CartoBoost Documentation

CartoBoost is a Python toolkit for structured regression, classification, ranking, and forecasting when time, location, route structure, or repeated IDs are part of the signal. These docs are organized around practical modeling workflows: preparing data, choosing an estimator, validating it without leakage, interpreting results, and deploying reproducible artifacts.

What It Is Good For

• Row-level tabular modeling for fare, duration, risk, or demand targets.
• Forecasting demand and count series with rolling-origin validation.
• Graph modeling when direction or relationship structure matters.
• Learned ID embeddings when stable identifiers carry repeated residual signal.
• Leakage-aware evaluation against strong baselines on the same split.

Choose Your Path

• First model: Getting Started covers installation, fitting, validation, baseline comparison, and persistence.
• Model selection: Choose A Model maps data and prediction tasks to the appropriate estimator family.
• Tabular and spatial ML: Boosting Model Guides cover regression, classification, ranking, categorical data, and structured splits.
• Time series: Forecasting covers validated frames, rolling-origin backtesting, metrics, artifacts, and CLI workflows.
• Production integration: Python API, CLI Reference, and Model Artifacts.
• Evidence: Benchmarks reports commands, datasets, splits, metrics, and limitations.

Model Families

• CartoBoost Boosting Model Guides: row-level tree models.
• CartoBoost Forecasting Model Guides: one guide per forecast family.
• Geo-Causal Experiment Models: synthetic DID, GeoLift-style design, and spillover diagnostics.
• CartoBoost Graph Model Guides: directed movement, link prediction, and graph features.
• CartoBoost Neural Model Guides: standalone ID embeddings and embedding features.
• Benchmark Overview: current benchmark evidence and limits.

Stable estimators are available from the package root. The cartoboost.models registry provides machine-readable metadata for tools that need to enumerate supported model surfaces.

Reference

• Forecasting: frame contracts, backtesting, artifacts, and shared forecast rules.
• Feature Catalog: full capability map.
• Sparse Features: sparse sets, H3/S2 point cells, and decoded route-cell encoders.
• Python API: public classes and methods.
• CLI Reference: command behavior.

Install

uv add cartoboost

Optional dependencies are installed directly:

uv add shap optuna polars onnx