It is possible (although not required) for the job seeker to register and log in. Veroeen fft lichu ngska lende r-Nav.html. The log files are stored irrespective of whether a visitor is logged in on the job In addition to job seekers, also firms are supported by employment agen- exchange website or not. cies in their search for suitable candidates. This is why the sum of unem- Multiple visits of a specific URL per day by the same visitor are possible ployed and vacancies is taken for normalisation purposes. The dynamics of EA and counted as such. I do not change substantially if the number of unemployed is used instead. 1 Page 4 of 16 C. Hutter at single URLs (for instance the one connected to the task Table 1 Summary statistics of  search intensities and search perimeter “generating a placement proposal”). However, again the correlation is never less than r = 0.9, which is no surprise JS F EA SP I I I since most of the necessary steps in the placement pro- Mean 1.8089 0.1651 0.8389 44.1953 cess are highly standardized, i.e. regularly performed, and Median 1.7783 0.1664 0.8716 44.2159 must be documented in VerBIS. Maximum 2.5255 0.2490 1.1456 44.7406 On the FEA’s job exchange, a job seeker can also cre- Minimum 0.9472 0.1077 0.4310 43.5653 ate a profile for job searching purposes. Besides informa - Std. dev. 0.3055 0.0261 0.1538 0.3307 tion about the desired job and the applicant’s educational Observations 52 52 45 32 attainments and skills, there is the possibility to enter a search perimeter (in kilometers) or to choose from JS F EA I , I , I : Intensities are measured as activated visits by job seekers / firms / employment agencies per working day per number of unemployed / vacancies default values (e.g. 35 km for apprentices, 50 km for / unemployed+vacancies. SP: Average search perimeter of job seekers in “normal” workers). This way, the job seeker can limit its kilometers search within a certain radius. Data on the search perim- eter reported in the search profiles are available from 2017:5 onwards. Since then, there have been a total of 140 million single entries ranging from 1 to 900 kilome- 2.8 tres. This information is exploited in order to measure the average monthly search perimeter (SP) and to empiri- 2.4 cally check how it changes alongside the economic or labour market cycle. 2.0 From a data quality perspective, the search intensities introduced in this paper have several advantages. They 1.6 are based on big data directly capturing online activity. u Th s, they can build on large samples and do not have 1.2 to rely on survey data or on counting actual applica- tions. Furthermore, they are available without any pub- 0.8 2015 2016 2017 2018 2019 2020 lication lag. This makes them especially valuable during Fig. 1 Search intensity of job seekers. Activated visits on the FEA’s times of big turmoil such as the COVID-19 crisis, while job exchange website by job seekers per working day normalised lagged standard data make it difficult to assess the cur - by number of unemployed. Structural breaks after periods of rent situation of the economy and the labour market. The missing data are eliminated by level shift dummies in ARMA models. novel data open up a new range of possibilities. Policy Seasonally adjusted data makers might use this high frequency measure to adjust specific policies more quickly. It could also help nowcast - ing the economy more accurately, which is important for not be representative for all job seekers since on-the-job researchers in that area but also for the government for searchers could use other search channels, too. How- budget reasons. ever, for an important subgroup of the job seekers (the This being said, the novel data also have their limita - unemployed), the FEA is the (legally required) contact tions. Netmind provides only aggregate data that can- point and a central intermediary. In view of the match- not be merged to individual or administrative data sets. ing function that will be discussed in Sect.  3, the data at It is possible to distinguish between visits and activated hand should be a very good proxy for search intensity of visits and whether the job seekers’ or the firms’ part of the unemployed. Consequently, it can serve as a valuable the online job exchange was accessed. However, Netmind additional ingredient for the matching function (which does not provide access to interesting other information does not model on-the-job search but the part of matches such as the free texts entered by the job seekers in the that stem from an outflow from unemployment). search masks, e.g. on the desired occupation. Further- Figures  1, 2 and  3 depict the development of the more, the job seekers of the online job exchange might resulting monthly search and placement intensi- ties since they are available (2015:11). Table  1 shows the respective summary statistics. All three intensity They are available not via Netmind but through a special data delivery after measures show relevant variation during the sample. a formal request at the FEA’s statistical service. Unfortunately, there is no In the COVID-19 crisis, they experience a dramatic access to data on the free text entered in the search masks with respect to the drop, reflecting the firms’ reluctance to hire and dif- desired occupation or the applicant’s educational attainments and skills due to data protection restrictions. ficulties for employment agencies and job centers in Cyclicality of labour market search: a new big data approach Page 5 of 16 1 0.26 45.5 0.24 45.0 0.22 0.20 44.5 0.18 44.0 0.16 0.14 43.5 0.12 0.10 43.0 2015 2016 2017 2018 2019 2020 2017 2018 2019 2020 Fig. 2 Search intensity of firms. Activated visits on the FEA’s job Fig. 4 Search perimeter of job seekers. Average search perimeter exchange website by firms per working day normalised by number of job seekers’ search profiles on the FEA’s job exchange website. of vacancies. Structural breaks after periods of missing data are Structural breaks after periods of missing data are eliminated by level eliminated by level shift dummies in ARMA models. Seasonally shift dummies in ARMA models. No seasonal pattern was found in adjusted data the data 3 Theoretical considerations Search and matching theory (e.g. [19, 21]) provides 1.2 guidance for what can be expected regarding the cycli- cal behaviour of search activities. It states that vacancies 1.1 (V) and unemployed (U) form matches (H for hirings) 1.0 through a Cobb-Douglas production function. After log- 0.9 linearisation, the matching function reads 0.8 ln(H ) = µ + αln(V ) + (1 − α)ln(U ), t t t−1 t−1 (1) 0.7 where α and (1 − α) are the elasticities of new matches 0.6 with respect to vacancies and unemployed, respec- 0.5 tively, under the assumption of constant returns to scale. 