Drawing on rich administrative data from the Greek Public Employment Service (DYPA) and the ERGANI employment register, this chapter first analyses the profiles of registered vulnerable jobseekers, considering both the methods currently used by DYPA to identify them and alternative approaches. It then applies advanced statistical methods to pinpoint the characteristics most strongly associated with a greater distance from the labour market. It also assesses the extent to which DYPA’s active labour market policies (ALMPs) reach those most in need, highlighting gaps in service provision. Finally, building on this evidence, the chapter presents the concept for a new digital tool, setting out its key functionalities and implementation steps, including its integration into counsellors’ workflows and its role in guiding referrals to ALMPs and external services.
Strengthening Individualised Support for Jobseekers Furthest from the Labour Market in Greece
Abstract
Vulnerable jobseekers are individuals who face high barriers to employment and require intensive support to integrate into the labour market. These barriers can include limited work experience, long-term unemployment, low educational attainment, care responsibilities, disabilities, or other social disadvantages. In Greece, a substantial proportion of registered jobseekers falls into vulnerable categories, either as defined by legal frameworks (Law 4430/2016), or due to repeated or prolonged periods of unemployment. These individuals require personalised support to overcome specific employment barriers and to successfully transition into employment.
DYPA has implemented various strategies and measures aimed at identifying and supporting vulnerable jobseekers. Currently, DYPA relies on a profiling questionnaire, and predefined legal categories to determine the level of support required by each jobseeker. However, the current profiling approach has limitations: first, it covers only around 35% of registered jobseekers; and second, it misses a significant share of jobseekers in need of intensive support, while at the same time targeting some who might not need it. Additionally, despite offering a wide range of active labour market policies (ALMPs), provision is still limited, as only about 4% of registered jobseekers engage in ALMPs. Moreover, DYPA continues to face challenges in effectively engaging vulnerable and special-category jobseekers, who participate in ALMPs at even lower rates than other groups.
To address these challenges, this note proposes a new digital tool designed specifically to strengthen DYPA counsellors’ ability to identify, prioritise, and assist vulnerable jobseekers effectively. The proposed solution consists of a digital dashboard that leverages existing administrative data and advanced statistical methods to flag vulnerable jobseekers at the earliest stage possible. Furthermore, the tool supports counsellors by clearly highlighting individual employment barriers and providing recommendations for participation in ALMPs or referrals to relevant external social support services. Overall, the tool seeks to prevent prolonged unemployment spells and improve overall effectiveness in moving vulnerable clients toward employment.
This chapter is structured as follows: Section 3.2 identifies and describes the characteristics of vulnerable jobseekers in need of intensive support, using comprehensive administrative data and statistical analysis. Section 3.3 reviews the range of services DYPA currently provides to jobseekers, evaluating participation rates and identifying critical gaps. Section 3.4 details the proposed digital tool, outlining its key functionalities and advantages. Section 3.5 provides a practical implementation roadmap, detailing the essential steps for successful tool implementation, including stakeholder engagement, technical design, pilot testing, and national rollout. Section 3.6 concludes the chapter by summarising key findings and recommendations for DYPA to improve support for vulnerable jobseekers effectively.
Designing a more effective support system for vulnerable jobseekers requires the ability to systematically identify those who are most at risk of poor labour market outcomes. Setting up data-driven mechanisms for such identification processes rely on the availability of high-quality administrative data. The Greek Public Employment Service (DYPA) has developed multiple tools to identify those in need of more intensive support, including legal classifications of vulnerability and a profiling system that segments jobseekers relying primarily on self-declared questionnaire responses. Yet, as this chapter shows, these tools still fail to identify a considerable share of jobseekers who face substantial employment barriers and at times target others who might not have needed intensive help.
The analysis draws primarily on administrative data from DYPA and the Greek employment register (ERGANI). These sources contain data on jobseekers’ registration histories, demographic characteristics, participation in support programmes, and employment outcomes. These data allow for a nuanced assessment of who is currently flagged as vulnerable in DYPA’s systems and how well the profiling system aligns with actual employment trajectories. These data also reveal the limits of the current tools and pave the way for the development of complementary tools that can aid in the timely identification of jobseekers needing intensive support.
The chapter describes the dataset, explains how DYPA currently identifies vulnerable jobseekers, and highlights the strengths and limitations of the current system in accurately capturing the needs of those at risk. The chapter takes a data-driven approach to uncover the common characteristics of jobseekers who experience long-term unemployment or fail to exit to employment. The analytical approach identifies key predictors of labour market disadvantage and provides insights into which groups may require more targeted and intensive support.
3.2.1. Comprehensive administrative data provide the opportunity to identify vulnerable clients
A prerequisite for the development of a new data-driven digital tool to accurately identify jobseekers needing intensive support is the availability of detailed, high-quality individual-level data. Such data are needed to better understand which jobseeker characteristics are associated with poor employment outcomes, what kind of support is currently provided to vulnerable jobseekers, and ultimately, what can be done to provide more efficient support to the most vulnerable. The empirical analysis underlying the proposed tool draws primarily on administrative data from DYPA and ERGANI (Table 3.1). The datasets cover the period from January 2017 to January 2025 and provide a comprehensive overview of each jobseeker’s unemployment histories, service participation, and employment trajectories.
Note: ERGANI records do not include employment contracts for public servants and self-employment. Occupational programmes include programmes for the retention of jobs, public works, training via work experience, and subsidies for job creation.
The analysis presented in this note integrates these datasets using pseudonymised identifiers that protect individual privacy and still allow for the accurate linking of records across datasets. This integration supports robust statistical modelling to identify jobseekers needing intensive support based on their observable characteristics, past employment and unemployment histories, and their patterns of engagement with DYPA services.
The resulting database contains detailed information on approximately 3 million unique individuals and more than 15 million unemployment spells. It includes comprehensive records of approximately 300 000 participation spells in active labour market policies (ALMPs) involving nearly 269 000 jobseekers. To see details on data preparation, see Annex 3.A.
Three distinct data samples are defined in this report to meet the needs of the different research questions:
A snapshot sample including 950 192 active registrations as of 31 January 2025. This sample is used to capture detailed and up-to-date descriptive statistics of the scenario for jobseekers actively registered with DYPA at the time of data extraction. It provides an accurate picture of the current caseload, including its scale, composition, and key jobseekers’ characteristics.
A post‑2018 sample which includes about 5 million unemployment spells from approximately 2.2 million jobseekers who registered after the implementation of DYPA’s profiling system, and one year before the data extraction to allow for a one‑year follow-up period (between 12 November 2018 and 31 December 2023). While this sample covers a shorter time span, it offers a broader set of variables, including profiling information and participation in group job counselling. It is used when the analysis requires the richest set of individual-level data and serves as the basis for examining the association between different vulnerability indicators and labour market outcomes (Section 3.2), as well as jobseekers’ participation in ALMPs (Section 3.3).
An extended history sample comprises approximately 14.1 million unemployment spells for nearly 3 million individuals whose registrations began at least one year before data extraction. This sample is used to examine long-term unemployment dynamics, as it captures individuals whose unemployment began well before 2017 and continued through the observation period. Its longer timeframe enables analysis of the full duration of unemployment spells and transitions over time, but it can only be used when variables such as profiling scores or ALMP participation are not required.
3.2.2. DYPA currently takes a multi-faceted approach to identify those most in need of intensive support
DYPA relies on legal definitions and a profiling questionnaire to identify jobseekers who are in a vulnerable labour market situation and may require intensive and individualised support.
DYPA flags jobseekers with social and behavioural vulnerability
The legal definition (Law 4430/2016) of vulnerable social groups comprises “Vulnerable groups” and “Special groups”. Vulnerable groups include jobseekers whose access to employment is limited by physical, mental, or behavioural factors, specifically persons with disabilities, former addicts, young offenders, and former prisoners. Special categories encompass jobseekers who experience labour market difficulties due to their personal or social circumstances, including immigrants, refugees, asylum seekers, single parents, victims of domestic violence, homeless individuals, people from marginalised cultural groups (such as Roma or Pomaks), as well as other specific groups like victims of human trafficking, transgender individuals, and young adults exiting child protection facilities.
During registration with DYPA, jobseekers indicate if they belong to a special or vulnerable social group. Registration in these groups requires official supporting documentation, such as disability certification, rehabilitation certificates, or official documentation from relevant authorities. This information is recorded in DYPA’s register, and it is considered when assessing the level of support the jobseeker might need.
Vulnerable and special groups represented about 5.7% of jobseekers registered with DYPA on 31 January 2025. Among the vulnerable jobseekers, people with disabilities form the largest subgroup, followed by former prisoners and young offenders. Within special groups, immigrants are the largest group with 1% of the total jobseekers. Asylum seekers, refugees, Roma, and other special cultural groups also represent significant proportions.
Compared to other registered jobseekers, jobseekers from vulnerable and special groups have lower educational attainment. While only 27% of jobseekers not in a vulnerable or special group have completed up to primary education or lower, 31% of those in vulnerable groups have such low educational attainment (Table 3.2). Annex Table 3.B.1 shows a detailed breakdown of these figures by specific vulnerable and special category groups. Former prisoners, former addicts, and homeless individuals show especially low educational levels, with above 40% having completed only primary education.
Demographic characteristics and educational and occupational background for vulnerable and special groups, January 2025
Note: Education levels are based on the International Standard Classification of Education (ISCED), where Low = ISCED 0‑2, Medium = ISCED 3‑4, High = ISCED 5‑8. Occupations are classified into three skill levels based on the International Labour Organization (ILO) ISCO‑08 standard: High (levels 3‑4: managers, professionals, technicians), Medium (level 2: clerical, service/sales, agriculture, craft, machine operators), and Low (level 1: elementary occupations). Armed forces occupations are not classified. “Vulnerable” jobseekers are those identified by DYPA as those whose inclusion in social and economic life is hindered by physical or mental causes, or due to delinquent behaviour (e.g. people with disabilities, individuals with addiction problems, young offenders, prisoners, and former prisoners). “Special category” includes individuals registered under specific DYPA priority groups because of disadvantages in terms of access to the labour market (e.g. persons with disabilities, single parents, etc.) “Other jobseekers” includes all remaining individuals registered with DYPA and not falling into the previous two categories.
