Regime-Aware Evidence Fusion for Calibrated Forecasting of Non-Stationary Time Series

Abstract

Non-stationary forecasting is not the same as forecasting in average terms; forecasting performance depends on the ability to forecast regime changes rather than the ability to fit perfectly in average cases. Regime shifts, calendar effects, policy interventions, and unreliable instrumentation constantly disrupt energy grids, transportation networks, and sensor-intensive urban systems, while even the most commonly-used energy- and transport-related learning-based forecasters typically assume that such systems are near-stationary or incorporate a fragile change-point detector as a pre-processing module. We choose a rather alternative path: discover the regime, gain insight into context, tailor or rate uncertainty, and continuously adapt. All these steps are learnable steps in the pipeline, from the beginning, not added on in a later phase. The five phases create this structure, which functions as follows. Under a causal-invariance constraint, a causal regime tokeniser (CaRT) is able to learn causal-invariance changes in tokens and small context pieces. A Wavelet–State-Space Dual-Path Predictor (WSDP) subsequently is used to predict in the short, medium and long-range combining regime conditioned wavelet decomposition with a neural path space model. A module called 'evidence-weighted fusion' (EviFuse) combines the two forecasts with a (reliability-based) 'weight' determined by the residual. A counterfactual, cross-regime validation suite (CF-XValid) is introduced to stress-test the entire pipeline system, but not just rolling splits to generate new counterfactual scenarios. Last, a drift-aware continual-learning module (DriftShield-ACL) is utilized to maintain honest uncertainty estimates in production with light-weight adapter updates and conformal online updates. The framework reduces mean absolute error by 12–18% compared with the best of three current frameworks on four multivariate, highly non-stationary benchmarks, and lowers the RMSE across long-horizon (96-step) testing by 10–15% compared to the best of three contemporary ones in the targeted 96-step horizon. Its error rate drops more so when up to 20% of exogenous inputs are unavailable, as compared to any baseline, and its 90% prediction intervals remain around nominal (88–89%) as compared to many baselines, which are one or more points below. With compound stress (fast changing levels, beds of high demand, loss of sensors), the worst-case error decreases by approximately 20% compared to the best baseline. Overall, the findings indicate that the regime awareness method, the multi-scale modelling approach, the evidence-based fusion approach and the calibrated adaptation method are more effective as an integrated design of a "combined system" rather than simply as a collection of fixes and more effective when tested within robustness and deployment trust, not only average accuracy.