Abstract
Anomaly detection in time series data is a critical challenge across variousdomains. Traditional methods typically focus on identifying anomalies inimmediate subsequent steps, often underestimating the significance of temporaldynamics such as delay time and horizons of anomalies, which generally requireextensive post-analysis. This paper introduces a novel approach for time seriesanomaly prediction, incorporating temporal information directly into theprediction results. We propose a new dataset specifically designed to evaluatethis approach and conduct comprehensive experiments using severalstate-of-the-art methods. Our results demonstrate the efficacy of our approachin providing timely and accurate anomaly predictions, setting a new benchmarkfor future research in this field.