C2FFormer: Coarse-to-Fine Time Series Imputation via Autoregressive Transformer

Time series imputation is essential in many scientific and industrial applications. However, real-world data often exhibit highly variable and block-wise missing patterns, posing challenges to existing methods that assume fixed missing rates and perform well only under random missingness. Inspired by autoregressive training paradigm of language models that emulate human reasoning, we propose a Transformer-based imputation framework with a coarse-to-fine reconstruction strategy. The input series is firstly discretized into multi-resolution patches, and the model then autoregressively predicts fine-grained sequences from coarser representations using causal attention. This progressive training enables the model to capture multi-scale temporal structures and supports zero-shot inference across varying missing scenarios. Experiments on four real-world datasets against 24 recent baselines show that, once trained, our method can be directly applied to diverse missing scenarios, consistently achieves high accuracy, and reduces imputation error by up to 95.09% in challenging block-missing settings. The source code is available at https://github.com/songkang77/C2FFormer .