Parametric Prior Mapping Framework for Non-stationary Probabilistic Time Series Forecasting

Researchers introduced Parametric Prior Mapping (PPM), a hybrid framework for probabilistic forecasting of non-stationary multivariate time series.

PPM uses a parametric estimator to construct a dynamic prior, then maps it into a generative model through a learnable transformation.

A hybrid training objective helps the model improve both predictive accuracy and uncertainty calibration.

Experiments show PPM outperforms strong baselines while remaining computationally efficient for real-world sequential data.