Catastrophe Loss Reserving Model for Storm Surges with XGBoost and Weather Telemetry
Insurers experience sudden, massive liability exposures immediately following El Niño winter storm surges. Traditional actuarial loss reserving relies on delayed loss reports that take weeks to accumulate. Predictive machine learning models using early weather radar signals and coastal elevation profiles can estimate final loss ratios within hours of storm landfall. In this challenge, you will develop a gradient boosting loss estimation model using XGBoost. You will extract spatial feature vectors from weather radar wind speeds, central barometric pressures, property exposure databases, and coastal surge heights to predict aggregate portfolio claims. Your model will incorporate cross-validation across historical El Niño storm events, produce calibrated loss distributions, and extract SHAP feature importance for actuarial transparency.
What you are building
The core problem, expected build, and operating context for this challenge.
Build an XGBoost supervised model to forecast catastrophe loss ratios from early post-landfall storm surge weather telemetry.
How work is evaluated
Evaluates model loss ratio prediction accuracy across test catastrophe events.
Challenge datasets
Public datasets, plus the private datasets you have access to.
How submissions are scored
These dimensions define what the evaluator checks and which criteria separate a passable run from a strong one.
valid_range_check
Loss ratio prediction falls within logical domain limits [0.0, 1.0]
This dimension contributes its full weight only when the submission satisfies the requirement. Partial credit is not awarded.
loss_ratio_mae
Mean Absolute Error on predicted portfolio loss ratio • target: 0.015 • range: 0-0.5
This dimension contributes its full weight only when the submission satisfies the requirement. Partial credit is not awarded.
What you should walk away with
Train and tune an XGBoost regression model on catastrophe exposure tabular data.
Engineer spatial features from raw barometric pressure and wind velocity data.
Interpret model predictions using SHAP (SHapley Additive exPlanations).
Evaluate loss ratio error across out-of-time El Niño storm events.
Reference links and supporting material
Synthetic historical dataset of storm landfall events, physical meteorology metrics, and aggregate insurance loss ratios.
Challenge setup
Evaluates model loss ratio prediction accuracy across test catastrophe events.
Challenge input
JSON with portfolio_exposure_usd, max_surge_m, min_pressure_hpa, peak_wind_kts
XGBoost
Standard high-performance library for tabular machine learning.
SHAP
Provides feature attribution for black-box regression models.
Evaluated output
JSON with predicted_loss_ratio, estimated_reserve_usd, primary_driver
- Loss ratio prediction falls within logical domain limits [0.0, 1.0]
- Mean Absolute Error on predicted portfolio loss ratio • target: 0.015 • range: 0-0.5
- Benchmark: CatLossBench
- Loss Ratio Mae target: 0.015
- 1 public reference case
- Python execution harness
- Python sandbox (unavailable on Versalist)
Summary shows the first task, up to three tools, and up to three evidence items. Open the technical recipe to view all configured items.
View technical recipe
Configured tools
- Mistral 7B · Required
- OpenAI · Optional
- Cohere Platform · Optional
Evaluation contract
- valid_range_check · Weight 1
- loss_ratio_mae · Weight 1
Metrics
- loss_ratio_mae · Target 0.015
Benchmarks
- CatLossBench
- general
Reference cases
- 1 public reference case
Workflow tasks
- Task 1
Input: JSON with portfolio_exposure_usd, max_surge_m, min_pressure_hpa, peak_wind_kts
Output: JSON with predicted_loss_ratio, estimated_reserve_usd, primary_driver
Execution harness: Python
Evaluation tools
- XGBoost · Gradient Boosting Framework
- SHAP · Explainability Library
Recipe freeze marker
No recipe freeze time is recorded.