// the find
salesforce/Merlion
Merlion: A Machine Learning Framework for Time Series Intelligence
Merlion is Salesforce's time series ML library covering forecasting, anomaly detection, and change point detection under a single API. It targets data scientists and ML engineers who want a single library instead of stitching together Prophet, statsmodels, and sklearn individually. The 2021 paper is still the primary citation, which tells you something about the pace of development.
The unified model interface is genuinely well-designed — switching from IsolationForest to LSTM-based detection is a config change, not a rewrite. The evaluation pipeline that simulates live retraining with configurable cadence is the most practically useful part; most frameworks give you metrics on a static split and call it done. The post-processing layer for anomaly scores (calibration, thresholding) addresses a real problem that competing libraries ignore entirely. PySpark backend for scaling the same model code to industrial data volumes without rewriting anything is a meaningful differentiator.
The last commit is over a year old and open issues are piling up — this is effectively in maintenance mode, and the 'To Dos' section still says 'stay tuned' about GPU support that never arrived. Some models require a JDK at runtime, which is an awkward dependency to carry for what is otherwise a Python data science package. The AutoML component is limited to SARIMA and Prophet variants; it won't touch the deep learning models, so the headline AutoML feature applies to maybe 20% of the model zoo. Installation complexity is real: optional extras, conda-specific OpenMP steps, and the ts_datasets sub-package that must be installed in editable mode — this will frustrate anyone not already living in a managed conda environment.