Forecasting Regional NEET Rates in Europe: A Spatio-Temporal Systems Analysis of Persistence, Socioeconomic Context, and Transferability


Kilimci Z. H., Çolak K., Dinçer E.

SYSTEMS, cilt.14, sa.11, ss.1-45, 2026 (Scopus)

  • Yayın Türü: Makale / Tam Makale
  • Cilt numarası: 14 Sayı: 11
  • Basım Tarihi: 2026
  • Doi Numarası: 10.3390/systems14111475
  • Dergi Adı: SYSTEMS
  • Derginin Tarandığı İndeksler: Scopus
  • Sayfa Sayıları: ss.1-45
  • Kocaeli Üniversitesi Adresli: Evet

Özet

Regional disparities in the share of young people not in employment, education, or training (NEET) remain an important challenge for European labour-market and social-inclusion policies. This study develops a leakage-aware spatio-temporal forecasting framework for regional NEET rates across 251 NUTS-2 regions in 30 European countries over 2005–2025. Predictive information is decomposed into temporal persistence (T), regional socioeconomic conditions (S), and external spatial context (P), and forecasting models are evaluated using expanding-window temporal validation, a locked 2022–2025 test period, unseen-region experiments, strict cold-region prediction, nested leave-one-country-out evaluation, and conformal uncertainty quantification. Under the locked future evaluation, the temporally selected ElasticNet model using T+S+P achieved the lowest observed RMSE among the evaluated models (RMSE = 1.685; = 0.867), corresponding to a 9.8% improvement over naive persistence. Ablation analysis showed that regional history was the dominant source of predictive information. Socioeconomic conditions produced a larger numerical improvement beyond temporal persistence, whereas the additional reduction in error from spatial context was small and was not statistically distinguishable from zero under either region- or country-cluster bootstrap resampling. Generalisation performance depended strongly on the evaluation setting: CatBoost performed best in spatial transfer when local history was available, whereas ElasticNet was strongest when transfer to unseen regions was combined with forecasting into the future test period. Under strict cold-start conditions without focal-region NEET history, CatBoost again achieved the strongest performance. Finally, standard conformal prediction under-covered the highest-NEET observations despite strong overall coverage, while predicted-risk-based Mondrian calibration increased coverage in the NEET group from 83.7% to 93.9%. The findings demonstrate that regional NEET forecastability depends not only on model choice but also on temporal persistence, information availability, validation design, and uncertainty calibration