Source: OECD calculations based on data from the Greek public employment service (DYPA).
Vulnerable groups have also more often worked previously on lower-skilled occupations. For instance, 78% of Roma jobseekers worked previously in low-skilled occupations, compared to 25% among other, non-vulnerable jobseekers.
More importantly, jobseekers from special groups are more likely to have missing or incomplete data in DYPA’s records. For instance, while education data are available for 91% of jobseekers not identified as from vulnerable or special groups, this rate drops to 53% among those flagged as members of special groups. The lack of information on these groups might reflect the broader challenges that exist in reaching out to and providing suitable services for these groups.
DYPA uses a questionnaire and a rules-based profiling tool to categorise jobseekers
Since November 2018, DYPA has used a profiling tool to systematically categorise registered jobseekers into five groups based on their proximity to the labour market. The tool is rules-based and relies on two inputs.
The first input to the profiling tool is a score, derived from a questionnaire completed by the jobseeker at registration. This questionnaire includes around 30 items covering characteristics relevant for the labour market, such as work experience, language proficiency, and digital skills. After the completion of the questionnaire, jobseekers receive a score from 0 to 100. Jobseekers with high scores (over 85 points) are placed in category 1, indicating immediate job readiness. Lower scores correspond to progressively higher barriers and risk of prolonged unemployment. Those with scores between 65 and 85 are allocated to category 2, scores from 40 to 65 to category 3, and lower scores to category 4.
The second input uses information on whether a jobseeker belongs to a special social group. These groups align closely but are not necessarily the same as the vulnerable and special categories by the legal definition (described in the previous section). The criteria regarding special social groups override the questionnaire score and (i) automatically exclude vulnerable jobseekers from lower-risk categories (for instance, young offenders or those declaring multiple employment barriers cannot be placed in categories 1, 2, or 3); (ii) assign low-risk jobseekers, specifically those aiming to start their own business in category 1; and (iii) allocate to category 5 those in vulnerable and special-category groups facing significant employment barriers (e.g. former prisoners, homeless individuals, or those with low Greek language proficiency).
Based on these inputs, the tool generates a suggested profiling category. This suggestion is not final, however. The profiling process is only complete after a follow-up meeting with a DYPA job counsellor, who reviews the case and can confirm or modify the suggested category based on further assessment and discussion with the jobseeker.
Even though profiling is mandatory for first-time registrations,1 many jobseekers do not complete the questionnaire, and among those who do, a large share still do not complete the profiling process. Roughly 48% of jobseekers receive a suggested profiling score, and only 35% complete the full process and receive a final score revised by a job counsellor (Annex Table 3.B.2). This may be due to jobseekers not completing all steps or limited availability of DYPA counsellors to finalise the process.
Completion rates are particularly low among vulnerable populations who might be harder to reach, such as asylum seekers, young offenders, and former prisoners. On average, vulnerable and special category groups are more likely than other jobseekers to start the profiling process. However, there is a lot of heterogeneity in profiling completion rates among those who are identified as belonging to a vulnerable or special category group. While people with disabilities are more likely than the average to complete the whole profiling process and receive a final score from a job counsellor (62% vs. 35% for non-vulnerable jobseeker), only about 20% of refugees and asylum seekers end-up receiving a final profiling score. Young offenders, former prisoners, the homeless, and immigrants also have slightly lower completion rates than average. This uneven coverage hinders the profiling system’s potential to guide the provision of targeted support.
Even among those who complete the profiling process, the current profiling model used by DYPA does not appear to meaningfully differentiate between high-risk and low-risk jobseekers. First, the variation in the categorisation is very limited (Figure 3.1). Most jobseekers (51%) are placed in category 3, which reflects moderate risk, while only 16‑18% are assessed to be in either category 2 or 4. Fewer than 10% are profiled in the outer categories, either as job-ready (category 1) or as facing significant barriers to employment (category 5). Only the small proportion of jobseekers in vulnerable and special categories show a more dispersed distribution, with a similar or larger share assigned to the highest risk category (category 5) as to the middle‑risk category (category 3). Second, profiling does not appear well aligned with actual labour market outcomes. The profiling distribution among long-term unemployed jobseekers is not substantially different to that of those unemployed for less than 12 months, with 52% and 50% respectively placed in category 3.
Distribution of jobseekers by profiling category, post 2018 sample (2018‑2023)
Note: DYPA profiling categories range from 1 (job ready) to 5 (special groups with major barriers). “Vulnerable” jobseekers are those identified by DYPA as those whose inclusion in social and economic life is hindered by physical or mental causes, or due to delinquent behaviour (e.g. people with disabilities, individuals with addiction problems, young offenders, prisoners, and former prisoners). “Special category” includes individuals registered under specific DYPA priority groups because of disadvantages in terms of access to the labour market (e.g. persons with disabilities, single parents, etc.) The data includes all registrations for which profiling was carried out from 12 November 2018 to 31 December 2023.
Source: OECD calculations based on data from the Greek public employment service (DYPA).
DYPA’s profiling and vulnerability criteria align with outcomes but overlook some jobseekers who need support and target others who may not
DYPA’s current profiling categories correlate with jobseekers’ likelihood of finding employment, but the accuracy level could be improved because a fraction of jobseekers with higher support needs remains unidentified (Figure 3.2). Jobseekers assigned to lower-risk categories still frequently experience long-term unemployment – they remain registered with DYPA for longer than 12 months. For instance, around 20% of those classified as immediately ready for employment (category 1) and 30‑40% of those with low to moderate barriers to employment (categories 2 and 3) remain unemployed for at least 12 months. Meanwhile, 43% of registrations from jobseekers that have been classified as furthest from the labour market (category 5) end within 6 months.
Share of registrations by unemployment duration and profiling category, post 2018 sample (2018‑2023)
Note: DYPA profiling categories range from 1 (job ready) to 5 (special groups with major barriers). The data include all registrations for which profiling was carried out from 12 November 2018 to 31 December 2023.
Source: OECD calculations based on data from the Greek public employment service (DYPA).
The profiling categories broadly correlate with the probability of entering employment. The proportion of registrations ending because of entering employment decreases as the profiling category increases (Table 3.3). Still, despite the correlation, a significant proportion of registrations from jobseekers classified as low- or moderate‑risk (categories 1‑3) do not lead to entering employment nor active labour market measures, but end due to administrative reasons, inactivity, or simply remain ongoing. Similarly, a quarter of jobseekers classified as furthest from the labour market (category 5) enter employment within 12 months from registering.
For registrations from jobseekers who belong to some specific vulnerable social groups, such as homeless individuals, former prisoners, and former addicts, profiling categories tend to accurately reflect their significant labour-market barriers. Jobseekers in these groups consistently show poor employment outcomes (Table 3.4). For instance, only 19% of registrations from former prisoners lead to entering employment within 12 months of starting their registration, while 42% of those from jobseekers not classified as vulnerable do.
Share of registrations by exit type and profiling category, post 2018 sample (2018‑2023)
Note: DYPA profiling categories range from 1 (job ready) to 5 (special groups with major barriers). The data include all registrations for which profiling was carried out from 12 November 2018 to 31 December 2023. ALMP – Active labour market policy. Administrative exit includes exiting the DYPA registry because of jobseekers’ failure to renew their digital registration card, and other forms of non-compliance.
Source: OECD calculations based on data from the Greek public employment service (DYPA) and ERGANI.
Still, only one group, victims of domestic violence, enters employment faster than the average jobseeker. Moreover, these specific vulnerable social groups represent only a small fraction of the total population. The remaining 94% of registrations from jobseekers not formally classified as vulnerable or special category led to highly heterogeneous labour market outcomes and thus the predictive power of the profiling tool for long-term unemployment is modest.
Unemployment duration by type of vulnerability, extended history sample (2017‑2023)
Note: “Other special category groups” refer to the following groups: victims of human trafficking, persons who still reside in Child Protection and Care Units after reaching adulthood, transgender people and persons identified as vulnerable under an earlier DYPA classification. ALMP – Active labour market policy. Administrative exit includes exiting the registry because of jobseekers’ failure to renew their digital registration card, and other forms of non-compliance. The data include all registrations from 1 January 2017 and starting before 31 December 2023.
Source: OECD calculations based on data from the Greek public employment service (DYPA) and ERGANI.
Taken together, these findings point to a substantial under-identification of employment barriers among registered jobseekers. The current classification systems miss key information about jobseekers’ real prospects in the labour market. More importantly, it demonstrates that there is a need for bringing in an additional lens, in particular, one that incorporates indicators based on labour market outcomes to better identify jobseekers needing intensive support. While profiling categories and formal vulnerability classifications are correlated with labour market outcomes, they are not sufficient to accurately identify jobseekers needing intensive support. On the one hand, a significant number of jobseekers not classified as high risk, either by the profiling tool or by not belonging to any vulnerable or special group, experience long-term unemployment or exit the registry without finding a job. On the other hand, some jobseekers categorised as high risk or vulnerable manage to transition into employment relatively quickly.
3.2.3. Jobseekers needing intensive support share some common features
This section examines labour market outcomes to identify jobseekers needing intensive support. More specifically, the section uses statistical methods to detect the observable characteristics most strongly associated with poor labour market outcomes.
A data-driven approach to identify key risk factors of poor labour market outcomes
Many OECD countries use statistical profiling to assess jobseekers’ likelihood of becoming long-term unemployed. These tools typically rely on administrative and survey data to identify patterns in jobseekers’ characteristics and labour market outcomes. Several countries, including Austria, Denmark, Italy, Sweden and the Netherlands have developed models based on Logistic regression, Random Forests, or other statistical techniques to rank jobseekers by risk and guide resource allocation (Desiere, Langenbucher and Struyven, 2019[1]). Some of these countries, such as the Netherlands and Estonia, also highlight to the counsellors those jobseeker characteristics that are contributing to the profiling result the most.
Somewhat similarly to the international examples of statistically profiling jobseekers, the analysis in this section uses statistical methods to find the most common observable characteristics across jobseekers that remain unemployed for more than 12 months.
The analysis uses the extended history sample (see Section 3.2.1 for details). The statistical approach is carried out in two steps. First, a data-driven approach identifies the observable characteristics most strongly associated with long-term unemployment (LTU), from now on referred to as risk factors. Second, survival curves are used to illustrate the effect of a selection of the risk factors on the probability of remaining unemployed over time.
Step one corresponds to a random forest classifier; a method used to predict outcomes and identify the strongest predictive factors. The main advantages of this approach are that it yields more accurate predictions than other common statistical approaches, and that the decisions of which factors to consider to be relevant are made in a data-driven way. In this application, the approach can point to the jobseekers’ characteristics, such as age, education, or employment history, that best predict the likelihood of remaining unemployed.
The second step of the analysis plots Kaplan-Meier survival curves based on the key risk factors identified by the random forest model. The curves provide an intuitive visualisation of how key characteristics influence the likelihood of remaining unemployed over time. More importantly, the method allows for comparing curves for different groups (for example, by age, gender, or education). Thus, it is possible to use the key predictors identified in step 1 to split jobseekers into groups and easily visualise the effects of the key predictors by comparing which groups tend to remain unemployed longer.
For additional technical details on the statistical methods, see Box 3.1
Random forest approach identifies key risk factors for poor labour market outcomes
Random forest is a data-driven method used to predict an outcome and identify the most influential factors contributing to that particular outcome. The method combines multiple simpler models, called decision trees, and aggregates their outputs to improve predictive accuracy (Breiman, 2001[2]).
A decision tree consists of a sequence of criteria that, at each step, separate individuals into two distinct groups based on their characteristics. For example, the initial criterion might split the group of jobseekers based on age, creating one group aged up to 50 years, and another group older than 50 years. Each resulting subgroup can then be further divided into two using additional criteria like gender or family status. This iterative splitting process continues until distinct groups emerge that clearly differentiate jobseekers who are likely to remain unemployed in the long-term from those who are likely to re‑enter employment quickly (or distinguish across groups with other labour market outcomes of interest).
This sequence of criteria can be used to predict if a newly registered jobseeker is likely to be long-term unemployed or not, or if they are likely to exit the registry to enter employment. Moreover, it also informs on what are the characteristics that are most predictive of the outcome. However, single decision trees have an important drawback, they are not robust to minor changes in data, and using a slightly different dataset may alter both, the sequence of criteria, and the predictions.
Random forests address this issue by constructing multiple decision trees, each based on different subsets of jobseekers and characteristics. The individual predictions of all trees are then aggregated, resulting in more stable and reliable overall predictions. Consequently, random forests provide a robust framework for consistently identifying the risk-factors of the labour market outcomes of interest.
Furthermore, random forests yield a straightforward measure of each risk-factor’s importance. Predictors with higher importance scores are those that consistently contribute to clearly distinguishing between different risk groups across the multiple decision trees. In turn, these key predictors can inform the construction of survival curves, to visually represent how various jobseeker traits affect the duration of unemployment or the likelihood of entering employment.
Kaplan-Meier survival curves illustrate the effect of key risk factors on unemployment duration
The Kaplan-Meier estimator is a statistical method used to measure how likely it is for individuals to remain in a specific state over time, including staying unemployed (Kaplan and Meier, 1958[3]). In the context of unemployment, Kaplan-Meier curves can show the share of jobseekers who remain unemployed after specific intervals (e.g. 3, 6, or 12 months). The curves start at 100% (at registration) and decline in a stepwise manner each time someone leaves the registry (Meyer, 1990[4]).
This method allows for straightforward comparisons between different groups defined by risk factors identified by the random forest method. Thus, survival curves provide an intuitive visual representation of how specific characteristics influence unemployment duration, clarifying which jobseekers typically face the greatest risk of prolonged unemployment.
Key predictors of long-term unemployment appear across demographic characteristics, skills, and labour market history
In a preliminary analysis, the Random Forest method identifies labour market history, demographic characteristics, jobseekers’ education, and their previous occupation as key predictors of long-term unemployment. Figure 3.3 presents key predictors in order of importance in estimating the risk.2
Information relating to labour market history ranks among the most important features. This information includes days since last registration, number of previous registered unemployment spells, number of prior long-term unemployment spells, and days in employment in the period right before the registration start. Demographic data follow second in importance, with age at registration, marital status, and gender, among the most important characteristics. Indicators of educational attainment and occupation-related variables, including preferred and previous occupation (ISCO category), also consistently show high predictive value.
Relative importance of key predictors of long-term unemployment, extended history sample (2017‑2023)
Note: Key predictors derived from the Random Forest. Variable importance scores are calculated based on how much each variable contributes to accurately distinguishing jobseekers who become long-term unemployed from those who do not. The scores are normalised so that the highest-ranking predictor is assigned an importance of 1, and all other predictors’ importance values are scaled accordingly. The data include all registration from 1 January 2017 and starting before 31 December 2023.
Source: OECD calculations based on data from the Greek public employment service (DYPA) and ERGANI.
Survival curves help to illustrate the impact of these risk factors over time. In these figures, the “start time” is the point when each jobseeker registers as unemployed, so everyone begins with a probability of being unemployed equal to 1. Then, the curve shows how this probability decreases as jobseekers in each group find employment.
Overall, the analysis shows that jobseekers in need of intensive support (those with poor employment prospects) tend to share common characteristics across three dimensions: demographics, skills, and labour market history. Figure 3.4 shows survival curves for factors associated to demographic characteristics and educational background, and Figure 3.5 focuses on labour market history.
Probability of remaining unemployed over time (Kaplan-Meier estimates), extended history sample (2017‑2023)
Note: Time measured in months since registration with DYPA as unemployed. Estimates account for right-censoring (individuals still unemployed at the end of the observation period). Panel (c) categories “Married and male” and “Married and female” include married and cohabitating jobseekers. Panel (d), Education levels are based on the International Standard Classification of Education (ISCED), where Low = ISCED 0‑2, Medium = ISCED 3‑4, High = ISCED 5‑8
Source: OECD calculations based on data from the Greek public employment service (DYPA).
Demographics such as age, gender, and marital status affect the probability of long-term unemployment. For example, the curve for jobseekers aged over 50 declines more slowly than for younger groups, indicating that older jobseekers remain unemployed for longer on average (Figure 3.4). Additionally, women, especially married women, also face notably higher risks of LTU compared to their male counterparts.
Educational attainment and occupational background also shape unemployment duration. Jobseekers with low to mid-level educational qualifications are at higher risk of LTU than those with higher education. Similarly, jobseekers pursuing employment in elementary occupations have a higher probability of remaining unemployed after 12 months of registration than those looking for work in sales and services.
Labour market history is another strong predictor of LTU (Figure 3.5). Even a single previous spell of long-term unemployment considerably increases the likelihood of future long-term unemployment. Additionally, jobseekers whose previous registrations ended between 6 to 12 months before the current unemployment spell, have significantly lower probability of staying unemployed at 12 months since registration start than those with more than 12 months between spells. This effect might be driven by jobseekers engaging in seasonal and tourism-related work who are less likely to spend both a full year of uninterrupted work, and a full year of uninterrupted unemployment.
The reason for exiting previous employment also matters. Jobseekers whose past employment ended due to dismissal exhibit notably higher risks of staying unemployed for longer than 12 months, whereas those whose employment concluded due to the expiration of a fixed-term contract are less prone to prolonged unemployment spells. Additionally, limited or no employment in the six months prior to registration considerably increases the risk of experiencing LTU.
Probability of remaining unemployed over time (Kaplan-Meier estimates), extended history sample (2017‑2023)
Note: This figure shows Kaplan-Meier survival estimates of the probability of remaining unemployed over time (in months) for different categories of jobseekers registered with DYPA. Time measured in months since registration with DYPA. Estimates account for right-censoring (individuals still unemployed at the end of the observation period). Panel (c) Contract end refers to the end of a fixed term contract. Panel (f) category “Other” includes all other categories from the International Standard Classification of Occupations (ISCO): Managers; Professionals; Technicians and Associate Professionals; Clerical Support Workers; Skilled Agricultural, Forestry and Fishery Workers; Craft and Related Trades Workers; Plant and Machine Operators and Assemblers; and Armed Forces.
Source: OECD calculations based on data from the Greek public employment service (DYPA).
This chapter describes DYPA’s active labour market policies (ALMPs) and analyses the extent of their use by different groups of jobseekers. ALMPs are key to support jobseekers, especially those facing significant barriers to employment. Given that programmes can be costly, it is important that this type of support reaches those who need it the most. The chapter first describes DYPA’s existing ALMPs, presenting detailed administrative data on programme characteristics and participation patterns. It then investigates the extent to which these programmes are reaching jobseekers needing intensive support according to three criteria: Belonging to vulnerable or special groups according to the legal definition, DYPA’s profiling categories, and experience with long-term unemployment.
The analysis combines descriptive statistics that show how participants in DYPA’s services differ from the general group of jobseekers, with logistic regression models that calculate the probability of participation in ALMPs by group, while controlling for other demographic, geographic, and time factors. By doing so, the analysis quantifies more rigorously the relationship between vulnerability and service uptake, which can provide helpful insights into whether and how effectively DYPA’s interventions target jobseekers needing intensive support.
