# Top high school team solution

I found this competition on LinkedIn. The question,
how soon will a fire reach people?
, felt worth spending time on. I had never used survival analysis before; I learned it while building this.
The problem.
316 fires (221 train / 95 test), each described by features from the first five hours after detection. The task is to predict, for four horizons (12/24/48/72h), the probability the fire reaches within 5 km of an evacuation zone. The data is right-censored: a fire that never threatened within 72h has an unknown true outcome.
Survival framing with proper censoring.
Each horizon is a separate binary head, but censored fires are masked out (
NaN
) at horizons where their outcome is genuinely unknown, instead of being treated as negatives. This is the part that makes it survival analysis rather than plain classification, and it's the choice I'm most confident about.
Leak-free OOF isotonic calibration.
I calibrate the probabilities with isotonic regression fit only on out-of-fold predictions. In my own evaluation, this was the single change that actually improved the score — which makes sense, since the Brier term carries 70% of the hybrid metric.
An ensemble I tested and chose not to rely on.
I built three diverse LightGBM configs (shallow / deep-regularised / extra_trees) and blended them by simplex search on OOF Brier. The theory is that diverse models average out their errors. But my out-of-fold results were honest with me: on a dataset this small, the blend didn't help — the best single config matched or slightly beat it. I kept the experiment in the notebook because reading your own results honestly matters more than the size of the model.
Robustness.
5-fold CV with fold ensembling (an honest estimate over 100% of the data), and an O(n log n) Fenwick-tree rewrite of the C-index.
What I learned.
With only 69 hits, almost all before 72h, the 72h horizon has essentially zero negatives — no algorithm can learn from data that isn't there. And the bigger lesson: I expected the ensemble to win, the numbers said otherwise, and trusting the numbers taught me more than any single trick did.