LeapTS: Rethinking Time Series Forecasting as Adaptive Multi-Horizon Scheduling
Researchers have introduced LeapTS, a novel framework that reformulates time series forecasting from a fixed mapping process into a dynamic scheduling mechanism. Traditional models often suffer from temporal decoupling and limited adaptability to evolving contexts. LeapTS addresses these limitations by organizing forecasting into multi-level decisions using a hierarchical controller for optimal prediction scale selection and continuous-time state evolution driven by neural controlled differential equations. This approach explicitly couples irregular temporal dynamics with discrete scheduling feedback. Extensive evaluations on real-world and synthetic datasets demonstrate that LeapTS improves overall forecasting performance by at least 7.4% compared to existing methods. Additionally, it achieves a significant inference speedup, ranging from 2.6x to 5.3x faster than representative Transformer-based models. The study also reveals how the model autonomously adapts its behavior to capture non-stationary dynamics by tracing scheduling trajectories. This advancement offers substantial improvements for applications requiring resource optimization and decision-making support, marking a significant step forward in machine learning architectures for time series analysis.
Wire timeline
LeapTS: Rethinking Time Series Forecasting as Adaptive Multi-Horizon Scheduling
Researchers have introduced LeapTS, a novel framework that reformulates time series forecasting from a fixed mapping process into a dynamic scheduling mechanism. Traditional models often suffer from temporal decoupling and limited adaptability to evolving contexts. LeapTS addresses these limitations by organizing forecasting into multi-level decisions using a hierarchical controller for optimal prediction scale selection and continuous-time state evolution driven by neural controlled differential equations. This approach explicitly couples irregular temporal dynamics with discrete scheduling feedback. Extensive evaluations on real-world and synthetic datasets demonstrate that LeapTS improves overall forecasting performance by at least 7.4% compared to existing methods. Additionally, it achieves a significant inference speedup, ranging from 2.6x to 5.3x faster than representative Transformer-based models. The study also reveals how the model autonomously adapts its behavior to capture non-stationary dynamics by tracing scheduling trajectories. This advancement offers substantial improvements for applications requiring resource optimization and decision-making support, marking a significant step forward in machine learning architectures for time series analysis.
cs.AI updates on arXiv.org