Researchers have developed a novel dynamic multi-step decoding strategy for time-series foundation models, aiming to improve forecasting accuracy. This approach adaptively selects the optimal decoding method for each forecasting instance, rather than relying on a fixed strategy. By training lightweight selector networks that rank candidate rollout schedules based on the backbone model's hidden states, the system can improve point-forecast skill, particularly for longer horizons. This method requires no fine-tuning of the forecasting backbone itself and has shown significant gains on benchmarks like fev-bench when used with models such as Chronos-2 and Moirai-1.1-R-small. AI
IMPACT This new decoding strategy could enhance the accuracy and adaptability of time-series forecasting models across various applications.
RANK_REASON Academic paper detailing a new method for time-series forecasting models. [lever_c_demoted from research: ic=1 ai=1.0]
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →
