merlion: Time Series Intelligence

merlion is a Python library for time series intelligence. We support the following key features, each associated with its own sub-package:

merlion.models: A library of models unified under a single shared interface, with specializations for anomaly detection and forecasting. More specifically, we have
- merlion.models.defaults: Default models for anomaly detection and forecasting. These are good models for getting started.
- merlion.models.anomaly: Anomaly detection models
- merlion.models.anomaly.change_point: Change point detection models
- merlion.models.forecast: Forecasting models, including those which support exogenous regressors
- merlion.models.anomaly.forecast_based: Forecasting models adapted for anomaly detection. Anomaly scores are based on the residual between the predicted and true value at each timestamp.
- merlion.models.ensemble: Ensembles & automated model selection of models for both anomaly detection and forecasting.
- merlion.models.automl: AutoML layers for various models
merlion.dashboard: A GUI dashboard app for Merlion, which can be started with python -m merlion.dashboard. This dashboard provides a good way to quickly experiment many models on a new time series.
merlion.spark: APIs to integrate Merlion with PySpark for using distributed computing to run training and inference on multiple time series in parallel.
merlion.transform: Data pre-processing layer which implements many standard data transformations used in time series analysis. Transforms are callable objects, and each model has its own configurable model.transform which it uses to pre-process all input time series for both training and inference.
merlion.post_process: Post-processing rules to apply on the output of a model. Currently, these are specific to anomaly detection, and include
- merlion.post_process.calibrate: Rules to calibrate the anomaly scores returned by a model, to be interpretable as z-scores, i.e. as standard deviations of a standard normal random variable. Each anomaly detection model has a model.calibrator from this module, which can optionally be applied to ensure that the model’s anomaly scores are calibrated.
- merlion.post_process.threshold: Rules to reduce the noisiness of an anomaly detection model’s outputs. Each anomaly detection model has a model.threshold from this module, which can optionally be applied to filter the model’s predicted sequence of anomaly scores.
merlion.evaluate: Evaluation metrics & pipelines to simulate the live deployment of a time series model for any task.
merlion.plot: Automated visualization of model outputs for univariate time series
merlion.utils: Various utilities, including the TimeSeries class, resampling functions, Bayesian conjugate priors, reconciliation for hierarchical time series, and more.

The key classes for input and output are merlion.utils.time_series.TimeSeries and merlion.utils.time_series.UnivariateTimeSeries. Notably, these classes have transparent inter-operability with pandas.DataFrame and pandas.Series, respectively. Check this tutorial for some examples on how to use these classes, or the API docs linked above for a full list of features.

The full API documentation is outlined below:

merlion.models package