FlexCast
A time-series foundation model framework that reformulates forecasting as variable-length masked prediction over time-series segments, disentangling the three entangled factors of autoregression (context conditioning, forecast-unit decomposition, and autoregressive rollout) and supporting multiple inference modes such as one-step forecasting, iterative refinement, and next-token generation.
Timeline (last 90 days)
The FlexCast framework decouples three factors in autoregressive training—variable context conditions, prediction unit decomposition, and autoregressive rollout—and proposes that sequential decoding is not performance-critical, while flexible training distributions are the true source of performance.
FlexCast reframes the prediction task as variable-length masked prediction over time-series segments, dynamically sampling visible context and masked target segments during pretraining without relying on sequential decoding.
In the inference stage, FlexCast supports three modes through a unified mask-predict interface: one-step prediction, confidence-based iterative refinement, and next-token-style generation.
In FlexCast's variable-length masked prediction, each pretraining step randomly selects several contiguous segments as prediction targets and the remaining segments as visible context, with both mask length and position varying dynamically.
FlexCast's ablation experiments show that contiguous target masking and context diversity together constitute the main source of performance improvement.
Evaluation on BLAST and multiple zero-shot benchmarks shows that FlexCast's accuracy is comparable to or even better than strong TSFM baselines, while inference latency is significantly reduced.
Once the model has learned flexible context-target relationships, left-to-right decoding does not bring consistent gains.
The research code for FlexCast has been open-sourced on GitHub, with the repository address BenchCouncil/FlexCast.
All Facts (8)
The FlexCast framework decouples three factors in autoregressive training—variable context conditions, prediction unit decomposition, and autoregressive rollout—and proposes that sequential decoding is not performance-critical, while flexible training distributions are the true source of performance.
50%UnverifiedFlexCast reframes the prediction task as variable-length masked prediction over time-series segments, dynamically sampling visible context and masked target segments during pretraining without relying on sequential decoding.
50%UnverifiedIn the inference stage, FlexCast supports three modes through a unified mask-predict interface: one-step prediction, confidence-based iterative refinement, and next-token-style generation.
50%UnverifiedIn FlexCast's variable-length masked prediction, each pretraining step randomly selects several contiguous segments as prediction targets and the remaining segments as visible context, with both mask length and position varying dynamically.
50%UnverifiedFlexCast's ablation experiments show that contiguous target masking and context diversity together constitute the main source of performance improvement.
50%UnverifiedEvaluation on BLAST and multiple zero-shot benchmarks shows that FlexCast's accuracy is comparable to or even better than strong TSFM baselines, while inference latency is significantly reduced.
50%UnverifiedOnce the model has learned flexible context-target relationships, left-to-right decoding does not bring consistent gains.
50%UnverifiedThe research code for FlexCast has been open-sourced on GitHub, with the repository address BenchCouncil/FlexCast.
50%