Combining Chronos forecasting with Discord-based Anomaly Detection

Time series data are pervasive across domains such as finance, healthcare, industrial monitoring, and climate science, where they support forecasting, classification, and anomaly detection tasks. Among these, anomaly detection is especially critical for ensuring reliability, safety, and efficiency, yet it remains challenging due to non-stationarity, noise, high dimensionality, and distribution shifts. Existing approaches include forecasting-based, reconstruction-based, encoding-based, and distance-based methods. Within the distance-based family, the Matrix Profile has become a widely adopted tool thanks to its efficiency, interpretability, and low parameter requirements. Extensions such as MERLIN and DAMP further improve scalability and robustness, particularly in large-scale or streaming settings. At the same time, recent progress in large language models (LLMs) has created new opportunities for time series analysis. Frameworks like SigLLM reformulate anomaly detection as a text-based task, while Chronos introduces a transformer-based architecture trained from scratch on time series data, achieving strong zero-shot forecasting performance. This thesis proposes a hybrid anomaly detection pipeline that bridges classical distance-based methods and LLM-inspired forecasting models. The approach combines DAMP, which efficiently identifies discords—subsequences most dissimilar to the rest of the signal—with Chronos, which provides accurate forecasts. Candidate anomalies discovered by DAMP are validated using Chronos residuals, enabling precise localization of anomalous regions while avoiding unnecessary computation on normal subsequences. Experiments on benchmark datasets such as Yahoo S5, NAB, and UCR evaluate the effectiveness of this integration, with DAMP serving as the primary baseline. Results are expected to show that the proposed method unites the scalability of distance-based algorithms with the accuracy of pretrained forecasting models, providing a novel and robust solution for time series anomaly detection.