sarkarghya/fast-st-us-geospatial-model

FAST-ST U.S. Geospatial Model

FAST-ST is a United States-only, connectable geospatial inference model. It combines a transferable temporal patch backbone, a sparse support-aware spatial adapter, and a joint low-rank Student-t output distribution. A typed housing artifact separately handles monthly Zillow ZORI rent and native-cadence annual Census ACS socioeconomic panels.

This is a numerical model system, not a live-data forecasting claim. Mutable sources are normalized into content-hashed snapshots before inference or training. Exact GIS/statistical tasks and event processes are routed to their own mathematical engines.

Released artifacts

• fast-st-us-v1: weather and traffic base (17,073,141 parameters)
• fast-st-us-housing-v1: monthly rent and annual socioeconomic model (18,799,208 parameters)

Frozen test results

The base improves aggregate MAE over persistence by 14.635% and over the temporal-only ablation by 5.867%. Its mean 80% interval coverage is 0.713562.

Monthly rent improves over the seasonal-naive baseline by 37.763%. Median household income improves over annual persistence by 70.69%.

Every checkpoint is published only when every manifest gate passes.

U.S. geography and data

The release covers 24 hard-coded U.S. metros including New York, Atlanta, San Francisco, Los Angeles, Chicago, Houston, Seattle, Boston, Miami, Washington, Philadelphia, Phoenix, Denver, Dallas, Minneapolis, Detroit, Portland, San Diego, Austin, New Orleans, Charlotte, Nashville, Las Vegas, and Salt Lake City.

Sources include pinned Hugging Face METR-LA and PEMS-BAY revisions, Open-Meteo/ERA5 weather, Zillow Research ZORI, and official Census ACS summary files. Source hashes and revisions are recorded in the manifests. Zillow data is not redistributed in this repository; use remains subject to Zillow's terms.

Limitations

• Predictive outputs are not causal effects.
• The base is validated on weather and two traffic networks, not every possible geospatial variable.
• Annual ACS releases are overlapping survey windows and remain annual.
• U.S. coordinate and country policy is enforced in ingestion and serving.
• Inspect per-series metrics: aggregate improvement does not mean every domain beats every baseline.

Files

Each model directory contains checkpoint.pt, manifest.json, and its generated model card. The manifest is the canonical record of data provenance, metrics, calibration, training hardware, and release gates.