Design, implement, and evaluate AI-driven underwriting risk scoring in general insurance.

Thesis Abstract

This study investigates the design, implementation, and evaluation of an AI-driven underwriting risk scoring system for general insurance, addressing the persistent gap between traditional underwriting heuristics and data-driven risk assessment in dynamic risk environments. The problem centers on limited transparent integration of multi-source data, model risk, and governance challenges that constrain underwriting profitability and consistency across portfolios. The aim is to develop a scalable, auditable risk scoring framework that improves predictive accuracy, supports decision-making, and satisfies regulatory and governance requirements. Specific objectives include (1) constructing a multi-source data architecture incorporating claims history, exposure data, geospatial risk factors, socio-economic indicators, and textual sources from underwriting notes; (2) designing an AI-driven scoring model that combines machine learning algorithms (gradient boosting, neural network ensembles) with traditional actuarial methods to produce calibrated risk scores and accompanying uncertainty estimates; (3) implementing governance mechanisms for model validation, interpretability (SHAP-based explanations), and ongoing monitoring; (4) evaluating performance against baseline underwriting scores using internal data from a general insurance portfolio; and (5) conducting a pilot integration study with underwriters to assess usability, decision impact, and operational feasibility. The methodological approach adopts a mixed-methods design. A longitudinal, retrospective cohort study will use 5 years of policy and claims data from a mid-sized general insurer, comprising approximately 120,000 policies across motor, property, and liability lines. A stratified random sample of 20,000 policies will be used for model development and validation, with a holdout period of six months for temporal validation. Quantitative analysis will apply a stacked ensemble framework combining gradient boosting machines (XGBoost), deep neural networks for unstructured text features, and logistic regression for calibration. Model performance will be evaluated via AUC-ROC, Brier score, calibration plots, and lift metrics, with statistical comparisons against existing underwriting scoring methods using DeLong's test for correlated ROC curves and net reclassification improvement (NRI). Feature importance will be probed using SHAP values to ensure interpretability for underwriters. A Bayesian approach will be employed to quantify predictive uncertainty, enabling threshold-based decision rules and risk-informed pricing adjustments. Qualitative insights will be gathered through semi-structured interviews with 12 senior underwriters to assess usability, workflow integration, and perceived trust in AI outputs, analyzed thematically to identify operational barriers and enablers. Data collection instruments include (i) a data integration pipeline aggregating structured policy data, claims history, external risk indicators, and unstructured underwriting notes; (ii) a model evaluation dashboard presenting performance metrics, calibration status, and SHAP-based explanations; (iii) a structured interview guide anchored in technology acceptance and decision-support theories; and (iv) an ethical and governance checklist aligned with model risk management standards. Validity and reliability will be ensured through rigorous data preprocessing, cross-validation, back-testing with pre-specified time windows, and external expert validation of model outputs. Ethical considerations address data privacy, consent where applicable, and bias mitigation across protected attributes. Expected findings include (1) improved discriminative power of the AI-driven score over baseline methods, evidenced by a 3–5 percentage point increase in AUC-ROC and a 10–15% improvement in net profit on a simulated portfolio, (2) well-calibrated probability estimates with reliable uncertainty quantification, (3) robust SHAP-based explanations enabling underwriters to validate risk drivers, and (4) positive underwriter reception regarding decision-support and workflow integration, with identified interface enhancements. The study contributes to knowledge by advancing the integration of explainable AI into underwriting practice, refining model governance and calibration procedures in insurance contexts, and providing a replicable framework for multi-source data fusion in risk scoring. The main conclusion anticipates that an AI-driven underwriting risk scoring system, when coupled with transparent explanations and strong governance, can enhance pricing precision, consistency, and underwriting productivity without compromising ethical standards. Practical recommendations include adopting a staged deployment with continuous monitoring, developing domain-specific calibration protocols, investing in underwriter training on AI interpretation, and strengthening data governance to sustain model relevance amidst evolving risk landscapes.