TORF: A New Framework for Mean-Preserving Probabilistic Time Series Forecasting
A new framework for probabilistic time series forecasting, named Two-stage Odd Residual Flows (TORF), has been developed by researchers to balance distributional flexibility with precise mean predictions. This method, outlined in an arXiv paper (2608.11114), separates mean forecasting from uncertainty assessment. Initially, a deterministic model provides an accurate mean prediction. Subsequently, a Restricted Normalizing Flow utilizing strictly odd functions models the residual distribution while maintaining the mean. Unlike traditional parametric techniques such as Mean Variance Estimation (MVE), which may lead to reduced point accuracy under joint Negative Log-Likelihood (NLL) training, TORF is tailored for risk-sensitive decisions, especially in long-term forecasting. The paper has been submitted to various venues, highlighting its potential to enhance decision-making in finance, energy, and supply chain management.
Key facts
- TORF stands for Two-stage Odd Residual Flows.
- The framework decouples mean forecasting from uncertainty estimation.
- First stage uses a pre-trained deterministic model for mean prediction.
- Second stage uses a Restricted Normalizing Flow with strictly odd functions.
- Addresses trade-off between distributional flexibility and mean accuracy.
- Traditional methods like MVE can suffer from degraded point accuracy under NLL objectives.
- Generative models like Normalizing Flows and Diffusion Models rely on costly Monte Carlo sampling.
- Paper available on arXiv with ID 2608.11114.
Entities
Institutions
- arXiv