SAGA: Synthetic Earnings Data for 50,000 Individuals with Demographic Effects

A synthetic dataset of 50,000 individuals generated from a parametric AR(1) model with education- and sex-specific fixed effects and an empirically calibrated Swedish age-earnings profile. It was created by Gustav Olaf Yunus Laitinen-Fredriksson Lundström-Imanov of Stockholm University for the SAGA manuscript submitted to IEEE TPAMI (2026). The deposit includes synthetic train/cal/test splits and moment validation statistics.

Use Cases

• Benchmarking probabilistic forecasting algorithms based on the described AR(1) model with fixed effects.
• Evaluating membership inference risks on synthetic data based on the reported AUC metric.
• Testing moment-matching validation techniques based on the reported 1.8% match within demographic subgroups.
• Developing adaptive temporal conformal prediction methods using the provided train/cal/test splits.

Strengths

• Contains 50,000 synthetic individual records.
• Membership inference AUC is reported as 0.512, consistent with random chance.
• Moment match is reported within 1.8% at every age within every demographic subgroup relative to empirical moments.

Limitations

• Column-level documentation is absent; field semantics must be inferred from the schema file after download.
• Row count for individual splits is unknown, which may limit suitability assessment.
• The data is synthetic and based on a specific parametric model; its utility for real-world forecasting requires validation.

Provenance

Source

Synthetic data generated from a model calibrated on Swedish earnings profiles; real LISA microdata from Statistics Sweden MONA project.

Collection Method

Parametric AR(1) model generation with demographic fixed effects.

Geography

Calibrated on Swedish data; synthetic individuals have no real geography.