0.4 Matching efficiency µ represents the productivity meas- 2015 2016 2017 2018 2019 2020 ure of this function. It depends on determinants such as Fig. 3 Placement intensity of employment agencies. Activated visits the institutional quality of employment services, search (performing genuine placement activities) on the VerBIS platform by employment agents per working day normalised by the sum intensity, willingness to take up work, or mismatch of unemployed and vacancies. Structural breaks after periods of (compare [3, 12, 16]). Since time variation in matching missing data are eliminated by level shift dummies in ARMA models. efficiency can be substantial (e.g. [12, 23]), matching effi - Seasonally adjusted data ciency is allowed to vary over time. Subtracting ln(U t−1 from both sides of the equation yields ln(jfr ) = µ + αln(θ ), t t t−1 (2) where jfr denotes the job finding rate and θ = V /U pursuing their placement tasks under corona condi- labour market tightness. tions and high priority attached to short-time work. This theoretical framework has several implications Job seekers’ search intensity decreased, too, both due for the cyclical behaviour of search intensity. From the to a decreasing nominator (activated visits) but also firms’ perspective, an upswing is—ceteris paribus—con - due to an increasing denominator (unemployment). nected to decreasing unemployment and hence a lower Figure  4 and the last column of Table  1 show the level of hirings (Eq.  1). However, the firms can react by development and summary statistics of the job seek- posting more vacancies and increasing search effort ers’ search perimeter. It moves within a rather limited (Eq. 2) in order to obtain the same level of hirings. In fact, range of 1.2 km. This is mainly due to the measure there is evidence in the literature supporting this reason- being a monthly average among all entries and needs ing. Davis et al. [3], for instance, find procyclical recruit - to be kept in mind when interpreting the regression ment intensity. Employers increase their hiring efforts in results in Sect. 4.2 with respect to economic relevance. 1 Page 6 of 16 C. Hutter stronger - and thus tighter - labour markets in order to 0.04 -0.9 fill their positions. 0.02 -1.0 From the job seekers’ perspective, applying the same reasoning leads to counter-cyclical search behaviour. A 0.00 -1.1 stronger labour market with lower unemployment levels -0.02 -1.2 leads—ceteris paribus—to higher job offer arrival rates -0.04 -1.3 (see, e.g. [19]), requiring less search effort from the job -0.06 -1.4 seeker to obtain the same level of hirings. In a similar reasoning, DeLoach and Kurt [5] argue that search effort -0.08 -1.5 can be countercyclical because it is intensified in down - -0.10 -1.6 turns to prevent declines in household wealth. A procy- 2015 2016 2017 2018 2019 2020 clical search effort on the firms’ side and countercyclical logofGDP,linearlydet rended (leftscale) search intensity on the job seekers’ side is in line with the logoflabour market tightness(right scale) compensation argument: Both market sides can com- Fig. 5 GDP and labour market tightness. GDP: Imputed data [6] using pensate for a tighter (firms) or worse (job seekers) labour quarterly GDP as anchor variable and monthly industrial production market via increased search efforts. This would also apply as auxiliary variable. Source: Federal Statistical Office. Labour market tightness: Number of vacancies divided by number of unemployed. for the search perimeter since job seekers’s willingness Source: Federal Employment Agency. Seasonally adjusted data to compromise probably increases as the labour market situation worsens. However, one could also think of a mechanism where lower chances to receive a job offer in the fact that the latter usually lags the former, Klinger times of labour market slackness can discourage search and Weber [13], for instance, document a sizeable [5]. Then, job seekers’ search intensity (and probably also decoupling between business cycle and the labour mar- search perimeter) would be procyclical if the discourage- ket in Germany, especially so during the last decade. To ment mechanism dominates. In this context, it is possible cover both, gross domestic product (GDP) and labour that the perceived severity and permanence of a down- market tightness ( θ ), defined as number of vacancies turn could influence the relative importance of the two divided by the number of unemployed, are used. Cal- competing channels. If a crisis is not being perceived as endar- and seasonally adjusted GDP (index: 2015=100) transitory, it could well be that the discouragement effect was taken from the Federal Statistical Office (FSO). It is dominates the compensation effect. only available at a quarterly frequency, which is why the The behaviour of the third actor, the employment imputation algorithm by Denton [6] was implemented agency, can also vary over time, although it is not clear a using industrial production (also from the FSO) as aux- priori whether it follows the cycle of the demand or sup- iliary variable in order to generate a monthly GDP time ply side, or none at all. One could think of procyclical series. Figure  5 shows that there was a rather stable placement behaviour if a tighter labour market requires upswing in GDP until the end of 2017 before the down- the agency to more intensively support firms in their turn in 2018 and 2019. Labour market tightness grew search for suitable candidates. By contrast, the agency strongly until the end of 2017, followed by a period of might intensify support for unemployed if the economic slower growth until mid of 2019, after which it steadily conditions worsen. It is conceivable, for example, that the decreased. Both variables dropped severely due to the job search could be made more comprehensive beyond COVID-19 crisis. All in all, both variables experienced the standard or that the placement officer to job seeker upswing and downturn during the sample even before ratio could be improved. Hence, the cyclical behaviour of COVID-19, which allows to investigate the cyclical pat- the search and placement activities is an empirical ques- terns of the search and placement intensities. tion