# WiDS Global Datathon 2026

My solution framed the challenge as a censor-aware survival prediction problem, building calibrated and monotonic wildfire threat probabilities for 12h, 24h, 48h, and 72h from only the first five hours of fire activity.
The final winning submission,
submission_domain_anchor_blend_95_5.csv
, used a conservative 95/5 blend of two survival-superlearner branches: a domain-specialist survival stack and an initial mainline survival stack. After blending, I enforced row-wise monotonicity so that each event’s predicted threat probability increased consistently across time horizons.
This approach prioritized stable generalization on a very small right-censored dataset, focusing on survival-aware modeling, calibration at the metric-heavy mid horizons, robust tabular ensembling, and strict probability validity.
Reproducibility materials are available here:
Kaggle Model:
[https://www.kaggle.com/models/luciayen/wids-2026-winning-survival-blend/ScikitLearn/winning-submission-package/3](https://www.kaggle.com/models/luciayen/wids-2026-winning-survival-blend/ScikitLearn/winning-submission-package/3)
Kaggle Reproducibility Notebook:
[https://www.kaggle.com/code/luciayen/wids-2026-winning-survival-blend-reproducibility](https://www.kaggle.com/code/luciayen/wids-2026-winning-survival-blend-reproducibility)
GitHub Repository:
[https://github.com/Yen-hub/Predicting-Time-to-Threat-for-Evacuation-Zones-Using-Survival-Analysis/tree/codex/winner-release-package](https://github.com/Yen-hub/Predicting-Time-to-Threat-for-Evacuation-Zones-Using-Survival-Analysis/tree/codex/winner-release-package)