Predicting
Flood Claim Severity
for the NFIP

Why Flood Claim Prediction Matters

The National Flood Insurance Program (FEMA) is the primary source of flood insurance in the U.S. With over 5 million active policies and $70B+ in historical payouts, accurately predicting individual claim severity is a high-value operational problem — it drives claims triage, reinsurance reserving, and portfolio risk management.

Target variable: amountPaidOnBuildingClaim — actual NFIP payout for structural damage, ranging from $0.01 to $250,000.

End-to-End ML Pipeline

FEMA NFIP Redacted Claims V2

Historical flood insurance claims from FEMA covering multiple decades, 50+ states, and every flood zone category. Starting from 73 raw columns, we retained a clean working dataset after removing leakage, high-cardinality identifiers, and invalid records.

8 Questions That Shaped the Model

Each EDA question was chosen to directly inform modelling decisions — not just describe the data, but test whether a feature is worth including and in what form.

EDA 01

What Does the Target Variable Actually Look Like?

Distribution of amountPaidOnBuildingClaim — raw, log-transformed, and by frequency tier.

Key Finding

The raw distribution is severely right-skewed — most claims are small ($500–$15K) but extreme claims up to $250K exist. Log-transformation converts this into a roughly bell-shaped distribution, which matters for Linear Regression's error assumptions. The small-to-moderate claims dominate frequency, meaning a model that only predicts large losses accurately would fail on the majority of real-world claims.

EDA 02

When Do Floods Cause the Most Severe Damage?

Temporal patterns: claim volume and severity by month, season, and loss year.

Key Finding

Claims are dramatically concentrated in August–October — peak Atlantic hurricane season. The 2005 (Katrina), 2012 (Sandy), and 2017 (Harvey, Irma, Maria) hurricane years are visible as distinct severity spikes. is_hurricane_season shows one of the strongest linear correlations with the target (0.239), directly justifying its inclusion as a binary feature.

EDA 03

Does Water Depth Actually Predict Claim Size?

Depth-damage relationship — raw vs. log-transformed, and the non-linear curve.

Key Finding

The relationship between water depth and claim amount is non-linear and concave — a classic depth-damage curve from flood engineering. The first few inches of water cause a disproportionate jump in losses; marginal damage diminishes at depth extremes. This non-linearity directly motivates the water_depth_log engineered feature, which improves linear correlation from 0.156 (raw) to 0.248 (log).

EDA 04

Do FEMA's Flood Zones Predict Losses?

Claim severity and volume by flood zone category — testing whether FEMA's risk hierarchy is reflected in actual payouts.

Key Finding

High-risk zones (AE, VE) do produce higher average severity, but Zone X (low-risk) generates the most total claim volume due to sheer policy count — the "volume paradox." Zone designation captures the broad severity signal, but physical hazard features (water depth, elevation) are essential complements. A claim in Zone X during Hurricane Harvey can exceed one in Zone AE during a minor rain event.

EDA 05

Do Elevation & Construction Era Actually Reduce Losses?

Elevated vs. non-elevated buildings, Pre-FIRM vs. Post-FIRM, and the freeboard threshold effect.

Key Finding

Building-level interventions reduce losses, but more nuanced than expected. Crossing above BFE cuts median claims by 63% (from $29K to $11K) — a threshold effect, not linear. Post-FIRM buildings show higher average claims than Pre-FIRM ($41K vs $30K) due to higher property values (TIV effect), not greater vulnerability. 81% of claims (1.6M) are non-elevated buildings, meaning the higher-severity group is also the most common type in the portfolio.

EDA 06

Where Are Flood Losses Geographically Concentrated?

State-level total paid, average severity, and a frequency vs. severity scatter for the top 20 states.

Key Finding

Losses are dramatically concentrated: Louisiana ($15.9B), Florida ($13.1B), and Texas ($12.5B) account for 63.3% of all NFIP building claim payments. LA leads on volume (~360K claims at $44K avg), FL leads on per-claim severity ($44,784). Mississippi is a hidden risk — moderate volume (~50K claims) but 3rd-highest average severity at $43,343, driven by Katrina concentration. NY/NJ appear in top-15 severity due entirely to Hurricane Sandy (2012).

EDA 07

How Are All Features Related to Each Other?

Correlation matrix across 20 key features and their individual correlations with the target.

Key Finding

Individual feature-to-target correlations are modest (max ~0.25), which is normal for insurance loss prediction. water_depth_log (0.248) and is_hurricane_season (0.239) are the strongest linear signals. Critically, freeboard_positive shows only −0.078 linear correlation, yet SHAP analysis reveals elevation as a top XGBoost feature — because its protective effect is conditional on flood zone and depth. This is the core insight: linear correlation cannot detect conditional interactions, which is precisely why XGBoost outperforms Linear Regression.

EDA 08

What Does a High-Severity vs. Low-Severity Claim Look Like?

Profile comparison: top 10% vs. bottom 30% claims across 13 features.

Key Finding

A high-severity NFIP claim has a consistent, physically sensible profile: 8.55 inches of water depth (vs 2.21 for low-severity — 3.9×), 39.2% in high-risk zones (vs 11.2% — 3.5×), 94.5% in coastal states (vs 71.9%), and $317K average TIV (vs $113K — 2.8×). Buildings average 39.9 years old vs 27.4 years. The profiles are coherent and grounded in flood science — the model is learning real patterns, not statistical noise.

Three Progressively Powerful Models

We trained three models in increasing order of complexity, using each as a benchmark for the next. All evaluated on a held-out test set of ~380K claims.

What Drives Flood Claim Severity?

Two complementary interpretation methods: XGBoost's built-in gain importance (global, which features are used most) and SHAP values (per-prediction, which features pushed each claim up or down).

SHAP Beeswarm — Per-Prediction Explanations

SHAP decomposes every individual prediction into the contribution of each feature. Unlike global importance, SHAP shows both direction (does this push the prediction up or down?) and magnitude for every claim in a 5,000-claim sample.

🔑 Key insight from SHAP: freeboard_positive shows only −0.078 linear correlation with the target, but SHAP reveals it as a top contributor — because its effect is conditional. When water depth is high AND the building is not elevated AND the zone is AE/VE, freeboard contributions are large and negative (reduces predicted loss when above BFE). Linear correlation averages this out to near-zero; SHAP exposes it.

Figure: SHAP beeswarm plot on 5,000-claim sample — Red = high feature value, Blue = low feature value, X-axis = impact on predicted claim ($)

Model Diagnostics

Actual vs. predicted scatter and residuals plot on a 5,000-claim holdout sample. A well-calibrated model should scatter around the 45° diagonal with residuals centered near zero.

Figure: Actual vs. Predicted (left) and Residuals Plot (right) — XGBoost v1 on 5,000-claim sample

The model under-predicts the largest catastrophic claims ($150K+) — a known limitation of tree-based models on heavy-tailed distributions. This informs the deployment recommendation: use the model's top 15% predicted segment (not just top 10%) as the triage priority queue to include a buffer for underestimated tail events.

Three Direct Applications for FEMA / NFIP

The XGBoost model predicts expected claim payout with R² = 0.669 and MAE = $16,474. Each recommendation below is a direct, immediate application of this predictive capability.

From 1.99M Claims to Actionable Predictions

Starting from raw FEMA data with 73 columns and 1.99 million records, this project delivered a production-ready flood claim severity model through a rigorous end-to-end pipeline.