Abstract
SUMMARY & CONCLUSIONSResilience engineering plays a critical role in ensuring the sustained performance of complex systems under disruptions, with predictive modeling serving as a cornerstone for proactive decision-making. Accurate long-term prediction of resilience metrics, such as recovery duration or performance degradation, remains a persistent challenge due to the dynamic behavior of systems and uncertainty in future influencing factors. Traditional predictive models often assume full knowledge of future covariates, which limits their reliability and practical applicability, especially in real-world forecasting scenarios. In this study, we address this gap by proposing a forecasting framework that improves prediction accuracy while accounting for the unknown nature of future covariates. Using a real-world dataset spanning ten years of crude oil prices and six historical covariates, we implement a disjoint sliding-window approach with TimeGPT, a state-of-the-art time series foundation model. Our framework exclusively utilizes historical covariates, ensuring that predictions are made under realistic constraints without access to future inputs. The results demonstrate stable and competitive predictive performance across various non-overlapping forecasting window configurations, with the best results achieved using the shortest window of 10 time points and a balanced 5-5 split between training and testing. In this setting, the model reached a mean absolute error of 1.2 and an R 2 value of up to 0.99, highlighting its robustness in long-term prediction under partial information.