that will be answered in the following. A visual impression of the behaviour of the search and placement intensities in the course of a business 4 Cyclicality and seasonality of search or labour market cycle can be obtained by contrast- and placement activities ing them with GDP and tightness in scatter plots. This 4.1 S catter plots on cyclicality provides a first idea about the pro- or countercycli - To get a first impression, this subsection analyses the cal nature of the search activities on the labour mar- cyclical behaviour of search and placement intensity via ket. Figures  6 and  7 show the pairwise relationships scatter plots. for GDP and labour market tightness, respectively. First, it needs to be addressed whether “cycle” means All scatter plots contain data until 2020:3 and hence the business cycle or the labour market cycle. Beyond exclude the extreme months of the COVID-19 crisis. Cyclicality of labour market search: a new big data approach Page 7 of 16 1 2.6 0.26 2.4 0.24 2.2 0.22 2.0 0.20 1.8 0.18 1.6 0.16 1.4 0.14 1.2 0.12 -2 -1 0 1 2 3 -2 -1 0 1 2 3 Gross domestic product Gross domestic product 1.2 44.8 1.1 44.6 1.0 44.4 0.9 44.2 0.8 44.0 0.7 43.8 0.6 43.6 0.5 43.4 -2 -1 0 1 2 3 -2 -1 0 1 2 3 Gross domestic product Gross domestic product Fig. 6 Scatter plots of search intensities and GDP. Gross domestic product: Linearly detrended data (i.e. a linear trend was subtracted from the original data). Search activity variables: The lags of GDP yielding the strongest correlation in the regressions of Table 2 were used in the scatter plots: JS F EA I and GDP , I and GDP , I and GDP , SP and GDP . The red lines visualise the fitted linear relationship stemming from OLS regressions t−1 t−3 t−2 t t−3 t t t 2.6 0.26 0.24 2.4 0.22 2.2 0.20 2.0 0.18 1.8 0.16 1.6 0.14 1.4 0.12 1.2 0.10 0.20 0.24 0.28 0.32 0.36 0.20 0.24 0.28 0.32 0.36 0.40 Labourmarkettightness Labourmarkettightness 44.8 1.2 44.6 44.4 1.0 44.2 0.8 44.0 0.6 43.8 43.6 0.4 43.4 0.20 0.24 0.28 0.32 0.36 0.40 0.20 0.24 0.28 0.32 0.36 Labourmarkettightness Labourmarkettightness Fig. 7 Scatter plots of search intensities and labour market tightness. Labour market tightness: Number of vacancies divided by number of JS F EA unemployed. The lags of θ yielding the strongest correlation in the regressions of Table 3 were used in the scatter plots: I and θ , I and θ , I and t t−1 t t t θ , SP and θ t−2 t t−1 Placement intens ity Job seekers’ search intensity Placement intensity Jobseekers’ search intensity Search perimeter Firms’ search intensity Search perimeter Firms’ search intensity 1 Page 8 of 16 C. Hutter 0.4 0.12 0.2 0.08 0.0 0.04 -0.2 0.00 -0.4 -0.04 -0.6 -0.08 -0.8 -0.12 -0.02 0.00 0.02 0.04 0.06 -0.02 0.00 0.02 0.04 0.06 Grossdomesticproduct Grossdomesticproduct 0.3 1.00 0.2 0.75 0.1 0.50 0.0 0.25 -0.1 0.00 -0.2 -0.25 -0.3 -0.50 -0.4 -0.75 -0.5 -0.6 -1.00 -0.02 0.00 0.02 0.04 0.06 -0.02 0.00 0.02 0.04 0.06 Grossdomesticproduct Grossdomesticproduct JS F EA Fig. 8 Scatter plots of search intensities and GDP—annual differences. The notes of Table 6 apply, except that annual differences of GDP, I , I , I , and SP have been used Both the placement intensity of employment agencies cycle variables lets the data speak about whether search and the firms’ search intensity seem to move pro-cycli - activity reacts immediately or after a delay to cyclical cally. The opposite holds for the labour supply side: Job movements. For each possible combination of search seekers’ search intensity and their search perimeter show activity and cycle variable, the following regression signs of counter-cyclical movements. equation is estimated: To the extent the variables are subject to persistence, there is a risk that the respective scatter plots might dis- ln(y ) = α + β ln(x ) + γ t + ǫ , t i t−i t (3) play pro- or countercyclicality due to spurious correla- i=0 tions. However, the following scatter plots based on the JS F EA where y ∈ (I , I , I , SP) , x ∈ (GDP, θ) , p is the lag annual differences of the variables show that this risk does length, and ǫ is a normally-distributed error term. Equa- not materialize here. Differencing, i.e. subtracting lagged tion (3) controls for a linear trend (γ t ). This way, β to β values, removes potential stochastic as well as deter- 0 p capture the cyclical effects and do not pick up correlation ministic trends from the time series and hence is a suit- stemming from linear trends in the variables. As in the able method to deal with persistent time series. Figures 8 scatter plots, the estimation period does not cover the and 9 show that in all cases, the scatter plots look similar COVID-19 crisis months 2020:4 and 2020:5 due to their to those of Figures 6 and 7. Hence, the visual impressions extreme values. regarding pro- or countercyclicality are confirmed for A further matter of interest are data properties such both the level and the differenced variables. as the persistence of variables in the given sample period. The scatter plots already gave a first impres - 4.2 Regression analysis sion that this might not be a decisive issue here. Indeed, The scatter plots gave a first impression. The follow - Augmented Dickey-Fuller (ADF) tests confirm that ing steps involve a more formal analysis in which also GDP does not show the persistence usually found in information about significance and lag structure of the longer samples: They find that GDP behaved as trend - cyclical relationships can be obtained. While so far stationary variable during the sample at hand. The it remained unclear how long it takes for the cycle to null hypothesis of GDP having a unit root is rejected materialise in the search activities on the labour mar- at the 1% significance level in an ADF test with a con - ket, regression analysis is suited to address this ques- stant and a deterministic trend as exogenous variables. tion. Including contemporaneous as well as lagged Placementintensity Jobseekers’ search intensity Search perimeter Firms’ search intensity Cyclicality of labour market search: a new big data approach Page 9 of 16 1 0.4 0.12 0.2 0.08 0.0 0.04 -0.2 0.00 -0.4 -0.04 -0.6 -0.08 -0.8 -0.12 -0.06 -0.02 0.02 0.04 0.06 0.08 -0.06 -0.02 0.02 0.04 0.06 0.08 Labourmarkettightness Labour market tightness 0.3 1.00 0.2 0.75 0.1 0.50 0.0 0.25 -0.1 0.00 -0.2 -0.25 -0.3 -0.50 -0.4 -0.75 -0.5 -0.6 -1.00 -0.06 -0.02 0.02 0.04 0.06 0.08 -0.06 -0.02 0.02 0.04 0.06 0.08 Labourmarkettightness Labour market tightness Fig. 9 Scatter plots of search intensities and labour market tightness—annual differences. The notes of Table 7 apply, except that annual differences JS F EA of θ , I , I , I , and SP have been used Table 2 Regression results: search intensities and GDP Dependent variable JS F EA SP I I I Constant 39.23*** (14.05) − 27.31∗ (14.39) − 77.84∗∗ ∗ (13.53) 7.18∗∗ ∗ (0.75) GDP 0.39 (0.81) 0.27 (1.14) 0.15 (1.29) 0.09** (0.04) GDP − 5.55*** (1.08) − 0.63 (2.35) 3.46** (1.74) − 0.13 (0.12) t−1 GDP − 1.65 (1.76) 0.07 (1.53) 7.22* (4.03) − 0.26** (0.11) t−2 GDP − 1.52 (0.99) 5.82*** (0.74) 6.03** (3.03) − 0.43*** (0.14) t−3 trend 0.45 (0.42) − 0.81** (0.33) − 2.69*** (0.45) 0.04** (0.02) R-squared 0.6684 0.1505 0.5104 0.4409 Estimated effects from OLS regressions with heteroscedasticity- and autocorrelation-robust standard errors (in parentheses). *, **, *** denote significance at the 10, JS F EA 5, 1% level, respectively. I , I , I : Search intensity of job seekers/firms/employment agencies. SP: Average search perimeter of job seekers in kilometers. All variables enter the regressions in logarithms Nonetheless, regressions on the annual growth rate of relationships prove to be significant in any of the regres - GDP were conducted as robustness check in Sect. 4.4. sions for at least one lag, which is remarkable given the Equation (3) is estimated using ordinary least squares limited number of observations. (OLS) with heteroscedasticity- and autocorrelation- Although I would recommend not to over-interpret robust standard errors. All variables appear in logs. As the lag structure at this early stage, one result seems evi- baseline, a lag length of p = 3 is chosen to allow delayed dent: On average, the reactions to labour market cycle effects up to a quarter (i.e. three months). A robustness movements are quicker than to business cycle move- check on the lag length is presented in Sect. 4.4. Tables 2 ments. For instance, a tighter labour market materialises and 3 show the results. instantaneously in higher search efforts of firms (second In general, they confirm the visual impression obtained column of Table 3) while it needs a while in case of GDP- F EA by the scatter plots: I and I move pro-cyclically while changes (second column of Table  2). While surprising at JS I and SP move counter-cyclically. Furthermore, these first glance, it could indeed be rational for firms not to Placementintensity Jobseekers’ search intensity Search perimeter Firms’ search intensity 1 Page 10 of 16 C. Hutter Table 3 Regression results: Search activities and tightness Dependent variable JS F EA SP I I I Constant 0.74 (0.51) − 1.41∗∗ ∗ (0.46) 1.38∗∗ ∗ (0.24) 3.76∗∗ ∗ (0.06) θ − 0.67 (1.33) 2.75∗∗ (1.38) 5.29∗∗ ∗ (1.36) 0.22 (0.29) θ − 2.28∗ (1.28) − 0.74 (2.01) − 6.81∗ (3.84) − 0.32∗∗ (0.12) t−1 θ 0.50 (3.07) − 2.19 (3.26) 5.79∗ (2.98) − 0.04 (0.25) t−2 θ 2.32 (3.67) 0.50 (2.74) − 3.20∗∗ (1.58) 0.11 (0.13) t−3 trend 0.94∗∗ (0.41) − 0.08 (0.38) − 0.96∗∗ ∗ (0.16) − 0.01 (0.06) R-squared 0.6050 0.1214 0.6863 0.2784 Estimated effects from OLS regressions with heteroscedasticity- and autocorrelation-robust standard errors (in parentheses). *, **, *** denote significance at the 10, JS F EA 5, 1% level, respectively. I , I , I : Search intensity of job seekers/firms/employment agencies. SP: Average search perimeter of job seekers in kilometers. θ : labour market tightness. All variables enter the regressions in logarithms Table 4 Regression results including both GDP and tightness Dependent variable JS F EA SP I I I Constant 63.08∗∗ ∗ (16.89) 6.22 (25.23) − 5.45 (20.72) 6.66∗∗ ∗ (1.82) GDP − 0.26 (0.66) − 0.48 (0.98) − 1.50 (1.24) 0.10 (0.08) GDP − 6.78∗∗ ∗ (2.10) − 4.47 (3.16) − 0.23 (2.11) − 0.06 (0.20) t−1 GDP − 3.26∗ (1.69) − 1.96 (2.38) 0.80 (3.64) − 0.24∗ (0.13) t−2 GDP − 2.98∗∗ (1.42) 5.21∗∗ ∗ (1.94) 2.44 (2.54) − 0.42∗∗ ∗ (0.14) t−3 θ 0.06 (1.35) 2.66∗ (1.34) 5.49∗∗ ∗ (1.30) 0.10 (0.22) θ − 0.53 (1.51) 1.52 (1.53) − 7.38∗∗ (3.49) − 0.22∗ (0.11) t−1 θ 1.31 (3.49) − 3.21 (3.14) 5.22∗ (3.03) 0.16 (0.26) t−2 θ − 0.19 (3.35) − 0.74 (2.89) − 2.19 (1.85) − 0.05 (0.29) t−3 trend 0.51 (0.58) 0.41 (0.69) − 1.35∗∗ (0.60) 0.04 (0.07) R-squared 0.6893 0.2330 0.7051 0.4636 Estimated effects from OLS regressions with heteroscedasticity- and autocorrelation-robust standard errors (in parentheses). *, **, *** denote significance at the 10, JS F EA 5, 1% level, respectively. I , I , I : Search intensity of job seekers/firms/employment agencies. SP: Average search perimeter of job seekers in kilometers. θ : labour market tightness. All variables enter the regressions in logarithms react immediately to (potentially short-lived) fluctuations placement intensity. Indeed, when estimating the third but instead to wait until an economic upswing or down- column of Table 3 with differenced log of theta instead of turn proves to be stable before making any decisions the level, the signs do not “jump” anymore, and β and β 0 2 with respect to their search behaviour. Furthermore, the are highly significant with estimated effects of 6 to 7%. labour market itself lags the real economy. At first glance, the effects seem to be less relevant in The significant effects are also relevant in size, although case of the search perimeter. They are much smaller, with to different extents: For instance, a positive 1% GDP the sum of the effects ranging between − 0.3 and − 0.5% F JS change increases I and decreases I by 5 to 6%, while it after 3 months (fourth column). However, the standard EA increases I by almost 17% after 3 months. In case of θ , deviation of SP amounts to only 0.33 km (or 0.75% in the effects are about half the size, which is compensated terms of its average). Hence, put into context, the esti- by the fact that during the sample, the variation of tight- mated effects are relevant after all. ness is much higher than that of GDP. Overall, the explanatory power of the trend and cycle With respect to placement intensity, the changing variables varies from 12 to 69%. While the search inten- signs (positive for lags of 0 and 2, negative for lags 1 and sity of firms is obviously influenced a lot by factors 3) indicate that the growth rate of tightness matters for beyond the aggregate business or labour market cycle, This can be seen when putting the absolute effects outside the brackets. Cyclicality of labour market search: a new big data approach Page 11 of 16 1 Table 5 Regression results: controlling for compositional effects Dependent variable JS F EA SP I I I Constant 78.01∗∗ ∗ (27.87) 50.75∗ (28.18) − 8.69 (24.00) 3.55∗∗ ∗ (1.06) GDP − 0.40 (1.01) − 1.16 (1.27) − 1.66 (0.97) 0.12 (0.08) GDP − 7.16∗∗ (2.90) − 5.54 (3.95) 0.47 (2.62) 0.02 (0.12) t−1 GDP − 3.53 (2.25) − 3.48 (2.62) 3.05 (3.65) − 0.02 (0.12) t−2 GDP − 3.31∗ (1.64) 4.09∗ (2.33) 1.46 (2.48) − 0.24∗∗ (0.12) t−3 θ 0.33 (1.54) − 0.99 (1.63) 7.29∗∗ ∗ (2.35) 0.36∗∗ ∗ (0.11) θ 0.02 (1.74) 5.06∗∗ (2.20) − 9.74∗ (4.01) − 0.34∗ (0.18) t−1 θ 0.60 (3.87) − 2.55 (3.59) 5.24 (3.72) 0.00 (0.17) t−2 θ − 0.70 (3.28) − 0.82 (2.86) − 1.63 (3.00) − 0.02 (0.14) t−3 trend − 0.38 (1.02) 0.06 (1.70) − 2.65∗∗ (0.97) 0.14 (0.12) R-squared 0.7668 0.3986 0.7968 0.6271 The notes of Table 4 apply. The control variables comprise the shares of high-education (college degree), low-education (neither vocational training nor high school degree), older (> 55 years of age), younger (≤ 25 years of age), female, and of foreign people among total inflow into unemployment two thirds of the variation in job seekers’ search intensity and hence influence the observed search intensity. For can be explained by the business cycle (the trend is irrel- instance, the composition of the monthly inflow into evant here). unemployment could (partly or fully) explain the cyclical In a more general setting, regression analysis allows patterns of the search activities, not only of job seekers both GDP and θ to play a role. By including both variables but also of firms and placement agents. To account for in the same equation, it is possible to know whether there such potential compositional effects, this subsection adds is a dominant cycle that influences search and placement several control variables to all regression equations. activities. Table 4 shows the results. The set of monthly available variables capturing rele - In case of the job seekers’ search intensity, the results vant characteristics of job seekers comprises the shares of show that the business cycle dominates the labour market high-education (college degree), low-education (neither cycle. The dependence of the firms’ search intensity on vocational training nor high school degree), older (> 55 the cycle variables remains rather unchanged when both years of age), younger (< = 25 years of age), female, and GDP and θ are included in the regression. It is still con- of foreign people among total inflow into unemployment. temporaneous labour market tightness and the 3-month The shares are taken from the FEO’s statistics. Table  5 lag of GDP that exert the procyclical effects. In addition, shows the results. also the size of the estimated effects barely change. In With respect to the job seekers’ search intensity, the case of placement intensity, it is the labour market cycle control variables are able to explain an additional share of that dominates the business cycle. The coefficients of θ the variation in search intensity (The R-squared increases do not change much, which means also the more com- by 8% points). Including them does not lead to decreas- plex lag structure found above remains. Both GDP and θ ing cyclical effects, though. The effects of job seekers’ remain significant factors for the search perimeter of the search intensity with respect to GDP remain basically job seekers even when both are added. Also the lag struc- unchanged. However, statistical significance is slightly ture remains unchanged compared to Tables 2 and 3. weaker when control variables are included. To sum up, for job seekers and firms, the economic In case of the firms’ search intensity, adding the control cycle seems to be the more relevant factor, while employ- variables increases the R-squared by a substantial amount ment agencies are mainly driven by the labour market (almost 17% points). However, the procyclical movement cycle. found so far does not disappear. Both GDP and θ remain significant factors, although the contemporaneous effect 4.3 C ontrolling for compositional effects of θ becomes insignificant and is replaced by the first lag While the results so far reveal cyclical patterns of the instead. search activities, one has to be careful in interpreting Also for the relationship between placement inten- these findings in terms of changing search behaviour. sity and cycle, the control variables do not play a game- Beyond the search behaviour, also the composition of the changing role. Although adding them increases the searchers could change during an upswing or downturn R-squared by approximately 9% points, the estimated 1 Page 12 of 16 C. Hutter cyclical coefficients do not change much. Again, their sta - market sides with increased search on the labour market. tistical significance is slightly weaker when control vari - Job seekers search more and widen their search perim- ables are included. eter in times of weaker economy and labour market while In case of the search perimeter, adding the control firms search more in times of stronger economy and variables increases the R-squared by 16% points. The labour market. Since the employment agencies are the countercyclical behaviour with respect to labour mar- intermediary between both market sides, their cyclical ket tightness disappears although it remains intact with behaviour is ambiguous from a theoretical point of view. respect to the business cycle. The results on the search The empirical results indicate that—like firms’ search perimeter, however, are based on 30 observations only, intensity—placement intensity increases when the labour so estimating 16 parameters might be a challenge that is market becomes tighter. A potential explanation could just a bit too high. Thus, probably a bit more time is war - be that in times of tighter labour markets there are shifts ranted in order to collect longer time series before even- within the FEA away from the placement of unemployed tually answering this question for the search perimeter. towards the employer service (“Arbeitgeberservice”, To sum up, there is some evidence that observable a department of the FEA supporting firms in finding compositional effects are able to explain a part of the employees). development of search and placement activities. How- ever, they do not substantially weaken the findings 4.4 Robustness checks regarding the cyclical behaviour. Of course, it is still pos- To check robustness of the results, richer specifications sible that there are unobservable compositional effects are estimated using 6 instead of 3 lags of GDP or θ to at play. For instance, the relation of unemployed search- allow for delayed effects up to half a year. The choice of ers versus on-the-job