CartoBoost SpatioTemporalGraphForecaster

Use SpatioTemporalGraphForecaster when node-time targets live on directed weighted edges and the graph is part of the forecast hypothesis. It is the generic deep-model facade for graph sequence forecasting.

Python Example

import numpy as np
from cartoboost.deep import SpatioTemporalGraphForecaster
from cartoboost.forecasting import GraphTemporalFrame

target = np.array(
    [
        [12.0, 8.0, 5.0],
        [14.0, 9.0, 6.0],
        [18.0, 12.0, 7.0],
        [21.0, 16.0, 9.0],
        [19.0, 17.0, 11.0],
        [16.0, 15.0, 12.0],
    ],
    dtype=float,
)

frame = GraphTemporalFrame(
    node_ids=["node_a", "node_b", "node_c"],
    timestamps=list(range(target.shape[0])),
    target=target,
    indptr=[0, 2, 3, 3],
    indices=[1, 2, 2],
    data=[0.7, 0.3, 1.0],
    horizon=2,
    frequency="hourly",
)

model = SpatioTemporalGraphForecaster(
    backbone="dcrnn",
    diffusion_steps=2,
    hidden_size=8,
)
model.fit(frame)
forecast = model.predict(2)

Browser WASM Example

SpatioTemporalGraphForecaster 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.

When To Use

• The target is a regular node-time matrix.
• Directed weighted edges are known before the forecast cutoff.
• Neighboring nodes plausibly influence each other over time.
• You need graph-aware forecasts rather than graph embeddings for rows.

Use When

Need	Better first choice
Node-time forecasting on directed weighted edges.	SpatioTemporalGraphForecaster
Forecasting with the explicit DCRNN API.	DCRNNForecaster
Row-level graph regression or link scoring.	Graph model guides
Panel forecasting without adjacency.	NeuralPanelForecaster or CartoBoostLagForecaster

Validation

Use rolling-origin validation and compare against seasonal naive, CartoBoostLagForecaster, and a panel neural model when available. Report errors by horizon and node, and keep the graph restricted to information known at the cutoff.

Limitations

• Graph construction and edge weights can dominate model behavior.
• Missing nodes or changing topology need explicit handling.
• Graph models add compute cost and should beat graph-free panel baselines on external origins.