Project / Spring 2026

Wildfire Proximity Prediction

Year

Spring 2026

Type

ML · Survival Analysis · Python

Skills

Python

Survival Analysis

XGBoost

Feature Engineering

Kaggle

Cox PH

Random Survival Forest

What I Built

Survival analysis predicting the probability a wildfire reaches an evacuation zone across four time horizons, from just 5 hours of early perimeter data.

Built a survival model in a Kaggle datathon to forecast wildfire evacuation-zone risk from the first 5 hours of fire data. Ranked 627 of 1,256 — middling — but the lessons about foundational ML knowledge have been the most durable ones I've picked up.

The dataset had 30+ features — fire growth metrics, centroid movement and direction, distance dynamics relative to evac zones, and temporal reliability flags marking which fires had sparse or unreliable observations. The outcome was right-censored: fires that didn't reach an evac zone within 72 hours had unknown time-to-hit values. My team's best score was 0.96157, ranking 627 out of 1,256. Not an impressive score — but the lessons from this project have been among the most durable ones I've picked up.

Decision Window

Forecasted risk 12–72 hours out from just 5 hours of fire data

Prediction Target

Probability a fire reaches within 5 km of an evacuation zone

Use Case

Supports earlier evacuation-risk triage and resource prioritization

POST-PROJECT REFLECTION

What I Took Away

What I learned. Ranges anywhere from technical skills, to general lessons.

My goal is to always learn from an experience, no matter how big or small.

I came into this class wanting to rival data scientists, thinking 7 weeks and some grit would close the gap. It didn't. The core issue was foundational ML knowledge. I was taking BUS316 (the supposed prerequisite) concurrently, so most of my modeling intuition was being built in real-time while I needed to use it. I didn't just lack advanced techniques — I lacked the vocabulary to diagnose why a model was failing. You feed data into XGBoost, the C-index improves 0.001, and then what? If you don't understand more complex ML concepts, asking an AI "how do I get a better score?" creates a loop: it suggests a technique, codes it for you, you submit, nothing changes, and you're back to square one. Because you don't know what lever to pull next.

In early March, I participated in a datathon at LMU — 4 hours, one problem, go. No weekly pacing, no "we'll get to it," just deadline pressure. My team ran a full ETL pipeline, EDA, model selection and refinement, and presentation all in 4 hours. The work was more thorough and more honest than what I produced across 7 weeks in class. Constraints forced prioritization. The lack of padding forced me to think instead of wander. That contrast sharpened a lesson I've tried to apply since: when you have a hard deadline, you find out what actually matters.

The meta-question that stuck: how does a business-focused student compete with more tech-heavy majors when the tool — AI — is equally available to both? The clear answer is knowing enough to evaluate what AI gives you, to recognize when it's hallucinating or over-fitting, to redirect when it's off track. Business intuition helps, but you still need the foundational knowledge to translate that intuition into model diagnostics. I didn't have a clean answer leaving the class. I still don't, entirely. But naming the question precisely is further than I was before — and I've been actively working on it since.

POST-PROJECT REFLECTION

What I'd Do Differently

Redos of project parts I was unsatisfied with, or simply things I would've done differently.

Redo 01

If I ran this project again, I'd structure it as three consecutive sprints with clear success criteria and hard milestones — not open-ended weeks. Sprint 1 (1 week): EDA & Feature Scoping — understand the data distribution, identify which features contain signal, decide which features to engineer. Success: a one-page summary of feature importance estimates (correlation, mutual information, or simple statistical tests) and a decision on which feature categories to prioritize. Sprint 2 (2–3 weeks): Model Exploration — test 3–4 candidates (Cox PH as baseline, RSF for nonlinearity, XGBoost AFT for gradient boosting) on a consistent validation setup. Success: a comparison table and a chosen model with documented rationale. Team play: assign each member a different model, compare results weekly. Sprint 3 (1–2 weeks): Refinement — hyperparameter tuning, feature interactions, calibration check.

Redo 02

On the technical side: I'd feature engineer more aggressively. We're not taught enough just how important feature engineering is. Engineered features demonstrate nonlinear relationships between variables better than any model can. The growth and distance categories were already the strongest predictors — there was more signal to extract there.

Redo 03

On the AI front: I'd "bag" the approaches. Run the same problem in 3–4 isolated AI chats, each with different system prompts or model contexts. Deliberately try opposing strategies — one asking for aggressive regularization, one for max performance, one for interpretability. Synthesize the best traits from each rather than iterating within a single compounding context where the AI builds on its own assumptions. That strategy didn't come from a textbook — it came from sitting with this problem and trying to reverse-engineer a fix.