3.3.1. DYPA provides a wide array of active labour market policies to support jobseekers
To better understand how different types of employment support are used by jobseekers, the analysis draws on detailed administrative data from DYPA on participation in ALMPs. The dataset covers participation in ALMPs for all jobseekers registered as unemployed at any point between January 2017 and January 2025 (see Table 3.1 for details). ALMPs included in the data can be broadly described as falling into two categories: (i) occupational programmes, which include programmes for the retention of jobs, public works, training via work experience, and subsidies for job creation; and (ii) in-house group job counselling services.3 More specifically, the data on DYPA’s ALMPs cover:
Employment incentives, such as wage subsidies, aiming to promote job retention or encourage employers to create new job opportunities (ΔΙΑΤΗΡΗΣΗΣ ΝΘΕ), with 14 programmes involving 36 317 participants between November 2015 and November 2018, and subsidies for creating new job positions (ΝΕΩΝ ΘΕΣΕΩΝ ΕΡΓΑΣΙΑΣ), comprising 48 programmes that supported 119 258 participants between October 2014 and March 2023. Some programmes specifically target the LTU, such as the “Subsidy Programme for Businesses to Employ 10 000 Long-Term Unemployed Aged 45 and Over, in High-Unemployment Areas – RRF (De Minimis Regulation 1 407/2013)”
Direct job creation initiatives (ΕΠΙΔΟΤΟΥΜΕΝΗ ΑΠΑΣΧΟΛΗΣΗ). This category involves 25 programmes with 84 257 participants from February 2012 until October 2022. Major examples include programmes such as “Public Works Programme in Municipalities, Regions and Social Welfare Centres for 30 333 Job Positions”.
Training, including work experience programmes (ΕΡΓΑΣΙΑΚΗΣ ΕΜΠΕΙΡΙΑΣ), such as on-the‑job training placements. DYPA offered 22 such programmes, which engaged a total of 66 140 participants since August 2016.
Labour market services, which include group counselling sessions and thematic hands-on workshops and seminars led by DYPA counsellors. Group counselling activities address practical aspects of job search and career development, such as preparing an effective CV, identifying and presenting skills, building a professional profile, using social networks for job search, preparing for interviews, and developing business ideas and business plans. Hands-on Job Search Workshops complement these activities by featuring guest speakers from companies and HR departments who share insights on recruitment practices and candidate expectations. Workshops can also focus on success stories and peer-to-peer exchanges, where entrepreneurs supported by DYPA present their experiences and lessons learned. In addition, during these events a “CV Corner” is set up to provide one-to-one advice.
3.3.2. Participation in ALMPs varies across different groups of jobseekers
ALMPs are varied, and some have the specific purpose of targeting vulnerable groups. Still, analysing ALMP participation by jobseeker characteristics can help better understand how much support different vulnerable jobseekers receive.
The statistical approach used to examine participation in DYPA’s ALMPs combines descriptive analysis and a logistic regression approach to assess how three vulnerability criteria (vulnerable and special social groups, profiling category, and long-term unemployment) relate to programme participation. The analysis uses the post‑2018 sample, this timeframe covers the period for which group job counselling and profiling score data are available, and it also allows for a follow-up period of at least one year after registration with DYPA (see Section 3.2.1 for details on the data).
The analysis is carried out at jobseeker level, thus, when one jobseeker has multiple registrations, only the latest available registration is used, whether ongoing or concluded. Both parts of the analysis distinguish between ex-post labour market outcomes such as the jobseekers currently being long-term unemployed, and ex-ante risk factors, such as having prior LTU spells, and time in employment before registering.
Still, it is important to note that differences in ALMP participation across different groups may reflect both differences in programme targeting by DYPA and variations in jobseekers’ willingness or ability to participate. Neither the descriptive statistics nor the regression analysis can fully disentangle these two underlying mechanisms, and the two sets of results should therefore be interpreted with this limitation in mind.
Summary statistics suggest that the most vulnerable often miss out on support
Descriptive statistics are used to compare the demographic and educational characteristics of ALMP participants with those of the overall population of registered jobseekers. This is informative as it highlights which groups are over-represented among ALMP participants relative to their share in the overall jobseeker population. In addition, the analysis examines the share of ALMP participants who fall under three vulnerability criteria. The first two criteria are those currently used by DYPA, i.e. (i) whether the jobseeker is, or has been, classified in DYPA’s vulnerable or special groups; and (ii) the jobseeker’s profiling category, as determined by the profiling tool. The third criterion is labour market history, in particular, whether the jobseeker is currently, or has previously been, long-term unemployed. The participation shares are then compared with the corresponding rates among all registered jobseekers.
Participation in DYPA’s ALMPs, including occupational programmes and group job counselling, remains low overall. Only 4.5% of all registered jobseekers engage in these services. Participation in occupational programmes and group job counselling differs remarkably by socio‑economic characteristics of jobseekers (Figure 3.6). Nevertheless, jobseekers that have characteristics of vulnerability are often less likely to participate in ALMPs.
Share of jobseekers by key demographic and vulnerability characteristics among ALMP participants and among all jobseekers, post 2018 sample (2018‑2023)
Note: Education levels are based on the International Standard Classification of Education (ISCED), where Low = ISCED 0‑2, Middle = ISCED 3‑4, High = ISCED 5‑8. Vulnerable social groups correspond to all jobseekers that fall into one of the categories outlined in Law 4430/2016. DYPA profiling categories range from 1 (job ready) to 5 (special groups with major barriers). The data includes the latest registration per jobseeker considering the window from 12 November 2018 to 31 December 2023.
Source: OECD calculations based on data from the Greek public employment service (DYPA).
Although older and lower-educated jobseekers are identified in the previous chapter as most in need of intensive support, younger jobseekers (under 30 years old) are overrepresented in both interventions (39% in both ALMPs vs. 26% of youth among all jobseekers). Participants are also more likely to hold higher education qualifications, especially in job counselling, where 41% have tertiary education compared to just 23% among all registered jobseekers. Yet, women, who have more difficulties to access employment, are slightly overrepresented among participants – 57% in occupational programmes and 65% in counselling, while the share of women among all jobseekers is 55%.
DYPA’s vulnerable or special category groups, and those in higher profiling categories are underrepresented in ALMP participation. Only 2% of jobseekers in occupational programmes and 3% in counselling services are classified as vulnerable or special category, while representing 5% of the jobseeker population overall. Similarly for profiling categories, jobseekers in category 3 are especially overrepresented in ALMPs (60% in occupational programmes and 49% in job counselling services). In contrast, jobseekers in category 5, those with the greatest distance from the labour market, are underrepresented in both types of interventions (1‑2% of participants).
Finally, jobseekers whose current (or most recent) spell is classified as LTU are more prevalent among participants in counselling services (78% compared to 41% of all jobseekers). However, the gap is much narrower when considering previous LTU experience: 54% of counselling participants have experienced one or more previous LTU spells, compared to 47% of all jobseekers. In contrast, jobseekers who were recently employed, that is, who had some employment in the six months prior to their registration are more likely to participate in ALMPs than those with no recent employment, even though recent employment is associated with a lower risk of long-term unemployment.
Regression analysis corroborates that most vulnerable jobseekers are less likely to participate in ALMPs, even when controlling for other factors
Three logistic regression models are used to examine the relationship between the different vulnerability criteria and ALMP participation. The approach estimates the likelihood of programme participation by group while accounting for other relevant characteristics. The dependent variable in all three regressions is defined as participation in any of DYPA’s ALMPs (see Section 3.1 for programme list). The regressions differ in their primary explanatory variables:
Regression 1: Includes indicators on whether the jobseeker belongs to a vulnerable social group
Regression 2: Includes indicators of DYPA’s profiling categories
Regression 3: Focuses on labour market history indicators, specifically it includes indicators for current and previous LTU spells, and recent employment.
To ensure robust estimates, all three regressions include controls for gender, age at registration, education level, and fixed effects for the jobseeker’s region and year of registration. These controls help isolate the relationship between vulnerability and ALMP participation from potential confounding demographic or geographic factors, as well as time effects.
Figure 3.7 presents predicted probabilities of participation based on each regression, with their corresponding 95% confidence intervals. The estimates from Regression 1 (at the top of Figure 3.7) show that the predicted participation probability for vulnerable groups is 4%, compared to 5% for non-vulnerable groups. Although modest, this difference is statistically significant, demonstrating that jobseekers in vulnerable and special categories are significantly less likely to participate in DYPA’s ALMPs even when controlling for relevant demographics, and regional and time effects.
Similarly, results from Regression 2 (middle of Figure 3.7) show that there is a misalignment between profiling categories, which should identify the level of need, and service delivery: jobseekers assessed as most in need of support are less likely to receive it. Those in categories 4 and 5 (groups facing major barriers to employment) have lower probability of participation than those in categories 2 and 3 (jobseekers at low and moderate risk of long-term unemployment). The predicted participation rates are 8% for category 4 and 4% for category 5, compared to 11% for categories 2 and 3.
Finally, estimates from the Regression 3 (bottom of Figure 3.7) suggest that current long-term unemployed jobseekers are slightly more likely to participate in ALMPs than those with shorter unemployment spells: 5% compared to 4%. Similarly, jobseekers who have previously experienced long-term unemployment are 1 percentage point more likely to participate in ALMPs than those who have not (5% and 4% respectively). Finally, those with more days of employment in the six months preceding their registration have similar participation rates to those with minimal or no recent employment experience (5% and 4% respectively).
Estimated probabilities of participation by group using three logit models, post 2018 sample (2018‑2023)
Note: The figure shows the predicted probability of participation in ALMPs for each vulnerability group based on three logistic regression analysis (one for each vulnerability criteria). Vulnerability estimates (at the top) correspond to Regression 1, profiling estimates (middle) correspond to Regression 2, and Labour market history estimates (Bottom) correspond to Regression 3. Vulnerable social groups correspond to all jobseekers that fall into one of the categories outlined in Law 4430/2016. DYPA profiling categories range from 1 (job ready) to 5 (special groups with major barriers). The analysis includes controls for demographic characteristics (gender, age, education level), regional effects, and year of registration. All controls are held at their average for the calculation of participation probability. The data includes the latest registration per jobseeker considering the window from 12 November 2018 to 31 December 2023.
