Conformal Counterfactual Forecasting with Reduced Uncertainty for Time Series Predictions

We propose a counterfactual framework for time series forecasting that targets uncertainty reduction rather than prescribed prediction ranges. The method, Conformal Interval Width Reduction (CIWR), generates counterfactual inputs that yield forecasts with statistically valid yet narrower conformal prediction intervals. CIWR couples conformal prediction with gradient-based counterfactual search via a differentiable width proxy, enabling optimization directly on inputs while preserving finite-sample coverage. We provide a variance-normalized warm-start that exactly aligns the initial proxy with the original conformal width and yields stable gradients. On forecasting reformulations of the UCR GunPoint and Italy Power Demand datasets using a GRU forecaster, CIWR achieves substantial width reductions while maintaining validity and proximity.