Title: Metric-Driven Hybrid Forecasting Strategy for the CTF4Science Lorenz Challenge

Abstract:

This paper outlines our methodology for the CTF4Science Lorenz challenge, a benchmark designed to test short-horizon prediction, long-term distribution alignment, and trajectory reconstruction across nine distinct task pairs. Our primary finding indicates that no single model architecture excelled across all evaluation metrics. Consequently, we developed a metric-aware hybrid framework that deploys specific predictors tailored to different metric categories: synthetic-pretrained denoisers were utilized for complete trajectory reconstruction; Lorenz ODE fitting combined with trajectory shooting was applied to the initial 20 forecasting steps; and histogram-tail substitution, leveraging synthetic Lorenz libraries, was employed for long-term performance assessment. A mature iteration of this system achieved a score of 83.83551 on the public leaderboard, while a subsequent, smaller implementation of the same principles reached 83.85529. This article concentrates on the intermediate system, which offers a clear representation of the full methodology while maintaining simplicity for reproduction and analysis, whereas the final submission serves as a conservative extension of this core architecture.

Source: arXiv Generated at: 2026-06-04 00:00:00 UTC