FlexCast: Flexible Context Learning for Time Series Foundation Models
FlexCast matches or beats strong forecasting baselines with lower inference latency, while finding no consistent gain from left-to-right decoding.
The framework trains on time series patches, varying which patches are visible and which must be predicted. A single pretrained model can then forecast in one step, refine predictions iteratively, or generate sequentially. On BLAST and multiple zero-shot benchmarks, the authors report that context diversity and contiguous target masking drive performance. Their experiments suggest flexible training, rather than sequential inference, explains the main benefit of autoregressive-style models.
ArXiv · AI/CL/LG's note
The framework trains on time series patches, varying which patches are visible and which must be predicted. A single pretrained model can then forecast in one step, refine predictions iteratively, or generate sequentially. On BLAST and multiple zero-shot benchmarks, the authors report that context diversity and contiguous target masking drive performance. Their experiments suggest flexible training, rather than sequential inference, explains the main benefit of autoregressive-style models.
ArXiv · AI/CL/LG's note
score 4