Online Conformal Prediction for Non-Exchangeable Panel Data
The Weighted Temporal Quantile Adjustment (W-TQA) is an online method for making predictions on panel data where units are not exchangeable over time. It uses two states—unit similarity weights and an adaptive miscoverage level—to create a stepwise coverage bound. This approach ensures good prediction accuracy, especially for the most difficult-to-predict units, by adjusting the prediction set dynamically based on feedback.