searchers could change during an the lag length is supported by statistical tests and infor- upswing or downturn. Similarly, it is possible that job mation criteria. For instance, in no case do the Schwarz seekers and firms change their preferred search channel criterion and the Akaike information criterion recom- during a business cycle, which could potentially change mend using a lag length of more than 6 months. Tests the composition of job seekers or firms using the FEA’s based on the Ljung-Box Q-statistics reveal that the null online job exchange. Although the aggregate search activ- hypothesis of no serial correlation in the residuals is ity data at hand do not allow to look into the details of not rejected at the 5 percent significance level in case of who accesses the websites but only how often the websites p = 3, and even at the 10 percent level in case of p =  6. are accessed, there is little evidence that changing search Table 6 shows the results for p = 6. channel preferences on the firms’ side play a decisive role They confirm the general results of Tables  2 and  3 on during the sample period. According to the job vacancy the cyclical behaviour of the search and placement activi- survey of the Institute for Employment Research (see, e.g. ties. In many cases, also the same lags of the cycle vari- [17]), the FEA’s online job exchange was mentioned as ables are significant—and they are similar in size. For relevant search channel in one third of the cases in every I , the 6th lag of GDP is significant, too, substantially single year since 2015, which does not leave much room increasing its procyclical response. Thus, the smaller for major compositional effects. model with 3 lags can be considered conservative. For EA Furthermore, one could think of other potential con- I , the bigger model prefers more delayed reactions to trol variables. For instance, the reservation wage could GDP-changes, and the sum of the effects is a bit higher be a relevant factor for the job seekers’ search behav- than in the specification with 3 lags. In case of SP , the iour. However, the reservation wage cannot be observed significant effects of the 4th and 6th lag of GDP balance directly but at most be modelled as a function of labour each other out so that the total effect does not change market tightness, the structure of the unemployed (both much. However, the countercyclical reaction of SP to of which are accounted for in the regressions), or of the θ disappears in the specification with 6 lags, a phenom - net replacement rate of the unemployed. With regards to enon already found when control variables were added. the latter, annual OECD data show that the net replace- In a further robustness check, the regressions are per- ment rate basically remained unchanged since 2015 at formed on the annual growth rate of GDP. The pro- or 59%, so again no major effects can be expected here. countercyclical patterns as well as the lag structure do JS Against the background of the theoretical reason- not change substantially, though. In case of I , the big- ing in Sect.  3, the results indicate that job seekers and gest effect is still stemming from the first lag of GDP firms seem to compensate difficulties on their respective growth with an estimated effect of −  4.79. For I , the highest effect of 3.48 is estimated to occur at the third The share never drops below 32% and never exceeds 34%. 9 10 See https ://data.oecd.org/benwa ge/bene fits-in-unemp loyme nt-share -of- However, linking the VerBIS data to other data sets to control for the com- previ ous-incom e.htm. position of the placement agents using VerBIS is not (yet) possible. Cyclicality of labour market search: a new big data approach Page 13 of 16 1 Table 6 Robustness checks on lag length Dependent variable JS F EA SP I I I GDP Constant 26.23∗∗ (12.69) − 46.00∗∗ ∗ (12.98) − 95.16∗∗ ∗ (5.04) 8.00∗∗ ∗ (0.93) GDP 0.01 (0.61) − 0.45 (0.98) − 1.12 (1.23) 0.07∗∗ (0.03) GDP − 6.80∗∗ ∗ (1.69) − 2.52 (1.89) − 0.85 (3.19) − 0.16 (0.15) t−1 GDP − 2.51∗∗ (1.02) − 1.18 (1.43) 3.51 (3.31) − 0.34∗∗ (0.13) t−2 GDP − 1.85 (1.72) 5.40∗∗ ∗ (1.16) 5.19∗ (3.07) − 0.48∗∗ ∗ (0.11) t−3 GDP 2.40 (2.63) 0.86 (1.01) 4.53∗∗ (2.19) − 0.24∗ (0.13) t−4 GDP 0.18 (1.35) 1.64 (3.26) 3.39∗∗ (1.42) − 0.03 (0.16) t−5 GDP 3.05 (2.08) 5.83∗∗ ∗ (1.80) 5.99∗∗ (2.83) 0.27∗∗ ∗ (0.08) t−6 Trend 0.04 (0.40) − 1.42∗∗ ∗ (0.32) − 3.42∗∗ ∗ (0.17) 0.04∗ (0.02) Tightness Constant 1.74∗ (0.89) − 0.53 (1.11) 1.99∗∗ ∗ (0.64) 3.87∗∗ ∗ (0.05) θ 0.37 (1.29) 3.64∗∗ (1.43) 6.33∗∗ ∗ (0.84) 0.42∗∗ (0.20) θ − 2.42∗∗ (1.17) − 0.68 (1.60) − 8.04∗∗ (3.09) − 0.39∗ (0.22) t−1 θ − 0.51 (3.51) − 2.91 (3.27) 5.24∗ (2.85) − 0.09 (0.26) t−2 θ − 0.52 (3.50) − 0.38 (3.98) − 6.44∗∗ ∗ (2.24) − 0.25 (0.18) t−3 θ 2.11 (2.30) -1.27 (4.81) 6.59∗∗ ∗ (1.85) 0.01 (0.15) t−4 θ 0.33 (3.33) 0.71 (2.69) − 3.48∗∗ (1.49) 0.13 (0.27) t−5 θ 1.12 (1.09) 1.76 (2.69) 1.23 (2.61) 0.21 (0.20) t−6 Trend − 1.73∗∗ (0.68) − 0.69 (0.82) − 1.49∗∗ (0.61) − 0.11∗∗ ∗ (0.04) Estimated effects from OLS regressions with heteroscedasticity- and autocorrelation-robust standard errors (in parentheses). *, **, *** denote significance at the 10, JS F EA 5, 1% level, respectively. I , I , I : Search intensity of job seekers/firms/employment agencies. SP: Average search perimeter of job seekers in kilometers. θ : labour market tightness EA 0.4 activated visits per unemployed per working day lag of GDP growth. In case of I , the second lag of GDP below the trend-cycle level) and highest in February growth exerts the strongest effect with an estimated (around 0.5 activated visits above the trend-cycle). effect of 4.24. And SP reacts strongest to the third lag of Note that the depicted months are not calendar GDP growth with an effect of − 0.19. All reported effects months but months between the counting days (see are found to be statistically significant. Sect. 2). Thus, a typical January covers the period from Finally, the monthly available index of industrial pro- mid of December to mid of January, a typical February duction was used instead of imputed GDP. However, the F EA goes from mid of January to mid of February, and so results do not change substantially. I and I still move JS forth. Consequently, the period of lowest search inten pro-cyclically while I and SP move counter-cyclically. sity on the job seekers’ side usually covers the holiday season of Christmas and New Year’s Eve. Another, less 4.5 The seasonal patterns of search behaviour pronounced, local minimum in the seasonal pattern