Machine learning in employment services: long-term data science partnership with Töötukassa (2019-…)
Since 2019, CITIS has been an integral research and data science partner to the Estonian Unemployment Insurance Fund (Töötukassa) through a long term data science partnership between the latter and Nortal AS. The same question runs through all of it: can the state use its own administrative data to reach people before their situation worsens, and can we tell whether it actually helped?
2019–2021: Building OTT
OTT is a machine learning prediction model built by us for Töötukassa together with Nortal, and in use since 2020. It was the first data-driven decision support tool in Estonia’s employment service. Before it, counsellors worked from experience and manual profiling across roughly 45 indicators, which is a lot of information to weigh consistently across thousands of cases.
When someone registers as unemployed, OTT estimates how likely they are to find work and how likely they are to become unemployed again afterwards, and shows the counsellor the ten factors that most affect that person’s prospects. The first version produced this assessment on day 35 after registration, drawing on five years of data about registered jobseekers from Töötukassa’s records and other state registers, alongside information jobseekers provide themselves.
The prediction is not the point. The conversation that follows it is. Counsellors carry large caseloads, and a risk estimate with its reasons attached helps them see quickly where someone needs more than standard job matching. The counsellor decides; OTT informs.
In 2021 OTT was named the best data-based digital service in the Estonian public sector.
2019-present: MALLE
Built alongside OTT from 2019 and in use ever since, MALLE estimates the impact and cost-effectiveness of the labour market services Töötukassa provides to registered jobseekers. It uses counterfactual impact evaluation, comparing people who took a service against a matched group who did not, to measure effects on employment and earnings for up to three years afterwards.
It reports aggregate effects and breaks them down by subgroup: age, target group, year of service. It also compares services against one another, showing which have the larger effect, which pay for themselves soonest, and which have the better cost-benefit ratio over a longer horizon. That gives Töötukassa’s service owners a basis for deciding which services to redesign and which to scale up or down.
2019–2022: Predicting unemployment before it happens
Registered unemployment is the point at which the risk has already materialised. As part of the RITA programme project MAITT, Andres Võrk and Taavi Unt modelled the other side of the problem: the probability that an employed person becomes unemployed, and the factors driving it. We also evaluated how Töötukassa’s services for employed people affected their later employment and earnings. Read more about our contribution from a previous post and the project report itself.
2022: Testing the model against expert judgement
OTT became research material in its own right. Our colleague Peeter Leets defended a master’s thesis at the Johan Skytte Institute, supervised by Mihkel Solvak and Andres Võrk, on augmenting OTT with expert knowledge. It tested whether feeding caseworkers’ and data scientists’ domain knowledge into the model produced something that both generalised better to new data and stayed closer to the rules practitioners actually work by. That tension between accuracy and interpretability runs through everything we have built here.
2023–2025: Impact analyses for Töötukassa
Parallel to concrete tools, our team hasworked with Töötukassa’s own analysis department on individual evaluations.
A 2023 analysis of temporary work during registered unemployment covered the measure’s first two years, September 2020 to August 2022, examining how taking short-term work while registered affects the duration of unemployment, later employment and wages, using Töötukassa and Tax and Customs Board data.
In 2025 Andres Võrk worked on an analysis of how the duration of unemployment insurance benefit affects median unemployment duration, months worked in the year after returning to work, and average annual wage. It again used counterfactual impact evaluation, forming participant and comparison groups from recipients with different entitlement durations, and set the findings against the costs and revenues of paying the benefit from the combined perspective of the insurance fund and the state budget. Learn more about the detailed methodology of the analysis here.
2024: Matching people to jobs, and jobs to people
Job titles and skill descriptions rarely line up neatly between a jobseeker’s profile and a vacancy. The same role appears under different names, and the same words mean different things across sectors. Our colleague Peeter Leets led the development and testing of OSVALD, a custom fuzzy text matching algorithm designed to work through that mismatch and suggest vacancies for jobseekers and candidates for vacancies.
2024–2026: Evaluating action plans with a language model
Our team researched whether a generative language model can assess the quality of jobseekers’ action plans, in partnership with EXAI, the Estonian Centre of Excellence in Artificial Intelligence, where e-governance is one of the applied focus areas. Every registered jobseeker agrees an action plan with their counsellor within 30 days, setting out their education, skills, experience and job preferences, and it is revisited at every meeting. Because the plans are free text they are rich but hard to analyse at scale. Because they also contain personal data, sending them to a commercial model hosted elsewhere was not an option. We worked with open-weight models such as Meta’s Llama that run on infrastructure the data owner controls, keeping the text inside the secure environment. The effort was led by Gregor Nepste and contributed by Peeter Leets, Mihkel Solvak and the rest of our team.
2024–2026: Evaluating the preventive services
Our colleagues Andres Võrk and Taavi Unt worked with RAKE to carry out a larger evaluation of Töötukassa’s services for employed people, commissioned by the Ministry of Economic Affairs and Communications together with the Fund. It asks whether the services reach the right people, what effect they have on employment and wages, and whether they are cost-effective, using linked register data from 2017 to 2024 inside Statistics Estonia’s secure environment. Details about the study can be found on the ministry’s homepage.
2026: OTT2
A single estimate at registration ages quickly. Someone completes a training course, or misses an activity they agreed to, and the picture they presented on day 35 is no longer the picture in front of the counsellor.
Led by Nortal, we launched OTT2, now in use in Töötukassa bureaus across Estonia. It replaces the one-off assessment with one that updates as circumstances change, and it shows not just the factors behind a prediction but how they have shifted over time.
The Fund is now piloting a change that rests on that accuracy: jobseekers with the strongest employment prospects, around a fifth of new registrations, are no longer required to attend counselling, which frees counsellor time for people at real risk of long-term unemployment. Building a model is the straightforward part. What decides whether it is any use sits around it: obtaining lawful access to linked register data, choosing methods a caseworker can interpret, designing an interface that shows a counsellor why the model says what it says, and evaluating afterwards whether the service changed anything.
Learn more about OTT2 at the OECD AI Policy Navigator.