Source: OECD calculations based on data from the Greek public employment service (DYPA).
Although the analysis does not fully disentangle the mechanisms driving ALMP participation, one potential explanation is that LTU jobseekers, by definition, remain unemployed for longer periods and therefore have more time to engage with available programmes. The findings also suggest that some of the higher participation among LTU jobseekers results from targeted programme interventions aimed specifically at individuals already experiencing prolonged unemployment. Such targeted efforts are important; however, they happen only after jobseekers have faced adverse labour market outcomes. Combining these efforts with the targeting of jobseekers at risk of LTU at an earlier stage may help to prevent such negative outcomes from occurring in the first place.
Informed by the analysis of microdata presented in Chapters 2 and 3, this chapter proposes a concept for a new digital tool designed to strengthen DYPA counsellors’ capacity to identify, prioritise and support vulnerable jobseekers. The concept describes and discusses a range of potential functionalities that the new digital tool could offer along with the key steps DYPA would need to take to implement these functionalities effectively.
3.4.1. The new digital tool will pursue three key objectives
The aim of the new digital tool is to support DYPA’s employment counsellors in assisting vulnerable groups. The tool will help counsellors identify more effectively those jobseekers who are furthest from the labour market, prioritise them to ensure that time and resources are focused on those who need them the most, and provide a means for counsellors to better understand each jobseeker’s situation and propose appropriate actions. Importantly, the concept for the tool builds on three key findings from the analysis of microdata in Chapters 2 and 3 and the note on potential new digital tools (Chapter 2):
DYPA’s current profiling approach is effective in identifying some of the most vulnerable jobseekers. However, Chapter 2 shows that some jobseekers who are far from the labour market still slip through the net. While DYPA is in the process of reviewing its profiling approach, the exact design of the revised system is yet to be defined. Regardless of the approach ultimately adopted, early identification of vulnerable jobseekers is essential to ensure that they receive timely and appropriate support.
Individualised counselling combined with tailored support is essential to help vulnerable jobseekers move into employment. While profiling aims to identify those furthest from the labour market, it does not provide direct insights into the specific employment barriers that a jobseeker faces.
Chapter 3 shows that participation in ALMPs provided by DYPA is low among jobseekers needing intensive support. Moreover, supporting vulnerable jobseekers often requires a holistic approach that goes beyond the services DYPA provides. Close co‑operation with other institutions and organisations that offer support to vulnerable groups is therefore essential.
Building on these three findings, the proposed concept for the new digital tool outlines functionalities in three key areas to help close gaps in the support provided to vulnerable groups: improving the identification of vulnerable jobseekers, supporting individualised counselling, and facilitating referrals to relevant measures. While the technical specifications of the new digital tool fall outside the scope of this project, the tool could take the form of a dashboard designed for employment counsellors and it could be integrated into the existing interface that counsellors use to support their interactions with jobseekers.
3.4.2. The tool will enhance the early identification of vulnerable jobseekers
The tool will support the early identification of vulnerable jobseekers by flagging jobseekers needing intensive support and enabling counsellors to focus their time and resources on those most in need, while reducing deadweight costs linked to providing services to individuals who are likely to find employment without additional support. Importantly, this process should take place as early as possible, ideally at the point of registration, so that counsellors can prioritise meetings with vulnerable jobseekers and provide timely support. Early intervention is indeed crucial to improving outcomes for vulnerable jobseekers.
In practice, the tool could flag newly registered jobseekers needing intensive support and suggest meeting times based on counsellor availability. Counsellors would then review these suggestions and either confirm the proposed meeting or amend it.
The profiling process can be strengthened through better use of data
As described in Chapter 2, since November 2018 DYPA has used a rules-based profiling tool to assess the employability of jobseekers. The current tool relies on information gathered through a profiling questionnaire but does not apply statistical profiling, which would make use of historical jobseeker data to understand how individual jobseeker characteristics are linked to labour market outcomes.
A recent DYPA project with the World Bank assessed the existing profiling tool and proposed a new statistical profiling model that uses machine learning techniques and jobseeker data to estimate the likelihood of long-term unemployment (World Bank, 2024[5]). At the time of writing this chapter, DYPA’s plans regarding the future design of its profiling tool are still to be determined. To complement the work undertaken by DYPA and the World Bank, the concept for the new digital tool developed under the OECD-DYPA project treats the underlying profiling model as a given. This approach ensures that the proposed concept can be implemented irrespective of whether DYPA chooses to maintain its current profiling model or adopt the new statistical profiling model proposed by the World Bank.
If DYPA decides to maintain its current profiling system, it could be augmented by integrating information on previous unemployment spells. As shown in the analysis in Chapter 2, individuals who have experienced periods of LTU in the past are significantly more likely to become long-term unemployed again. This information is readily available in DYPA’s Integrated Information System and could be retrieved for any new jobseekers registering with DYPA. In addition, DYPA could integrate information on jobseekers’ previous employment histories from ERGANI, which, as shown in the analysis in Chapter 2, is also a strong predictor of distance from the labour market. To capture transitions into employment more accurately, DYPA could consider establishing systematic data exchanges – similar to the one currently in place with ERGANI – with other institutions such as the Greek Social Security EFKA. This would allow DYPA to observe the transition of jobseekers into self-employment, which is not recorded in ERGANI. Incorporating these data would enhance the accuracy of profiling.
Regardless of how DYPA decides to move forward regarding the profiling methodology, the profiling process could be carried out at the point of registration, with the results used to prioritise meetings between jobseekers and counsellors. While this approach is already applied at DYPA to some extent, the process could be further strengthened and streamlined. In particular, the tool could support counsellors in prioritising clients in the least job-ready groups, while postponing meetings with carefully selected jobseekers who are closer to the labour market and likely to find employment without immediate support from a counsellor.
Profiling must comply with data protection regulations
When processing personal data, such as those used for profiling, it is essential to comply with fundamental data protection principles. The EU General Data Protection Regulation (GDPR) governs the use of personal data, including for profiling purposes (Article 22). For example, the principle of fair and transparent processing requires that individuals be informed about the existence and implications of profiling whenever it is applied. More broadly, individuals have the right not to be subject to decisions based solely on automated processing, without human intervention.
This implies that if profiling results are automatically used to schedule first meetings with counsellors, jobseekers should be informed and given the option to opt out. Alternatively, or in addition, the automated meeting prioritisation generated by the tool should be reviewed and, where necessary, adjusted by a counsellor. As such, the decision on how to prioritise meetings with counsellors would not be fully automated but would involve human oversight.
Face‑to-face interactions remain essential
While statistical profiling can help DYPA identify vulnerable jobseekers and prioritise those furthest from the labour market more efficiently, it is essential to emphasise that face‑to-face interactions remain critical for supporting vulnerable clients (OECD, 2021[6]). Many vulnerable jobseekers may lack the digital skills required to complete online processes (e.g. digital registration) or may not have access to the necessary digital infrastructure (e.g. internet connection, computer, or smartphone) to complete such steps independently.
As discussed in Chapter 2, although almost all jobseekers begin the profiling process, only roughly one in two receives a suggested profiling score, and only one in three obtains a final profiling score validated by a counsellor. While the completion rates are higher among vulnerable groups, substantial variation exists across specific subgroups. For example, over 60% of persons with disabilities and heads of single‑parent families obtain a final profiling score, compared with only 22% of young offenders and 38% of homeless individuals. Although different factors can explain these low completion rates, they may partly reflect low digital skills and limited access to digital infrastructure among some groups. This highlights the importance of maintaining the option for jobseekers to register and complete the profiling questionnaire in person. Doing so ensures that DYPA’s services remain accessible to all vulnerable jobseekers, including those facing digital barriers and who are more difficult to reach through digital channels.
3.4.3. The tool will support individualised counselling by highlighting jobseeker-specific risk factors and employment barriers
Profiling tools are effective in identifying jobseekers who are furthest from the labour market, but they generally do not enable counsellors to gain a deeper understanding of the underlying risk factors (or combinations of factors) that contribute to a jobseeker’s profiling score, nor of the employment barriers they face.
To support tailored and personalised counselling, which is essential to assist vulnerable groups, the new digital tool could provide in a dashboard a summary of the individual risk factors contributing to each jobseeker’s profiling score along with their relative importance and the employment barriers associated with these factors. This dashboard would help counsellors gain a better understanding of the specific challenges each jobseeker faces and manage their one‑to‑one meetings with jobseekers more effectively.
Other PES across the OECD are already using similar dashboards. For example, counsellors in the Estonian PES employ the decision-support tool OTT, which allows them to view not only the profiling score (i.e. the estimated probability of moving into employment) but also the factors influencing this probability and the direction of their impact. In the example below, a low vacancy-to‑unemployment ratio in the occupation in which the jobseeker is seeking employment in is a factor that decreases the probability of employment, while a high level of digital skills increases this probability.
Decision Support Tool OTT in Estonia
Source: (Leinuste, 2025[7]), Work-focussed counselling and Decision Support Tool.
There are various ways for DYPA to assess how individual jobseeker risk factors (or combinations of factors) influence the probability of moving into employment, each requiring different levels of resources and depending on DYPA’s future plans for its profiling approach. First, DYPA could directly draw on the results presented in this report. Table 3.5 summarises the factors that according to the analysis in Chapter 2 are most predictive of the probability of long-term unemployment and indicates the direction of their impact. Based on this information, the new digital tool could be programmed to highlight key risk factors for each jobseeker, such as age or limited work experience, allowing counsellors to better visualise which factors are reducing and which factors are increasing the jobseekers’ chances of exiting unemployment. While this option can be implemented by DYPA immediately based on the findings of the current reports, it would not fully exploit the potential of DYPA’s administrative data, unlike the implementation of a fully-fledged statistical profiling model.
