SEMINAR: sparsity enhanced mamba with neural iterative regression for predicting unemployment dynamics

Forecasting unemployment rates using multivariate time series is a tough problem due to its complex temporal dependencies, demographic complexity and need for uncertainty quantification that is interpretable. In labour economics and macro-econometrics, the demographic complexity of the unemployment rate dynamics has been well studied; the recent approaches for deep learning forecasting barely take into account the demographic complexity of the unemployment dynamics. We present SEMINAR: sparsity-enhanced Mamba (SEM), economic priors integration layer (EPIL), adaptive cross-variable attention (ACVA), seasonally-naive anchor-based iterative refinement regression block (IRRB) and probabilistic uncertainty quantifier (PUQ), giving heteroscedastic, horizon-aware confidence intervals. This proposed model is tested against autoregressive integrated moving average (ARIMA), long short-term memory (LSTM), non-sparse Mamba, and transformer in three progressively harder settings using international data for unemployment rate forecasts (2006–2024) in a leak-free chronological protocol with 12-step-ahead multi-horizon forecasting and statistically significant performance evaluation across eight random seeds. In the base setting, SEMINAR gives the best mean absolute error of all five models (128.6 vs . 130.2–149.0 in baselines), beating ARIMA ( p = 0.008) and transformer ( p = 0.039). Under cross-country transfer, SEMINAR is the strongest neural model on the Korea/Chile/Russia task mean absolute error (MAE) (134.6 MAE), although the classical ARIMA baseline remains substantially more accurate there (86.0 MAE); on the short, weakly seasonal Bolivia/Dominican Republic task SEMINAR is outperformed by all four baselines, a failure mode we trace to the seasonal-anchor assumption. Both boundaries are reported rather than omitted. SEMINAR also gives calibrated 80% prediction intervals, a feature that none of the baselines provides in their native form. The complete architecture consists of around 44,401 parameters; an ablation study with seven variants has been done.