Automatic robust estimation for exponential smoothing: Perspectives from statistics and machine learning

•M-estimators, boosting and bagging are evaluated for exponential smoothing.•M-estimators and machine learning approaches show improvements in accuracy.•The Pseudo–Huber loss provides the best accuracy and bias reduction.•Inverse boosting is comparable to M-estimators outperforming conventional boosting.•Bagging achieves poor forecast bias compared to benchmark maximum likelihood.