SAGA: A Sequence-Adaptive Generative Architecture for Multi-Horizon Probabilistic Forecasting with Adaptive Temporal Conformal Prediction
Researchers have developed SAGA, a novel decoder-only transformer architecture designed for multi-horizon probabilistic forecasting on irregular tabular panel sequences. This model, trained on extensive Swedish longitudinal data, significantly improves upon existing methods in predicting annual labor earnings up to thirty years into the future. SAGA demonstrates superior performance in reducing prediction errors and provides reliable prediction intervals, outperforming traditional parametric models and other machine learning baselines. AI
IMPACT Introduces a new architecture for improved long-term probabilistic forecasting, potentially impacting financial modeling and economic analysis.