CartoBoost RegimeMoEForecaster

Use RegimeMoEForecaster when one global model hides materially different geo-temporal regimes. The public surface exposes six named experts: stable recurring pattern, sparse cold-start, high-volume hub, volatile shock, long-distance pair, and low-signal fallback.

Python Example

from cartoboost.deep import RegimeMoEForecaster

model = RegimeMoEForecaster()
model.fit(
    features_train,
    duration_train,
    entity_ids=pickup_zone_ids,
    time_features=hour_day_features,
    recent_volatility=rolling_duration_volatility,
    graph_centrality=pickup_graph_centrality,
)

parts = model.predict_components(features_holdout, entity_ids=pickup_zone_ids_holdout)
prediction = parts["combined_prediction"]

GeoTemporalMixtureOfExperts, PairRegimeRouter, and EntityRegimeRouter are aliases for this first-cut MoE surface.

Browser WASM Example

RegimeMoEForecaster Wasm example

Runs the Rust-backed browser export on a substantial multi-route taxi panel with a held-out prediction window.

Ready to run in this page.

Use When

Use this model when distinct, repeatable regimes are plausible and expert usage can be inspected. Prefer one model when the router collapses.

Validation

Report router entropy, expert usage, combined RMSE, and a single-expert comparison under the same split. Treat degenerate expert usage as a failed MoE claim even if the aggregate error is acceptable.

Limitations

• Mixtures add router instability and expert-identifiability risk.
• Aggregate error can hide unused or redundant experts.
• Regime interpretation requires stability across seeds and cutoffs.