Automated financial time series anomaly detection via curiosity-guided exploration and self-imitation learning

Financial multivariate time series anomaly detection has received widespread attention, which is crucial for the reliable and safe operation of the financial system. Based on this, multi-variable time series anomaly detection methods based on different network architectures have emerged, but these models are either based on a single anomaly hypothesis (such as point anomalies) or based on labeled data (such as supervised methods). This goes against the multimodal nature and complexity of real-world data, making these methods ineffective. To remedy these shortcomings, we propose a model selection framework based on curiosity search and experience replay mechanisms.This framework includes multiple anomaly assumptions and base models in a carefully designed search space. The agent explores this search space to guide each base model, leveraging its strengths in exotic types of time series data. The framework subsequently employs reinforcement learning to select the optimal model from these candidate models. Through multi-dimensional, multi-index experimental verification of four different types of data sets, our model search framework can select the optimal base model suitable for the current timestamp from the base model pool, thereby achieving superior detection performance and better than the current A model selection framework based on reinforcement learning.Additionally, we introduce self-imitation learning and utilize an experience replay mechanism to enhance the effectiveness of samples, thereby saving detection time. To better extend our approach to more complex real-world applications, we conduct interpretability analysis on the output of anomalies. This research assists in troubleshooting and analyzing anomalies.