In the short to medium term, a relatively simple and cost-effective way for DYPA to make better use of available data would be to apply the approach and methodology proposed in this report, incorporating additional risk factors. Specifically, data collected through the profiling questionnaire such as self-assessed language skills, digital skills, and motivation to search for a job could be used. These factors are typically strong predictors of distance from the labour market. Importantly, they are also malleable and can be influenced through interventions such as training or group counselling. Making these factors visible to employment counsellors could help inform decisions on appropriate ALMPs for each jobseeker.
Impact of the most predictive long-term unemployment risk factors
LTU: long-term unemployment.
Note: The risk factors are selected using the results presented in Figure 3.3. The direction of the impact of each factor is determined based on Figure 3.4 and Figure 3.5.
In the medium to long term, DYPA could choose to augment and progressively expand the approach proposed in this report. This would enable DYPA to independently determine which data to use and to regularly update the model to reflect changing labour market conditions. To extend the methodology described in this report, DYPA should take into account three key aspects. First, DYPA would need to decide which data sources to use. Second, DYPA would need to analyse the data to identify the factors most predictive of employment outcomes (or other outcomes, such as the risk of long-term unemployment). Third, DYPA would need to assess the direction of the impact of each risk factor. Importantly, each of these steps may depend on the overall profiling model adopted by DYPA. The following sections provide a more detailed discussion of each step and the critical aspects to consider. This will enable DYPA – should it choose to revise the overall profiling model – to ensure that any new profiling approach meets the specific needs of vulnerable jobseekers and the employment counsellors who support them.
Choosing the set of data to use
The first step is to identify the jobseeker outcomes that best reflect distance from the labour market. Many statistical profiling tools used by PES across OECD countries focus on the probability of long-term unemployment, although the definition of long-term unemployment varies between countries. For example, Australia, Italy and the Netherlands use a threshold of more than 12 months of unemployment; Belgium and Sweden use more than 6 months; and Denmark uses more than 26 weeks. These outcomes form the basis of what the statistical profiling tool is designed to predict.
While unemployment duration is a useful indicator of a jobseeker’s distance from the labour market, exiting unemployment does not always mean a successful labour market integration. For example, some jobseekers may leave unemployment and move into inactivity rather than employment. For this reason, some countries use alternative outcomes, such as the probability of entering unsubsidised employment (e.g. Austria) or the probability of transitioning from unemployment to employment within 12 months (e.g. Ireland). However, selecting such outcomes requires linking unemployment data with employment data.
As different outcomes capture distinct aspects of labour market integration, some countries use multiple outcomes in their profiling models to obtain a more comprehensive picture of a jobseeker’s distance from the labour market. For example, as noted earlier, the Estonian PES uses both the probability of moving into employment and the probability of returning to unemployment (i.e. re‑registering with the PES). This approach takes into account not only how quickly a jobseeker finds employment but also the sustainability of that employment. Similarly, Luxembourg has piloted statistical profiling using three distinct classification models, estimating the probability of reemployment within 3, 6, and 12 months.
DYPA already shares some data with the employment register ERGANI through a dedicated web service, providing a solid foundation for accessing data on jobseekers’ employment outcomes. Alternatively, DYPA could rely solely on its own data, for example by using the probability of becoming long-term unemployed or the probability of returning to unemployment – both of which are already available within DYPA’s information systems. While these measures may not fully capture successful labour market integration, they are readily available and do not require linking to external data sources.
Once the outcomes have been defined, DYPA should identify the set of risk factors that will be put in relation to the chosen outcomes to determine distance from the labour. These factors typically include (Desiere, Langenbucher and Struyven, 2019[1]; Immervoll and Scarpetta, 2012[8]):
Demographic characteristics of jobseekers (e.g. age and gender)
Motivation to search for or accept a job (e.g. willingness to take up employment, wage expectations)
Job readiness, including education, skills, disabilities, and care responsibilities
Opportunities (e.g. regional labour market conditions)
Some of these data, such as demographic characteristics and indicators of job readiness, are available in DYPA’s registries and were shared for the purposes of this project. Other information, particularly on behavioural aspects such as motivation to look for work or willingness to start employment, is collected through the profiling questionnaire. While these behavioural data were not included in the data shared for this project, DYPA could consider continuing using them for profiling purposes.
Determine which factors are associated with distance from the labour market
To analyse how individual risk factors are associated with jobseekers’ outcomes, DYPA could apply a range of analytical approaches. One approach would be to use machine learning-based models, such as random forests. The project carried out by the World Bank in collaboration with DYPA proposes such a model for profiling jobseekers. These models are particularly well-suited for this purpose, as they can handle large sets of variables and capture complex, non-linear relationships between risk factors and outcomes in a highly flexible way. However, they are also more computationally intensive than traditional models. In addition, random forests operate as a kind of “black box”: while they produce an estimate of a jobseeker’s distance from the labour market, they do not directly show how individual risk factors affects this estimate.
If DYPA decides to revise its overall profiling model and adopt the approach proposed by the World Bank, it could use the factors identified in that model as strong predictors of distance from the labour market. Alternatively, if DYPA chooses to maintain its current profiling tool, it could apply random forest methods to the data identified under Section 3.4.3 to flexibly identify factors associated with distance from the labour market. One relatively simple and widely used alternative is to employ classical probability models, such as probit or logit regression. These models are well-suited to examining the relationship between multiple risk factors and a binary outcome variable – such as being long-term unemployed or moving into employment – where the outcome takes a value of either 0 or 1. The advantage of using a traditional model is that it is relatively easy to implement, and the direction in which individual risk factors influence the outcome is directly visible from the estimated model coefficients. However, the main disadvantage is that such models are less flexible in capturing complex relationships between risk factors and outcomes.
Determine in which direction single risk factors are affecting the prediction
As mentioned above, when using traditional models such as logit or probit, the direction in which risk factors influence the outcome is clearly indicated by the sign of the estimated coefficients. In contrast, more sophisticated machine learning algorithms, such as random forests, provide information on the relative importance of different factors in determining the outcome but do not indicate the direction of their effect. For example, the analysis in Chapter 3 shows that the reason for the termination of a previous contract – e.g. the end of a fixed-term contract – is a relatively strong predictor of the probability of moving into employment within 12 months of registration. However, the model does not directly show whether individuals whose contracts ended due to the end of a fixed-term contract are more or less likely to enter employment compared to those whose contracts ended for other reasons.
One straightforward way to understand how different factors affect the probability of entering employment is to use survival curves, as shown in Section 3.2.3. By plotting these curves, it is possible to analyse whether jobseekers with certain characteristics enter employment faster or slower than others and, therefore, understand the direction in which a particular factor influences employment outcomes. Another approach is to use conditional inference trees (Hothorn, Hornik and Zeileis, 2006[9]). These operate similarly to decision trees (described in Box 3.1), but rather than focussing solely on predicting a specific outcome (e.g. the probability of long-term unemployment), they also use statistical tests to determine which variables to split on. This makes it possible to identify the direction and influence of each risk factor on the prediction. As a result, this method makes it possible to identify how specific risk factors (or combinations of factors) relate to better or worse employment outcomes, providing clear, interpretable rules that show the direction of their impact.
If DYPA decides to upgrade its current profiling tool and adopt the machine learning-based model proposed by the World Bank, it could complement this model with conditional inference trees to gain clearer insights into the direction in which each risk factor affects a jobseeker’s distance from the labour market.
Regardless of the approach adopted – whether using traditional probability models or machine learning-based methods – DYPA could generate estimates of the relative importance of individual risk factors in determining a jobseeker’s distance from the labour market, as well as the direction of their impact. The proposed digital tool could use these estimates to show counsellors not only the overall profiling score but also the key characteristics driving the profiling score (either increasing or decreasing the distance from the labour market) similar to the tool used by the Estonian PES (Figure 3.8). In addition, the tool could display a list of employment-related barriers linked to each risk factor, helping counsellors to better understand the challenges faced by jobseekers and to identify the measures needed to support them. This would allow for a more effective and targeted use of counselling time.
3.4.4. The tool will help counsellors referring jobseekers to tailored services
Vulnerable jobseekers often face multiple barriers to employment. To create effective pathways into employment for this group, a range of integrated social supports is required, going beyond the services currently offered by DYPA. These services may range from more targeted and intensive employment counselling to broader social support, such as housing and healthcare services. Some, like individual and group counselling, are already provided by DYPA. Others, such as intensive job coaching, could be contracted out by DYPA to external providers while broader support services, such as housing and healthcare, fall under the responsibility of other ministries or are delivered by NGOs. Chapter 4 discusses the procurement of intensive job placement support in more detail.
The analysis in Section 3.3 highlights some bottlenecks in the services DYPA provides to the most vulnerable jobseekers. Specifically, while the overall level of service provision is already low, with only around 5% of all registered jobseekers participating in ALMPs, participation is even lower among vulnerable jobseekers reaching only 2.1%. This may reflect both a lack of tailored ALMPs for this target group and the greater difficulty in reaching and engaging these jobseekers.
The current profiling tool already provides a list of potential ALMPs to which jobseekers can be referred to based on their profiling score, thereby supporting counsellors in making referrals to ALMPs. The new digital tool could build on this functionality and expand it in two ways. First, it could generate for each jobseeker a tailored list of available and recommended ALMPs that best match their individual profile. While the current system suggests ALMPs based primarily on the profiling score, the new tool could fully leverage DYPA’s data to identify the most suitable programmes for each jobseeker. To achieve this, the tool could integrate the results of the component of the project that DYPA has carried out with the World Bank on the Feedback Mechanism, a system designed to leverage information on jobseeker characteristics and preferences, the specific barriers they face, the availability of programmes in a specific region and the local labour market conditions to support more effective referrals to ALMPs.
Second, the new tool could incorporate a mapping of available services at the regional level for vulnerable individuals, based on their specific needs, and provided by organisations other than DYPA. As an example, the Lithuanian PES has developed a social assistance map, which includes more than 500 providers of various services relevant to vulnerable groups across Lithuania. The map allows users to search for services by region, type of service, or both. This tool can be used by PES case managers to inform vulnerable jobseekers about the availability of external services that may be helpful to them, or directly by jobseekers themselves to search for services available in their area.
