Temporal Conformal Prediction (TCP): A Distribution-Free Statistical and Machine Learning Framework for Adaptive Risk Forecasting

TL;DR

This paper introduces Temporal Conformal Prediction (TCP), a distribution-free framework for adaptive, well-calibrated risk forecasting in nonstationary time series, combining quantile forecasting with conformal calibration and online adjustments.

Contribution

TCP is a novel framework that integrates modern quantile forecasting with conformal calibration and online updates, enabling reliable risk intervals under distribution shifts.

Findings

TCP achieves near-nominal coverage across datasets.

TCP intervals are slightly wider than historical simulation but well-calibrated.

The online Robbins-Monro update has minimal impact on calibration.

Abstract

We propose \textbf{Temporal Conformal Prediction (TCP)}, a distribution-free framework for constructing well-calibrated prediction intervals in nonstationary time series. TCP couples a modern quantile forecaster with a rolling split-conformal calibration layer; its \textbf{TCP-RM} variant adds an online Robbins-Monro offset to steer coverage in real time. We benchmark TCP against GARCH, Historical Simulation, Quantile Regression (QR), linear QR, and Adaptive Conformal Inference (ACI) across S\&P 500, Bitcoin, and Gold. Three results are consistent. First, QR baselines yield the sharpest intervals but are materially under-calibrated; even ACI remains below the 95\% target. Second, TCP achieves near-nominal coverage, yielding intervals slightly wider than Historical Simulation (e.g., S\&P 500: 5.21 vs.\ 5.06). Third, the RM update changes calibration only marginally at default hyperparameters. Crisis-window visualizations (March 2020) show TCP promptly expanding and contracting intervals as volatility spikes. A sensitivity study confirms robustness to hyperparameters. Overall, TCP bridges statistical inference and machine learning, providing a practical solution for calibrated risk forecasting under distribution shift.