SAGA: A Sequence-Adaptive Generative Architecture for Multi-Horizon Probabilistic Forecasting with Adaptive Temporal Conformal Prediction

Researchers introduced SAGA, a decoder-only transformer for irregular panel data that forecasts multi-horizon labor earnings and lifetime income distributions.

On Swedish LISA register data, SAGA outperformed traditional parametric models and baseline methods in earnings prediction.

The model also produced calibrated prediction intervals using split conformal calibration, improving statistical reliability.

Its probabilistic forecasts can strengthen microsimulation tools used by central banks and finance ministries for policy planning.