Similarly to the Lithuania PES, DYPA could consider mapping the various services available to vulnerable jobseekers across Greece and integrate this information into the tool, enabling counsellors to refer jobseekers to relevant services when appropriate. However, the development of such a tool would require close collaboration between DYPA and other institutions and organisations supporting vulnerable people. To be effective, it would also require up-to-date information on the services provided by these organisations. DYPA has already taken an important first step by signing several memoranda of understanding (MoU) with institutions involved in supporting vulnerable groups. It could therefore start by mapping the services offered by these organisations, gradually expanding the map as new partners are identified and additional MoU are signed.
The implementation plan for the new digital tool consists of three key components. First, it outlines seven key steps DYPA should take on its journey towards implementing the new tool. Second, it sets out a provisional implementation timeline. Third, it defines the resources and technical skills required at each stage of the implementation.
Step 1: Define DYPA’s future profiling approach and how the new tool will complement it
As a first step, it is important to define the strategic context for introducing the new tool. To ensure that the new digital tool meaningfully complements the overall profiling tool without creating overlaps, DYPA should first clarify the future direction of its overall profiling process. The implementation of the tool will depend critically on whether DYPA intends to revise its current profiling approach and, if so, what form the revised approach will take. In addition, it is important to situate the tool within DYPA’s broader digitalisation strategy.
It is also useful to define the tool’s specific objectives and intended users. For example, it should be clarified whether the tool will form part of DYPA’s standard profiling process (whether current or revised), or whether it will function as a stand-alone solution used only in DYPA’s local offices for vulnerable jobseekers. Although the tool has been designed with the needs of vulnerable jobseekers in mind, it could, in principle, also support counselling for more employable jobseekers.
Step 2: Identify relevant stakeholders and engage them in the implementation process
Engaging relevant stakeholders from the beginning is essential to ensure that all critical dimensions – such as usability, technical feasibility, and data protection – are addressed and to avoid costly adjustments later in the implementation. Early and regular involvement not only ensures the tool responds to actual needs but also strengthens ownership and promotes uptake among counsellors.
Key stakeholders include employment counsellors in DYPA’s KPAs for vulnerable groups (and, possibly, those in regular KPAs), as they will be the tool’s primary users. Jobseekers – particularly those from vulnerable groups – should also be engaged to ensure the tool effectively captures their individual labour market situation. In addition, internal or external IT specialists responsible for the tool’s technical implementation should be involved to confirm its technical feasibility, alongside data protection officers to ensure compliance with data protection regulations. It is equally important to involve DYPA’s data analysts, as well as any external consultants working with DYPA’s data.
Further relevant stakeholders include other institutions and organisations in Greece that support vulnerable groups, such as Community Centres and NGOs providing local services. Their involvement is key to identifying external support services relevant to the needs of vulnerable jobseekers. One way to facilitate stakeholder engagement throughout the implementation is by establishing an expert steering committee to guide and oversee the tool’s development and implementation.
Step 3: Define the technical and operational design of the tool
Building on Step 1 (defining DYPA’s future profiling approach) and Step 2 (identifying and engaging key stakeholders), Step 3 focusses on specifying the detailed design of the new digital tool. This involves clarifying the data sources and methods to be used, as well as determining how the tool will best support employment counsellors in referring jobseekers to appropriate services.
Below are illustrative questions to be addressed in this step, presented for two possible scenarios.
Scenario 1: The current profiling approach is maintained with minor revisions
Can the existing profiling tool be enhanced to incorporate additional risk factors (in addition to those collected through the profiling questionnaire), such as prior spells of long-term unemployment from DYPA’s internal data and employment histories from ERGANI?
If so, what technical modifications are needed in the web service between DYPA and ERGANI?
Should the additional information be integrated into the calculation of the profiling score and, if yes, how?
Scenario 2: The current profiling approach is revised and statistical profiling is adopted.
Does the methodology adopted provide information on the most important risk factors and direction of effect? If not, how can they be determined?
Which data sources should be used to determine individual risk factors (e.g. data from the profiling questionnaire, DYPA data on unemployment histories, ERGANI data on employment histories)?
Which jobseeker outcomes should the profiling model aim to predict as indicators of successful labour market integration (e.g. likelihood of entering employment, risk of long-term unemployment, or likelihood of returning to unemployment)?
How frequently should the model parameters – reflecting the importance of different risk factors – be updated, who will be responsible for these updates, and what resources will be allocated for this purpose?
Questions relevant to both scenarios
Which services should be mapped into the tool to enable counsellors to refer jobseekers (formally or informally) to suitable support?
Which organisations should be involved in mapping these services?
Which services should DYPA provide directly, and which could be outsourced to external providers (see also Chapter 4 of this project)?
Which services are already offered by other institutions and NGOs?
How will the mapping of services be kept up to date? Who are the relevant stakeholders, and how should DYPA engage and co‑ordinate with them?
Step 4: Develop a training strategy for DYPA counsellors
An important step involves building the capacity of employment counsellors to use the new tool effectively. This can be supported through a combination of in-person and virtual training sessions, developing practical guidelines for the tool’s use, sharing examples of good practice, and organising Q&A sessions to address common challenges. Another widely used training approach is the “Train the Trainer” model. In the context of the new digital tool, DYPA could adopt this approach by initially training a small group of counsellors in each KPA. Ideally, these counsellors should be experienced, motivated, and well-respected by their peers, enabling them to act as facilitators and transfer their knowledge of the tool to other colleagues within the same KPA.
Training and supporting materials should be regularly updated to reflect the current status and functionality of the tool. This will help ensure that training remains relevant to users’ needs and maximises the productivity gains achieved through the tool’s adoption. In addition, training should help counsellors understand both the capabilities and the limitations of the tool, supporting its appropriate and effective use in practice.
Step 5: Establish a monitoring and evaluation framework
The implementation of the new tool should be systematically monitored and evaluated. This requires setting up a Monitoring and Evaluation (M&E) framework that enables DYPA to identify and address implementation challenges in a timely manner and assess whether the tool is achieving its intended objectives. The design of the M&E framework will be addressed in detail in Chapter 5.
Step 6: Pilot the new tool
Once the technical and operational details of the tool are defined, DYPA should pilot the new tool. A pilot phase will allow DYPA to assess the tool’s practical impact, observe how employment counsellors use the tool in practice, verify whether it meets the objectives set out in Step 1, and make adjustments before national roll-out.
If the tool is intended for use exclusively in KPAs for vulnerable groups, DYPA could pilot it with a selected group of counsellors in these KPAs. If, instead, the functionalities of the new tool are to be integrated into the broader profiling tool used across KPAs in Greece, the pilot could be implemented in a subset of KPAs. One effective approach to piloting is to conduct a randomised controlled trial (RCT), where the use of the tool is randomly assigned across counsellors or KPAs. If properly designed and implemented, an RCT can provide robust evidence on the impact of the tool.
Step 7: Refine the tool and roll it out nationally
The pilot phase should help identify any shortcomings and areas where adjustments to the tool are needed. Once these issues have been addressed and the tool refined accordingly, DYPA can proceed with the national roll-out.
Indicative resources needed for the implementation plan and timeline
Figure 3.9 presents an indicative timeline for the key steps in the implementation plan of the new digital tool, starting shortly after the completion of this project. The steps are grouped into four broader phases: planning, development, piloting, and roll-out. For each step, DYPA will need to mobilise specific resources. Table 3.6 provides an indicative overview of the skills and level of resources required for each stage. The resources required are estimated on a scale from 1 (low) to 5 (high). The overall cost of implementing the tool will largely depend on its final design and the extent to which synergies can be realised through a potential revision of the overall profiling system. Costs will also vary depending on the extent to which DYPA procures the technical development externally. To estimate the cost of technical development and integration into DYPA’s existing IT infrastructure, DYPA could draw on the experience and budgets of similar digital tools that have been recently introduced. Training costs are expected to be broadly in line with those of previous training sessions delivered to employment counsellors.
The first activities include deciding on the future direction of DYPA’s overall profiling process, defining the objectives of the new digital tool (Step 1), and analysing and engaging relevant stakeholders (Step 2). While it is essential to complete Step 1 before proceeding with the technical development of the tool, stakeholder engagement should be ensured throughout the implementation. Although Step 1 requires relatively few operational resources, it involves a strategic, high-level decision on the future approach to profiling within DYPA. In contrast, defining the tool’s objectives requires a strong understanding of how employment counsellors (particularly those working with vulnerable jobseekers) operate.
Once the future profiling process is determined and the tool’s objectives are defined, DYPA can define the details of the new digital tool (Step 3). This stage requires both a deep understanding of counsellors’ work and expertise in data science, including econometric methods. While parts of the tool’s development (particularly IT development) could be outsourced to external contractors, it is important that DYPA has sufficient in-house skills and resources to guide the work of external contractors and ensure the tool can be updated over time.
In Step 4, DYPA should develop a training strategy to ensure that employment counsellors are properly equipped to use the tool. Training material should be prepared during the development phase, so that it can be tested and, if necessary, updated during the pilot.
The development of a M&E framework (Step 5) is a critical step to identify implementation challenges and assess whether the tool is meeting its objectives. The M&E framework will be developed as part of Chapter 5.
In Step 6, DYPA could pilot the new tool to test its functionality in a controlled environment and make any necessary adjustments before scaling it up nationally. This step is particularly resource‑intensive and requires a diverse set of skills, including expertise in piloting, monitoring and evaluation, and strong communication skills to ensure effective collaboration with counsellors and the IT specialists developing the tool.
Finally, once the results of the pilot are available, DYPA should assess, in Step 7, whether further adjustments to the tool are needed and, if so, define and implement these adjustments before the national roll-out.
