Modeling Information Blackouts in Missing Not-At-Random Time Series

Traffic sensor networks frequently experience “blackouts,” i.e., contiguous intervals of missing observations. This preprint evaluates two tasks: (1) blackout imputation (reconstructing values inside blackout windows) and (2) post-blackout forecasting at horizons +1, +3, and +6 steps on a 5-minute grid. We compare a MAR linear dynamical system (Kalman filtering with RTS smoothing) against an MNAR extension that treats the missingness mask as an informative observation channel via a logistic missingness model conditioned on the latent state. The repository includes code, evaluation-window manifests, and notebooks for experiments on the Seattle Loop dataset and the METR-LA dataset.