## Problem

When a wildfire ignites, emergency managers and responders must decide which communities to warn, when to warn them, and where to position scarce resources. Decisions must be made before certainty is available. The response requires both prioritization (which fires are most urgent soon) and calibrated risk estimates (how likely a fire is to threaten evacuation zones within actionable time windows). 

This competition turns that operational need into a survival analysis challenge. You will generate calibrated probability forecasts across multiple time horizons to support real-world decisions.

## Real-World Context
Many wildfire forecasting approaches reduce the task to a single question. Will this fire become dangerous. Emergency response needs more. Decisions are time-bound and comparative. How soon is a threat likely. How confident should we be. Which incident should move to the front of the queue when crews, aircraft, and messaging capacity are limited.

This competition addresses that gap by framing the task as survival analysis using real early-incident signals. The dataset captures wildfire perimeter dynamics and their spatial relationship to evacuation zones. Features are computed strictly from the first five hours after the initial perimeter observation, t0. The target is the time from t0 plus 5 hours until the fire comes within 5 km of any evacuation zone centroid.

## Survival Analysis Framing and Censoring
This is right-censored survival analysis:
- If a fire hits within the 72-hour window, event = 1 and time_to_hit_hours is the observed time from t0+5h.
- If a fire does not hit within 72 hours, event = 0 and time_to_hit_hours is the last observed time in the window (<= 72).

Participants do not predict a single time. They submit survival probabilities that a hit occurs by 12h, 24h, 48h, and 72h. The strongest solutions will deliver two outcomes at once. They will rank fires correctly by urgency for triage and produce calibrated probabilities that can support threshold-based decisions.