How we Automated Forecasting at Chime

By Li Wang, Senior Data Scientist at Chime^®

Every company forecasts, whether it’s revenue, member counts, or transaction volume. For a long time, Chime’s forecasting process relied on manual notebooks and spreadsheets. Forecasts took hours to run, were difficult to reproduce, and required significant hands-on work from model builders. 

Our goal was to change that. We wanted a forecasting system that could run automatically, scale across many different models, and enable data scientists to focus on insights rather than infrastructure. This would allow more metrics to be forecasted with high accuracy in a shorter time and with a limited headcount, which in turn would empower business stakeholders to make data-driven decisions ahead of the game. That vision became AutoForecast. 

AutoForecast is a modular, multi-algorithm forecasting engine designed to automate model selection, tuning, and deployment within Chime’s ML ecosystem. 

My team, DSML (Data Science and Machine Learning), deploys models using MLKit, Chime’s internal machine-learning platform. MLKit standardizes how models are trained, evaluated, and deployed using configuration files,  making development reproducible and scalable. But forecasting was not originally supported within MLKit. 

To address growing forecasting needs, enable automation, and work efficiently with limited resources, we created a four-stage roadmap to gradually integrate forecasting into MLKit.

Stage 1: Laying the Foundation 

In Stage 1, we deployed MLKit’s first forecasting model using a lightweight, config-driven workflow to support Chime’s top-down forecasting architecture. Along the way, we built processes for data exploration, rigorous backtesting, and performance visualization, and we fully automated pipeline execution with data-quality checks. 

This stage replaced hours of manual work with an automated, reproducible forecasting pipeline—our first major step toward scalable forecasting at Chime.

Stage 2: Enabling Hierarchical Forecasting 

As forecasting use cases expanded, Stage 2 introduced support for hierarchical forecasting, allowing model builders to define multiple segments directly in configuration files. During this stage, we also developed the first version of AutoForecast, which was quickly adopted across several model projects. 

AutoForecast’s initial release supported multiple forecasting algorithms and automatically tuned them, enabling more consistent and efficient modeling across diverse forecasting tasks. 

Stage 3: Integrating Forecasting into MLKit 

Stage 3—an ongoing phase of the process—focuses on integrating AutoForecast into MLKit as a supported learner, enabling fully configuration-based exploration, training, and deployment. This is the most technically complex stage, requiring AutoForecast to adopt MLKit’s abstracted learner interface, maintain backward compatibility, and pass all unit and integration tests. 

Once complete, forecasting will use the same production workflows as any other ML model,  including versioning, observability, and standardized deployment. This will standardize forecasting within MLKit, eliminating the need for custom project-specific workflows and reducing friction for model builders. 

Stage 4: Prototyping Intelligent Automation 

Stage 4 introduces a proof-of-concept system that automatically evaluates multiple forecasting algorithms and selects the one that performs best based on backtest metrics. In one real deployment, the prototype generated the majority of forecasts automatically, reducing manual tuning and highlighting the potential for scalable, consistent forecasting. 

While this prototype is not yet fully productionized, it demonstrates a promising direction for further automation and efficiency. 

Closing

Building AutoForecast through the first two stages and prototyping the more advanced stages has provided valuable insights into designing scalable forecasting systems, integrating diverse model families, and aligning forecasting workflows with MLKit’s platform standards. 

Although the remaining work will continue evolving with future contributors, the foundation is now in place for Chime to expand its forecasting capabilities with greater speed, accuracy, and automation. AutoForecast establishes a flexible framework that future teams can build upon as forecasting needs grow across the company.