The Federal Reserve Board published a Finance and Economics Discussion Series paper proposing two complementary approaches that replace suspect measurements with missing values to reduce outlier-driven distortions in linear state-space filtering and forecasting. The paper introduces supervised missing data substitution (MD), which sets observations to missing when they exceed a Huber threshold to improve the performance of Huber-based linear filters when same-sign outliers are clustered over time. It also develops an unsupervised method, missing data substitution via exogenous randomization (RMDX), which averages filtered or forecasted targets over measurement series with randomly induced missing-data subsets at a chosen randomization rate, creating a bias-variance trade-off that can be tuned via cross-validation. Monte Carlo simulations indicate both methods can materially improve robust filtering, particularly when combined, and an empirical application reports consistently attractive performance when forecasting inflation trends affected by clustered measurement outliers.
Federal Reserve Board 2025-01-31
Federal Reserve Board research proposes supervised and randomized missing data substitution to strengthen robust filtering and forecasting in state-space models
The Federal Reserve Board's paper proposes two methods to address outlier-driven distortions in linear state-space filtering and forecasting. The supervised missing data substitution (MD) method improves Huber-based filters by setting observations to missing when exceeding a threshold. The unsupervised method, missing data substitution via exogenous randomization (RMDX), averages targets over series with induced missing-data subsets. Monte Carlo simulations show both methods enhance robust filtering, especially when combined, with promising results in forecasting inflation trends.