is The data on search and placement activities used until visible in August and September, most probably due to now were seasonally adjusted. Beyond the cyclical move- summer vacation. ments, also the usual seasonal behaviour in the search The minima of the firms’ search intensity occur in and placement intensities could be of interest as it might January and September, too. However, the highest provide explanations for seasonal patterns in aggregate efforts can be detected from June to August, most likely unemployment or employment. Figure 10 shows the sea- due to additional efforts to duly recruit apprentices sonal patterns in the novel data. They are obtained by before the vocational training cycle starts. The firms applying the additive X12-ARIMA seasonal adjustment seem to be supported by the intermediary since July is procedure. also the period in which employment agencies under The graphs show how many additional activated take the highest placement efforts. The other maxi - visits usually occur in a specific month, beyond the mum of placement intensity is in February, matching trend-cycle level. For instance, job seekers’ search the maximum of job seekers’ search intensity. And also intensity usually is lowest in January (approximately 1 Page 14 of 16 C. Hutter 0.6 140 0.4 0.2 0.0 -0.2 -0.4 Jobseekers’searchintensity -0.6 2015 2016 2017 2018 2019 2020 M2 M3 M4 M5 M6 M7 0.03 0.02 0.01 Job seekers’ search intensity 0.00 Firms’ search intensity Placement intensity -0.01 -0.02 Fig. 11 Search and placement during the COVID-19 pandemic. Daily -0.03 data using a 7-day moving average to eliminate day-of-the-week effects; holiday-adjusted using ARMA models with dummies; index -0.04 Firms’ search intensity (January 6 2020 = 100) -0.05 2015 2016 2017 2018 2019 2020 0.08 the labour market. Consequently, January typically is the month with the highest unemployment rate in the 0.04 course of a year. 0.00 No seasonal pattern could be detected for the search perimeter (neither significant nor relevant in size). This -0.04 is no surprise given the development shown in Fig. 4. The maxima and minima do not occur at the same months. -0.08 Placementintensity -0.12 2015 2016 2017 2018 2019 2020 5 Search and placement during the COVID‑19 Fig. 10 Seasonal patterns of search and placement intensities. The pandemic seasonal patterns for search and placement intensities stem from So far, the COVID-19 months were excluded in the additive X12-ARIMA seasonal adjustment analysis. This is justified by the extreme disruption this pandemic has caused, as can be seen in Figs.  1, 2, 3, 4 and  5. Including such extreme values in scatter plots or regression analysis would dominate the results and make the minima of placement intensity closely match those it difficult to investigate reactions as they would occur in of the labour demand and supply sides: They occur in normal times. At the same time, the long-term conse- January, September, and November. quences of this crisis are not yet clear. It is possible that Measuring the extent of the seasonal pattern relative long-lasting shifts in the economy and the labour market to the mean search intensity, the seasonality on the job will remain even when the virus is under control. seekers’ side is found to be a bit higher (ranging from However, the data at hand allow for a descriptive analy- − 25 to +  30% of its mean search intensity) than that sis of search and placement in the COVID-19 pandemic. on the firms’ side, ranging from − 25 to +  15% (com- Since during the first lockdown in March 2020 the news pare also [4] who find that seasonality is much bigger situation changed almost on a daily basis, the fact that for applications than postings). the data at hand allow for a daily analysis becomes all the All in all, the seasonal patterns seem to be largely more valuable. Figure 11 shows the daily development of influenced by holiday seasons and the vocational train - the search and placement intensities of all three agents ing cycle. Since the seasonal patterns on both market on the labour market. For scaling purposes, the lines are sides and the employment agencies largely coincide, indexed so that they all start at a value of 100. they most probably reinforce each others’ effects on Cyclicality of labour market search: a new big data approach Page 15 of 16 1 Between March 5th and March 25th, the firms’ search activities bounce back more quickly after the COVID-19 intensity dropped substantially by 42%. Until the end of crisis. the sample (July 13th), it recovered again, reaching 92.5% In the future, further data from the BA job exchange of the pre-crisis level. There was also a sharp (− 35%) and could be exploited for scientific use, especially data on even faster (until March 17th) slump of placement inten- search behaviour. Beyond the search perimeter, other sity, whereas the subsequent recovery proceeded much entries on the job exchange could be made accessible more slowly. Until the end of the observation period, for research. For instance, the FEA plans a change from only 71% of the pre-crisis level had been reached again. Netmind to Matomo, after which information from the Placement intensity declined strongly since employment search masks such as the desired profession will be saved agents had to prioritize the processing of short-time work and could then be accessed for research purposes. Future requests, in the course of which the placement activity projects may make it possible to merge the search activity that would be usual in normal times could be carried out data to administrative data. Then, it could be investigated only on a considerably reduced scale. how characteristics such as qualification, occupation or The job seekers’ search intensity displays a more com - region affect the search duration. It would also be con - plex development which was hidden in the monthly fig - ceivable to analyze the relationship between the duration ures. Within ten days, it initially declined sharply before of unemployment or the period of time a job vacancy strongly recovering and even overshooting pre-crisis-lev- exists and the search intensity. els. However, it quickly declined afterwards to very low Acknowledgements levels and has not recovered since. A possible explana- I thank Hermann Gartner, Tobias Hartl, Ortwin Herbst, Maren Müller, Clemens tion for this development is that an intensified search in Usbeck, Enzo Weber, and participants of the annual meeting of the FEA’s sta- tistics department 