PaP-NF: Probabilistic Time Series Forecasting with LLMs and Normalizing Flows
A team of researchers has introduced PaP-NF, a probabilistic forecasting framework that integrates continuous time series with a static large language model (LLM) through a Prefix-as-Prompt approach. This framework also conditions a normalizing flow decoder based on the overall context provided by the LLM. Its effectiveness is assessed using the Continuous Ranked Probability Score (CRPS) on various long-term forecasting benchmarks, showcasing a strong ability to represent uncertainty accurately.
Key facts
- PaP-NF uses a Prefix-as-Prompt mechanism to align time series with a frozen LLM.
- A normalizing flow decoder is conditioned on global context from the LLM.
- Evaluation uses CRPS, a standard probabilistic forecasting metric.
- Tested on multiple long-term forecasting benchmarks.
- Addresses limitations of deterministic models in uncertain environments.
- Proposed by authors on arXiv (2605.23219).
- Focuses on probabilistic rather than single-point predictions.
- Aims to quantify and represent uncertainty in forecasting.
Entities
Institutions
- arXiv