Quiet seasons deserve honest numbers

How we hold the numbers

Precise about what is calibrated

The fire and grid component scores are calibrated model probabilities. The compound score is a weighted index over those components: it ranks regions, it is not a probability and it is not calibrated. Prediction intervals are conformal, so their coverage comes from the method rather than from a claim we make about it.

Compound, not single-hazard

Fire, grid stress, and drought interact. RiskMesh trains a separate model for each component, then combines the three into one weighted index that rises when more than one hazard is elevated at the same time. That combined view is the one that matters when it matters.

Explainable, not magic

Every prediction has receipts: which features moved the score, by how much, and what the inputs looked like. SHAP attributions, not vibes.

Honest about novelty

If the inputs look unlike anything the model trained on, the OOD flag fires. The score still renders; you simply know to be skeptical of it.

How it works

1. 01Pull upstream feedsWeather, vegetation, drought monitor, fire detections, and grid load from authoritative sources. Cached with TTLs that match each feed's update cadence.
2. 02Compute features per regionVPD, dry-day counts, soil moisture, wind anomalies, grid load deviation, drought category: about 40 features per component model.
3. 03Score with three component modelsFire, grid, and drought each get their own XGBoost regressor. Conformal residuals from the last few thousand predictions become the interval.
4. 04Combine into the compound indexThe three component scores are folded into a single weighted index. It is the number that fronts the dashboard, and it is a ranking rather than a calibrated probability.
5. 05ReceiptsTop-3 SHAP features and raw inputs are stored alongside the score. Anyone can audit any prediction at /receipts/{id}.

Built on open data and open source

Component models

XGBoost · gradient-boosted trees

Compound score

Weighted index over component scores

Intervals

Split conformal · 80% and 90%

OOD

Mahalanobis distance · 99th-pctile cutoff

Explainability

TreeSHAP · per-prediction attributions

Backend

FastAPI · Postgres · Railway

Frontend

Next.js 14 · React 18 · Vercel

Data

NOAA · USDM · FIRMS · EIA · Open-Meteo

Code

Repository withheld