2018, the workshop “Genesis of information from online job the first days after the lockdown was not seen as a pri - advertisements” 2019, and the IWH/IAB workshop 2020 for valuable support ority due to extraordinary challenges to cope with the and helpful suggestions. situation, and the search intensity therefore initially sank. Authors’ contributions Subsequently, a compensation effect responsible for the I am the sole contributor to the writing of this article. The author read and usual counter-cyclical search pattern (see Sect. 3) appears approved the final manuscript. to have been dominant until the end of March, while the Funding discouragement effect could have been the dominant Not applicable. driver relatively quickly from April onwards. Availability of data and materials The data on online visits of the job exchange platform of the Federal Employ- 6 Conclusion ment Agency and its internal placement software VerBIS are initially stored This article introduces innovative online data allowing in server log files. In an aggregated form, they can be downloaded via the the instantaneous measurement of search and—for the software Netmind. Access is provided by the Federal Employment Agency. The other aggregate data used in this article stem from the Federal Statistical first time—placement intensity in the labour market in Office (in case of GDP and production index), and from the Statistics Depart - form of online activity. These data are used to estimate ment of the Federal Employment Agency (in case of unemployment, vacan- their behaviour during the economic and labour mar- cies, and the control variables). They can be downloaded from the respective websites. Alternatively, I can provide the data upon request. ket cycle, as well as their usual seasonal patterns. The results show that firms’ and employment agencies’ search Competing interests and placement intensity displays a pro-cyclical pattern I declare that there are no competing interests, neither financial nor non-financial. while—at least before the COVID-19 pandemic—job seekers’ search intensity is counter-cyclical. Received: 2 July 2020 Accepted: 15 December 2020 In the COVID-19 crisis so far, the data reveal that the placement intensity of employment agencies and the firms’ search intensity dropped substantially. Looking at the daily data, the job seekers’ search intensity displays a References 1. Barnichon, R., Figura, A.: Labor market heterogeneity and the aggregate more complex development. Here, the procyclical driv- matching function. Am. Econ. J. 7, 222–249 (2015) ers seem to eventually dominate during the COVID-19 2. Carrillo-Tudela, C., Gartner, H., Kaas, L.: Recruitment policies, job-filling pandemic, contrary to the counter-cyclical pattern found rates and matching efficiency. IAB-Discussion Paper 15/2020, 51 pages (2020) for normal times. The overall drop in search and place - 3. Davis, S.J., Faberman, R.J., Haltiwanger, J.C.: The establishment-level ment activities naturally will have adverse impacts on behavior of vacancies and hiring. Q. J. Econ. 128(2), 581–622 (2013) the labour market, especially on the job finding rate [9]. 4. Davis, S. J., de la Parra, B. S., 2018, Application Flows (online presenta- tion) https ://stati c1.squar espac e.com/stati c/5e2ea 3a809 7ed30 c779b The most critical labour market effects of the crisis may d707/t/5e315 71592 58077 0732d ac3b/15802 91866 530/Appli catio arise not via the separation but via the hiring margin [18]. n+Flows +Febru ary+2018.pdf It is therefore important that the search and placement 1 Page 16 of 16 C. Hutter 5. DeLoach, S., Kurt, M.: Discouraging workers: estimating the impacts of 15. Kuhn, P., Skuterud, M.: Internet job search and unemployment durations. macroeconomic shocks on the search intensity of the unemployed. J. Am. Econ. Rev. 94(1), 218–232 (2004) Labor Res. 34, 433–454 (2013) 16. Launov, A., Wälde, K.: The employment effect of reforming a public 6. Denton, F.T.: Adjustment of monthly or quarterly series to annual totals: employment agency. Eur. Econ. Rev. 84, 140–164 (2016) an approach based on quadratic minimization. J. Am. Stat. Soc. 66(333), 17. Mario, B., Nicole, G.A.: The IAB job vacancy survey: design and research 99–102 (1971) potential. J. Labour Market Res. 12, 54 (2020) 7. Faberman, R.J., Kudlyak, M.: The intensity of job search and search dura- 18. Merkl, C., Weber, E.: Rescuing the labour market in times of COVID-19: tion. Am. Econ. J. 11, 327–357 (2019) don’t forget new hires. VoxEU 07 April (2020) 8. Gomme, P., Lkhagvasuren, D.: Worker search effort as an amplification 19. Mortensen, D.T.: Job search and labor market analysis. In: Ashenfelter, mechanism. J. Monet. Econ. 75, 106–122 (2015) O., Layard, R. (eds.) Handbook of Labor Economics, vol. 2, pp. 849–919. 9. Hartl, T., Hutter, C., Weber, E.: Neueinstellungen in der Krise. Makronom, Elsevier, New York (1987) https ://makro nom.de/coron a-arbei tsmar kt-auswi rkung en-neuei nstel 20. Mukoyama, T., Patterson, C., Sahin, A.: Job search behavior over the busi- lunge n-in-der-krise -36315 (2020) ness cycle. Staff Reports 689, Federal Reserve Bank of New York (2014) 10. Holzner, C., Watanabe, M.: Do intermediaries reduce search frictions? 21. Pissarides, C.: Equilibrium unemployment theory, 2nd edn. The MIT Press, https ://www.pe.econ.uni-muenc hen.de/perso nen/postd ocs/holzn er/ New York (2000) forsc hung/pea_searc hfric tions 01.pdf (2019) 22. Sahin, A., Song, J., Topa, G., Violante, G.L.: Mismatch unemployment. Am. 11. Hutter, C., Weber, E.: Mismatch and the forecasting performance of Econ. Rev. 104(11), 3529–3564 (2014) matching functions. Oxford Bull. Econ. Stat. 79(1), 101–123 (2017) 23. Sedlacek, P.: Match efficiency and firms’ hiring standards. J. Monet. Econ. 12. Klinger, S., Weber, E.: Decomposing beveridge curve dynamics by cor- 62, 123–133 (2014) related unobserved components. Oxford Bull. Econ. Stat. 78, 877–894 (2016) Publisher’s Note 13. Klinger, S., Weber, E.: GDP-employment decoupling in Germany. Struct. Springer Nature remains neutral with regard to jurisdictional claims in pub- Change Econ. Dynam. 52, 82–98 (2020) lished maps and institutional affiliations. 14. Krueger, A.B., Mueller, A.: Job search, emotional well-being and job find- ing in a period of mass unemployment: evidence from high-frequency longitudinal data. Brook. Papers Econ. Activ. 42, 1–81 (2011)

Journal

Journal for Labour Market Research – Springer Journals

Published: Jan 23, 2021

There are no references for this article.