Post-hoc residual correction is a common trick for squeezing extra accuracy out of time-series forecasters, but the authors of a new arXiv paper argue that applying it unconditionally can backfire. When residual signals are weak or noisy, an expressive corrector tends to overfit, sometimes making the overall forecast worse than the original baseline.
Their proposed framework, CRC, addresses this by treating correction as a deployment decision rather than just a fitting problem. A hybrid corrector combines a conservative ridge regression "floor" with conditional nonlinear refinement, but corrections are only applied when held-out validation data indicates that the updated forecast will reduce error relative to a frozen baseline forecaster.
Across standard long-sequence benchmarks and multiple backbone models, CRC improved horizon-averaged MSE and MAE in most dataset-backbone pairs, with especially large gains on complex datasets—including a relative MSE reduction of over 20% on Traffic. The paper also reports a non-degradation rate (NDR) as a stability diagnostic and ablates the validation-calibrated "firewall" that prevents harmful updates. The work is accepted at NeurIPS 2026.