2025-01-12-美联储-线性状态空间模型中增强鲁棒滤波和预测的缺失数据替换（英）_38页_2mb

Summary

This paper introduces two complementary methods, Supervised Missing Data Substitution (MD) and Unsupervised Missing Data Substitution via Exogenous Randomization (RMDX), to enhance the outlier-robustness of filtering and forecasting in linear state-space models. MD improves existing Huber-based filters by replacing outliers exceeding a Huber threshold with missing data, directly eliminating outlier-induced errors. RMDX further addresses smaller undetected outliers by averaging filtered or forecasted values from measurement series with randomly missing data, leveraging a regularization effect via cross-validation to optimize a randomization rate. Both methods are validated through simulations and real-world inflation forecasting. The combined approach, especially RMDX-MD-RobKF, significantly improves performance by suppressing both detected and undetected outliers, offering practical and effective enhancements. This work bridges statistical and econometric methods in state-space modeling by applying time-series extensions of bagging and rational inattention principles.