Salesforce Research Introduces ‘Merlion’, An Open-Source Machine Learning Library For Time Series Intelligence


Time series are a way to visualize and analyze the behavior of complex systems. They can represent key performance indicators for computing resources such as memory utilization or request latency, business metrics like revenue per day active users (DAU), marketing campaigns in social media reactions from customers through clickthrough rate. It is important to forecast the trends and values of key metrics accurately and rapidly detect any anomalies in those numbers. In software industries, anomaly detection is a critical machine learning technique to automate the identification of issues and improve IT system availability. It notifies the operator in a timely manner when something unexpected happens and helps them resolve underlying issues.

Salesforce research team introduces Merlion, an open-source Python library for time series intelligence. Merlion offers an end-to-end machine learning framework comprised of loading data, transforming it into useable formats, and building models. Once training has been completed, post-processing can include many aspects such as evaluating model performance or making predictions on what will happen in the future based on historical trends with some degree of accuracy.


Merlion provides a unified interface for models and datasets to detect anomalies on both univariate time series as well multivariate ones, along with standard pre/post-processing layers. Merlion is very intuitive in usage, especially in visualization, anomaly score calibration to improve interpretability, AutoML for hyperparameter tuning and model selection, and model ensembling. Merlion’s evaluation framework simulates the live deployment and re-training of a model in production. It can provide engineers with a one-stop solution for developing models with specific time series needs.

Key Takeaways

  • The proposed framework is an open-sourced, standardized, and easily extensible solution for data loading, preprocessing and benchmarking tasks related to time series forecasting. This way it can be used in a wide range of applications, including anomaly detection where state-of-art methodologies are urgently needed.
  • This suite of models provides a library for both anomaly detection and forecasting, united under one interface.
  • This library provides a new way to tune and select the best model for your data. AutoML allows you to automate hyperparameter tuning, reducing time spent on manual adjustments that experts in this field often require.
  • It provides a new set of easy-to-use ensembles that combine the outputs from multiple models for robust performance.
  • It provides native support for visualizing model predictions.
  • Univariate / Multivariate Forecasting
  • Univariate / Multivariate Anomaly Detection