A comparative study of time series anomaly detection techniques for client transactions.

Detecting anomalies in client transactions is a critical task for identifying financial crimes such as money laundering. This study investigates the suitability of various machine and deep learning models for outlier detection in unlabeled time series data. A comparative analysis is conducted to evaluate the models' performance in terms of running time, interpretability, and consistency of the results as compared to other methods. Random samples of data are manually labeled by experts to conduct more tests and leverage the benefits of supervised learning validation. Results show that for monthly aggregated data, the Prophet model combined with z-score analysis of residuals is the most effective and robust approach. Furthermore, increasing the amount of data available favors deep learning models due to their ability to process complex and high-dimensional data. Overall, this study provides insights into the strengths and weaknesses of different anomaly detection methods and their potential applications in the finance industry.