A new arXiv paper introduces Bellman Conformal Inference (BCI), a framework that can be wrapped around any time-series forecasting model to generate calibrated prediction intervals. The work is positioned as an advance over existing conformal inference methods, which are widely used for uncertainty quantification but often struggle with the sequential dependencies in time-series data.
The abstract states that BCI is able to leverage something that existing methods cannot, but the description cuts off at that point, leaving the specific mechanism unclear from the excerpt alone. What is clear is that BCI is designed to be model-agnostic, meaning it can be applied on top of forecasts produced by any model, from classical statistical methods to deep learning architectures.
For practitioners, the appeal is practical: calibrated prediction intervals are essential for decision-making under uncertainty, and a framework that can be bolted onto existing forecasters without retraining them could lower the barrier to reliable uncertainty estimates. The paper is available on arXiv under identifier 2402.05203.