Robust Residual Correction via Selective Deployment for Time Series Forecasting
The proposed framework, Causality-Inspired Safe Residual Correction (CRC), aims to guarantee non-degradation in multivariate time series forecasting by safely correcting systematic errors. It uses a causality-inspired encoder to find directional structure, a hybrid corrector with a linear floor and MLP delta, and a four-fold safety firewall. This ensures the correction term never makes the forecast worse than the original baseline.