Parametric Prior Mapping for Non-Stationary Time Series Forecasting
A novel approach named Parametric Prior Mapping (PPM) has been launched for forecasting probabilistic multivariate time series. By integrating parametric structural priors into a generative modeling framework, PPM employs a parametric estimator to create a dynamic and adaptive prior. This prior aids in the learning of a complex predictive distribution through a learnable mapping, achieving a balance between expressiveness and robustness. It maintains the efficiency of parametric techniques while leveraging the expressive capabilities of generative models. When trained with a hybrid objective, PPM produces accurate forecasts accompanied by well-calibrated uncertainty estimates, surpassing current methods.
Key facts
- Parametric Prior Mapping (PPM) framework introduced for non-stationary probabilistic time series forecasting.
- PPM injects parametric structural priors into a generative modeling process.
- Uses a parametric estimator to derive a dynamic, adaptive prior.
- Guides learning of complex predictive distribution via a learnable mapping.
- Trained with a hybrid objective.
- Yields precise forecasts with well-calibrated uncertainty estimates.
- Balances expressiveness and robustness.
- Retains efficiency of parametric methods while exploiting generative models' expressive power.
Entities
—