FarmRisk IQ: Agricultural Insurance Analytics & In-Browser Litigation Risk Prediction

Introduction & Project Motivation

The primary business objective of this build is to demonstrate a production-ready solution that aligns with the advanced analytics frameworks championed by industry leaders like ValueMomentum. Specifically, this pipeline mirrors the architecture required to achieve a 15% reduction in legal spend by routing high-risk claims to specialized adjusters before costs escalate.

Tech Stack & Architecture

Core Systems Integration (Guidewire & Duck Creek)

By compiling the LightGBM model into the Open Neural Network Exchange (ONNX) format, this pipeline completely bypasses the need for a microservice API. The ONNX file acts as a portable binary. In a Guidewire integration, the model can be deployed directly into the UI layer or a localized edge node. This means when an adjuster inputs FNOL data, the prediction executes client-side with zero latency, eliminating network failure points and server scaling costs.

Methodology & Data Pipeline

Because agricultural insurance data is highly proprietary, I engineered a synthetic dataset calibrated precisely to the USDA Risk Management Agency (RMA) 2022 Annual Report benchmarks. The pipeline generated 15,000 individual claims across 8 lines of business, anchored to a baseline earned premium of $13.8B and a 2022 loss ratio of 97.7%.

Feature Engineering & Domain Logic

The litigation risk model relies on biologically and operationally grounded features. For example, livestock claims involve injury variables that drastically alter litigation probabilities compared to standard crop yield failures.

# Synthetic litigation label generation using domain-informed logistic constraints
logit = (-2.8
    + 1.6*(claims['claim_amount']>80_000).astype(float)
    + 1.2*(claims['days_to_report']>25).astype(float)
    + 1.0*(claims['prior_claims']>=2).astype(float)
    + 1.4*claims['injury_flag'].astype(float)  # High biological/legal risk
    + 0.7*(claims['coverage_type']=='Livestock Risk Protection').astype(float)
    + 0.6*claims['peril'].isin(['Disease','Flood']).astype(float)
    + 0.5*(claims['policy_tenure_yrs']<=2).astype(float)
    + 0.4*claims['multi_peril'].astype(float)
    + np.random.normal(0, 0.4, N))

prob = 1 / (1 + np.exp(-logit))
claims['litigation'] = (np.random.random(N) < prob).astype(int)

Model Training & Export

I utilized LightGBM due to its superior handling of tabular insurance data. The model was trained with 300 trees and a maximum depth of 6, utilizing class-weight balancing to handle the ~4:1 ratio of standard to litigated claims. The model achieved a 5-fold CV AUC of 0.700 and a Test AUC of 0.722.

# Exporting the LightGBM model to a portable ONNX binary
from onnxmltools.convert import convert_lightgbm
from onnxmltools.convert.common.data_types import FloatTensorType

initial_type = [('float_input', FloatTensorType([None, X_tr.shape[1]]))]
onnx_model   = convert_lightgbm(model, initial_types=initial_type, target_opset=12)

with open('litigation_risk.onnx', 'wb') as f:
    f.write(onnx_model.SerializeToString())

Dashboard Components & Interactive Analysis

1. Claims Intelligence & FNOL Litigation

This section monitors the influx of claims and deploys the AI model.

• FNOL Heatmap: Tracks submission volume by month and day of the week, allowing claims managers to optimize staffing during peak harvest or weather events.
• Live ONNX Scorer: Users can manually adjust input parameters (e.g., Claim Amount, Days to Report). The WASM-backed ONNX engine instantly recalibrates the litigation probability, outputting a discrete routing decision (LOW, MEDIUM, HIGH, CRITICAL) for adjusters.

1. Claims Intelligence

2. Underwriting Precision

This segment evaluates portfolio profitability across geographic and coverage variables.

• Risk Score vs. Loss Ratio Scatter: Visualizes individual segment performance. Bubble size dictates policy count, while the color (green/amber/red) instantly identifies unprofitable cohorts (Loss Ratio > 100%).
• Submission Funnel: Tracks the conversion pipeline from initial submission to in-force policies, identifying bottlenecks in the quoting process.

2. Underwriting Precision

3. Actuarial Analytics

Agricultural insurance develops faster than long-tail liability lines. This page provides actuarial visualization for reserving.

• Loss Development Triangle: Utilizes the chain-ladder method. Users can track how claims from a specific accident year develop over a 6-year lag. For instance, my calibrated data shows 72% of ultimate losses are paid in year 1.
• Reserve Adequacy: A gauge tracking whether current IBNR (Incurred But Not Reported) reserves meet the industry benchmark of ~94%.

3. Actuarial Analytics

4. Portfolio Stress Test

This tool handles capital adequacy planning against catastrophic events.

• Scenario Sliders: Users manipulate parameters for Drought Frequency, Claims Inflation, and Fraud Rate Deltas.
• Real-Time Simulation: Adjusting the sliders instantly recalculates the Stressed Combined Ratio and Solvency Margin. If the simulated drought mimics 2012 severity, the interface actively displays whether current capital remains adequate or enters a stressed state.

4. Portfolio Stress Test

Results & Conclusion

Future Directions

To scale this prototype into an enterprise-grade application, the following enhancements are planned:

1. MLOps Feedback Loop: Implementation of automated score distribution monitoring. The system will track week-over-week score drift and trigger a LightGBM retraining pipeline when the AUC drops below a 0.68 threshold on newly labeled data.
2. Generative AI for Adjudication: Integrating a local LLM to parse unstructured adjuster notes and automatically extract the injury_flag and multi_peril boolean values before feeding them into the ONNX scorer.
3. FHIR Integration: For livestock claims involving veterinary interventions, parsing electronic health records via SMART on FHIR to dynamically assess disease-based peril risk.