LeapTS: Rethinking Time Series Forecasting as Adaptive Multi-Horizon Scheduling
LeapTS: Rethinking Time Series Forecasting as Adaptive Multi-Horizon Scheduling
Authors: Sheng Pan, Ming Jin, Bo Du, Shirui Pan Date: 2026-05-11 Paper ID: openalex:2605.10292
Summary
LeapTS redefines time series forecasting by moving away from fixed history-to-target mappings, instead viewing the process as an adaptive scheduling task. It employs a hierarchical controller to determine prediction scales and advancement lengths, while utilizing neural controlled differential equations to maintain continuous-time state evolution. This approach explicitly couples irregular temporal dynamics with scheduling feedback, resulting in significant improvements in both predictive accuracy and inference speed compared to traditional Transformer-based forecasting models.
Key Contributions
- Introduces LeapTS, a framework that reformulates time series forecasting as a dynamic multi-horizon scheduling process.
- Implements a hierarchical controller for selecting optimal prediction scales and advancement lengths at each temporal step.
- Integrates continuous-time state evolution via neural controlled differential equations to couple irregular dynamics with scheduling feedback.
- Achieves at least 7.4% improvement in forecasting performance and 2.6x-5.3x inference speedup over Transformer-based baselines.
Open Questions & Future Work
Key Concepts
- adaptive-multi-horizon-scheduling: A forecasting framework that treats multi-horizon prediction as a dynamic scheduling task involving the sequential selection of prediction scales and advancement lengths.
Archivist Review
The approved concept identifies a foundational shift in how time series forecasting is modeled—moving from static mapping to dynamic scheduling—which is highly reusable. The approved open question addresses the critical transition from task-specific training to generalized scheduling foundations. Other potential candidates were excluded to maintain the vault’s high barrier for entry.
Approved Concepts
- Adaptive Multi-Horizon Scheduling: It shifts the paradigm of time series forecasting from a static mapping to a decision-making and scheduling process.
Approved Open Questions
- Universal Scheduling Forecasting Bottleneck: This addresses the bottleneck of domain-specific training and enables better generalization, which is critical for scaling forecasting capabilities to heterogeneous real-world applications.
Links
Metadata & Links
- url
- https://arxiv.org/abs/2605.10292
- paper_id
- 2605.10292
- paper_source
- openalex
- domain
- time-series
- tags
-
- architectures
-
- datasets
-
- concept_slugs
- adaptive-multi-horizon-scheduling
- dataset_slugs
-
- skill
- TimeSeriesSkill
- created_at
- 2026-05-14T06:08:36Z