Indicative timeline for the implementation of the new tool
Indicative skills and resources needed for the implementation of the new tool
3.5.1. Potential risks and mitigation measures
The implementation of a new digital tool comes with a number of operational and strategic risks. Table 3.7 summarises these risks, their potential impact on implementation and possible mitigation measures.
A first set of risks relates to whether the tool will be used effectively in practice and whether it will function as intended. Limited take‑up by counsellors (e.g. due to insufficient digital skills), the presence of algorithmic bias, limited predictive power of the underlying statistical model or a lack of appropriate data, including missing information on important vulnerability factors, would all reduce the tool’s ability to identify vulnerable jobseekers and to provide meaningful support to counsellors. These risks can be mitigated by involving counsellors early in the design and roll-out process, systematically monitoring the tool’s usage and providing targeted training so that counsellors understand its functionalities and can integrate these into their day-to-day work. These risks can also be mitigated by rigorously testing, monitoring and auditing the model, updating it regularly, and putting in place robust data exchange agreements to ensure timely access to high-quality data.
A second group of risks concerns the broader institutional and governance context in which the tool will operate. If information on support available to vulnerable groups in a given region and provided by other organisations is incomplete or not regularly updated, counsellors will struggle to inform vulnerable jobseekers about appropriate services outside DYPA. Likewise, if the tool is not well aligned with DYPA’s wider revision of its profiling approach and the development of new digital tools, there is a risk of duplicating or overlapping functionalities, leading to a fragmented patchwork of tools that is difficult and time‑consuming for counsellors to navigate. These risks can be reduced by maintaining and expanding co‑operation with relevant organisations, developing and updating a shared mapping of services, and ensuring close internal co‑ordination within DYPA so that the new tool coherently complements existing and planned digital tools.
Despite DYPA’s significant investments in digitalisation in recent years, which have improved the efficiency of service provision, important challenges remain, particularly in supporting vulnerable jobseekers who are far from the labour market and face multiple barriers to employment.
Building on the overview of digital tools used by PES across the OECD presented in Output 1, as well as the analysis of individual-level data from DYPA and the ERGANI register, this note proposes a concept for a new digital tool to help DYPA identify, prioritise, and support vulnerable jobseekers more effectively, with the overarching goal of promoting their transition into employment. The proposed tool takes the form of a dashboard that leverages data to flag jobseekers needing intensive support and displays the key risk factors most predictive of each jobseeker’s distance from the labour market, along with the potential barriers associated with these factors.
The risk factors identified with the data available for this report include: being over 50 years old; being a woman, especially if married; having low or medium levels of education; having experienced long-term unemployment in the past; having a previous job that ended in dismissal; limited or no employment in the six months prior to registration. Although the analysis can be expanded with additional data, these indicators already provide the foundation for a viable digital tool.
In addition, the tool is designed to support employment counsellors in referring jobseekers to appropriate ALMPs and other relevant services. To ensure a holistic approach to supporting vulnerable jobseekers, the tool enables counsellors to view services available in the region that are provided by other institutions and NGOs. Overall, the proposed digital tool would allow DYPA to make better use of data available to support the work of employment counsellors in supporting vulnerable jobseekers into employment.
The implementation of such a tool presents multiple challenges. First, its development and implementation should be closely co‑ordinated with any potential revision of DYPA’s overall profiling approach to avoid duplication and to ensure meaningful integration with DYPA’s broader digital infrastructure. Although the current profiling approach was recently revised as part of a project with the World Bank, no final decision has yet been made regarding the future of profiling at DYPA. Second, the integration of additional data in the profiling process would require extending the existing systematic data exchange with ERGANI and potentially expanding it to other institutions, such as EFKA. While the web services provided by the Ministry for Digital Governance offer a sound basis for such data sharing, the information retrieved from other registers requires further processing before it can be effectively used for profiling purposes. Third, although the new digital tool can help DYPA better prioritise meetings with jobseekers and allocate resources more efficiently, the caseload per counsellor remains high in Greece, limiting the scope for personalised and tailored counselling.
References
[2] Breiman, L. (2001), “Random Forests”, Machine Learning 45, pp. 5–32.
[1] Desiere, S., K. Langenbucher and L. Struyven (2019), “Statistical profiling in public employment services: An international comparison”, OECD Social, Employment and Migration Working Papers, No. 224, OECD Publishing, Paris, https://doi.org/10.1787/b5e5f16e-en.
[9] Hothorn, T., K. Hornik and A. Zeileis (2006), “Unbiased Recursive Partitioning: A Conditional Inference Framework”, Journal of Computational and Graphical Statistics, Vol. 15/3, pp. 651-674, https://doi.org/10.1198/106186006x133933.
[8] Immervoll, H. and S. Scarpetta (2012), “Activation and employment support policies in OECD countries. An overview of current approaches”, IZA Journal of Labor Policy, Vol. 1/1, pp. 1-20, https://doi.org/10.1186/2193-9004-1-9.
[3] Kaplan, E. and P. Meier (1958), “Nonparametric Estimation from Incomplete Observations”, Journal of the American Statistical Association, Vol. 53/282, pp. 457-481.
[7] Leinuste, K. (2025), Work-focussed counselling and Decision Support Tool, https://www.oecd.org/content/dam/oecd/en/about/programmes/dg-reform/bulgaria/Work-focussed-counselling-and-OTT-EUIF-Karina%20Leinuste.pdf.
[4] Meyer, B. (1990), “Unemployment insurance and unemployment spells”, Econometrica, Vol. 58, pp. 757-782.
[6] OECD (2021), “Building inclusive labour markets: Active labour market policies for the most vulnerable groups”, OECD Policy Responses to Coronavirus (COVID-19), OECD Publishing, Paris, https://doi.org/10.1787/607662d9-en.
[5] World Bank (2024), Component 2: Development of a Proposed Methodology for Evidence-based oJbseeker Profiling and Feedback Mechanisms : Machine Learning-Based Statistical Model for Jobseeker Profiling in Greece (English), World Bank Group., http://documents.worldbank.org/curated/en/099041825135514310.
The original DYPA registration dataset contained a very small share of duplicate entries (0.15%) with identical registration identification numbers. In these cases, the registration record containing the most complete profiling information was retained; when multiple records had an equal amount of profiling information, the earliest entry was preserved. Additionally, in the relatively rare instances (1.14%) where a registration ended due to either merging of the registration card or failure to renew the card, and a new registration began within one day, these consecutive registrations were merged into a single continuous unemployment spell.
DYPA’s ALMP data includes information on the jobseeker, the programme, the employer, and the job placement. Due to how employers are recorded in the system, around 24% of entries are duplicates for the same person and programme, often linked to multiple employer activity codes. In most of these cases, it is possible to retain a single observation by prioritising the employer’s primary financial activity. Each programme entry is then linked to the most recent unemployment spell before the programme began, in order to integrate it with the wider registration data. The resulting dataset includes participations in 305 972 programmes for 268 712 unique jobseekers. In addition, a second dataset provides information on participation in group job counselling sessions organised by DYPA. This dataset contains records for 6 842 participations by 6 046 individuals and includes details on the type of counselling programme and whether the jobseeker attended to all sessions. Together, these datasets provide a rich source of information on how ALMPs and support services are accessed by jobseekers over time.
As the unit of observation in the dataset is the registration, most analyses are carried out at the registration level. When the analysis is conducted at the individual level (Section 3.3.2) the latest available registration is used, whether ongoing or concluded. Information from earlier registrations is retained through constructed variables that capture the number and characteristics of previous unemployment spells.
Note: Education levels are based on the International Standard Classification of Education (ISCED), where Low = ISCED 0‑2, Medium = ISCED 3‑4, High = ISCED 5‑8. Occupations are classified into three skill levels based on the International Labour Organization (ILO) ISCO‑08 standard: High (levels 3‑4: managers, professionals, technicians), Medium (level 2: clerical, service/sales, agriculture, craft, machine operators), and Low (level 1: elementary occupations). Armed forces occupations are not classified. “Vulnerable” jobseekers are those identified by DYPA as those whose inclusion in social and economic life is hindered by physical or mental causes, or due to delinquent behaviour (e.g. people with disabilities, individuals with addiction problems, young offenders, prisoners, and former prisoners). “Special category” includes individuals registered under specific DYPA priority groups because of disadvantages in terms of access to the labour market (e.g. persons with disabilities, single parents, etc.) “Other jobseekers” includes all remaining individuals registered with DYPA and not falling into the previous two categories.
Source: OECD calculations based on data from the Greek public employment service (DYPA).
Note: “Other special category groups” refer to the following groups: victims of human trafficking, persons who still reside in Child Protection and Care Units after reaching adulthood, transgender people and persons identified as vulnerable under an earlier DYPA classification. Long-term unemployed refers to those with at least 365 consecutive days registered with DYPA. DYPA profiling categories range from 1 (job ready) to 5 (special groups with major barriers).
Source: OECD calculations based on data from the Greek public employment service (DYPA).
Note: The figure presents the 15 most important predictor variables in determining whether an unemployment spell lasts 12 months or longer, derived from the Random Forest. Variable importance scores are calculated based on how much each variable contributes to accurately distinguishing jobseekers who become long-term unemployed from those who do not. The scores are normalised so that the highest-ranking predictor is assigned an importance of 1, and all other predictors’ importance values are scaled accordingly. The data includes all registrations from 1 January 017 and starting before 31 December 2023.
Source: OECD calculations based on data from the Greek public employment service (DYPA) and ERGANI.
← 1. Jobseekers in some professions are not obliged to undertake the profiling (i.e. teachers, builders, seasonal workers, workers employed in motorised fishing vessels and emery mines)
← 2. Figure A B.1 provides similar information but focusses on predicting the likelihood of entering employment within one year from the registration with DYPA.
← 3. Registered DYPA jobseekers have access to a broader range of programmes, including other forms of training, supported employment and rehabilitation, and start-up incentives. However, these programmes are not included in the figures presented as their participation data were not accessed for the study.