KReF: Training-Free Retrieval for Probabilistic Long-Term Forecasting
So, there's this new framework called KReF, which stands for Knowledge Retrieval for Forecasting. It’s designed to improve long-term time-series forecasting without relying on traditional training methods. Instead of using trained models or building intervals around forecasts, KReF pulls in relevant historical data to create a predictive distribution on the spot. It incorporates smart preprocessing and uses historical data pairs to guide predictions. The method adjusts based on similarity weights, affecting various metrics like quantiles and continuous ranked probability scores. Plus, it adapts interval estimates using specific rates chosen during validation. This framework really tackles the issue of lagging feedback in long-term predictions. You can check out the full paper on arXiv, ID 2608.06748.
Key facts
- KReF is a training-free retrieval framework for probabilistic long-term time-series forecasting.
- It treats retrieved historical futures as a query-local empirical predictive distribution.
- It uses handcrafted statistics or frozen random Fourier features for embedding lookbacks.
- Similarity weights define predictive masses, quantiles, CRPS, and a weighted-mean point forecast.
- It constructs a probability-integral-transform map from the observed query lookback.
- It applies validation-selected expansion and shrinkage rates to adapt intervals.
- It addresses delayed feedback issues in sequential conformal methods at long horizons.
- The paper is available on arXiv with ID 2608.06748.
Entities
Institutions
- arXiv