Warning
This is a quick implementation over the weekend, expect bugs, unpolished and untested code.
Implementation of the peaks-over-threshold (POT) algorithm to detect extreme values in time series data, including Streaming POT (SPOT) and Streaming POT with drift (DSPOT).
The implementation follows the paper: Siffer, Alban, et al. "Anomaly detection in streams with extreme value theory." Proceedings of the 23rd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining. 2017.
Note
Unlike the original paper's implementation,
this implementation reuses L-BFGS-B minimization results to speed-up the Grimshaw procedure.
Note
If you encounter runtime warnings saying invalid values, it might be because of number underflow. Try casting input data to double-precision numbers. Additionally, if the time series values are too large, it might cause some issues with the upper and lower bounds of the Grimshaw procedure; in that case, try rescaling the time series to some smaller values depending on the domains.
Sample Results
Please check the examples folder for more results.
