Luca Baggi - How I used Polars to build built functime, a next gen ML forecasting library

Functime is a forecasting library built on Polars that enables efficient processing of multiple time series without requiring distributed systems

Key advantages of using Polars for forecasting:

• Multi-threaded query engine
• Lazy optimizations
• Apache Arrow format for fast I/O
• Query optimization leveraging relational database concepts

Global forecasting approach allows handling thousands of time series efficiently:

• Fits one model on the entire dataset instead of individual models
• Works well for 90% of common forecasting workflows
• Particularly effective for dozens to hundreds of time series

Core features of Functime:

• Scikit-learn compatible API
• Built-in diagnostics and visualization tools
• Automatic timestamp alignment
• Native support for panel datasets
• Conformal prediction intervals

Design principles:

• Minimal dependencies for easy deployment
• Focus on practical scale (hundreds/thousands of series) rather than extreme scale
• Emphasis on the entire forecasting workflow, not just model accuracy
• Automated feature engineering and preprocessing

Future developments planned:

• Probabilistic forecasting capabilities
• Rank-based interval forecasts
• Enhanced plotting features
• Integration with FLAML for hyperparameter tuning
• Improved feature extraction methods