This competition focused on predicting the probability that a wildfire would threaten an evacuation zone within 12h, 24h, 48h, and 72h using only the first five hours of wildfire behavior data.
\nI approached the problem as a right-censored survival analysis task since the dataset contained both observed event times and censored wildfire events.
\nThe final solution combined:
\nThe solution achieved 4th place globally.
\nThe most important observation was that wildfire threat progression was primarily driven by:
\nInstead of broad feature expansion, I focused on modeling the physical relationship between wildfire spread and evacuation-zone proximity.
\nI created a compact set of physics-inspired features designed to capture wildfire dynamics, including:
\nThese features helped the survival models better capture urgency and risk progression while avoiding overfitting on the small dataset.
\nThe final ensemble used three survival models:
\nGradient Boosting Survival Analysis (GBSA)
\nThe strongest standalone model, particularly effective at capturing nonlinear wildfire behavior.
\nRandom Survival Forest (RSF)
\nAdded model diversity and improved robustness across different wildfire regimes.
\nCoxNet
\nProvided a more linear survival perspective and performed especially well for far-distance fires.
\nI used 5-fold cross-validation to:
\nTest predictions were averaged across folds.
\nOne of the most impactful improvements came from stratifying fires into:
\nusing tertiles of the minimum evacuation-zone distance feature.
\nInstead of using one global ensemble weight, I optimized separate model weights for each distance bin using Optuna.
\nThis worked well because different models performed better in different distance regimes:
\nThe final solution applied a simple rule:
\nAlthough simple, this consistently improved leaderboard performance because the 72h horizon behaved differently from shorter horizons and was strongly skewed toward positive outcomes among observed events.
\nThe model predictions were still fully used for the 12h, 24h, and 48h horizons.
\nFinal predictions were constrained to ensure:
\nThis ensured logical consistency across time horizons.
\nSeveral approaches were tested but did not improve the final score, including:
\nThe simpler distance-stratified survival ensemble remained the strongest solution.
\nThis competition focused on predicting the probability that a wildfire would threaten an evacuation zone within 12h, 24h, 48h, and 72h using only the first five hours of wildfire behavior data.
\nI approached the problem as a right-censored survival analysis task since the dataset contained both observed event times and censored wildfire events.
\nThe final solution combined:
\nThe solution achieved 4th place globally.
\nThe most important observation was that wildfire threat progression was primarily driven by:
\nInstead of broad feature expansion, I focused on modeling the physical relationship between wildfire spread and evacuation-zone proximity.
\nI created a compact set of physics-inspired features designed to capture wildfire dynamics, including:
\nThese features helped the survival models better capture urgency and risk progression while avoiding overfitting on the small dataset.
\nThe final ensemble used three survival models:
\nGradient Boosting Survival Analysis (GBSA)
\nThe strongest standalone model, particularly effective at capturing nonlinear wildfire behavior.
\nRandom Survival Forest (RSF)
\nAdded model diversity and improved robustness across different wildfire regimes.
\nCoxNet
\nProvided a more linear survival perspective and performed especially well for far-distance fires.
\nI used 5-fold cross-validation to:
\nTest predictions were averaged across folds.
\nOne of the most impactful improvements came from stratifying fires into:
\nusing tertiles of the minimum evacuation-zone distance feature.
\nInstead of using one global ensemble weight, I optimized separate model weights for each distance bin using Optuna.
\nThis worked well because different models performed better in different distance regimes:
\nThe final solution applied a simple rule:
\nAlthough simple, this consistently improved leaderboard performance because the 72h horizon behaved differently from shorter horizons and was strongly skewed toward positive outcomes among observed events.
\nThe model predictions were still fully used for the 12h, 24h, and 48h horizons.
\nFinal predictions were constrained to ensure:
\nThis ensured logical consistency across time horizons.
\nSeveral approaches were tested but did not improve the final score, including:
\nThe simpler distance-stratified survival ensemble remained the strongest solution.
\n