Profit-aware classifiers tailored to operational limits improve detection of high-risk cold-start auto-insurance customers and raise expected portfolio profit relative to a standard baseline (p < 0.001). A balanced bagging ensemble gives the best practical trade-off for regulated deployment, while a lightweight Transformer delivers stronger robustness and generalization under data perturbations.

Main Finding

In a no-history insurance setting with extreme class imbalance and heavy-tailed losses, two business-aware classification strategies — a balanced bagging ensemble optimized for expected profit under customer-omission constraints, and a lightweight Transformer that learns richer feature representations — both substantially outperform the prior baseline in profit (p < 0.001). The balanced ensemble delivers the best practical trade-off (predictive performance, robustness, interpretability, computational efficiency) for regulated deployment, while the Transformer shows stronger robustness and generalization under data perturbations.

Key Points

• Problem context: classifying new customers at an insurer with no prior company-specific history, under severe class imbalance, heavy-tailed claim costs, and strict operational selection limits.
• Business-aware design: both methods explicitly incorporate the asymmetric financial cost structure of insurance and operational selection (omission) constraints rather than optimizing only global accuracy.
• Proposed approaches:
  • Balanced bagging ensemble: handles class imbalance and is tuned to maximize expected profit while respecting omission constraints.
  • Lightweight Transformer: learns richer feature representations to improve generalization and robustness.
• Evaluation:
  • Primary dataset: proprietary auto-insurance portfolio of 51,618 customers.
  • Robustness checks: nine synthetic datasets with controlled perturbations.
  • Metrics: business-oriented profitability metrics (expected profit under selection limits) and statistical tests (ANOVA, Friedman, pair-wise comparisons).
• Results summary:
  • Both approaches outperform the baseline in profit (statistically significant, p < 0.001).
  • Balanced ensemble: better balance of performance, interpretability, and computational efficiency — well-suited for regulated environments.
  • Transformer: stronger robustness to data perturbations and better generalization in stressed scenarios.
• Practical recommendation: use the balanced ensemble when interpretability, runtime, and regulatory constraints dominate; consider the Transformer when expecting significant distributional shifts and needing stronger generalization.

Data & Methods

• Real data: proprietary autos portfolio, n = 51,618 customers (no prior company-specific behavior histories).
• Synthetic validation: nine datasets designed to test robustness and generalization under various perturbations (e.g., distributional shifts, label noise, class imbalance severity).
• Models:
  • Baseline: previously published methodology optimized for global predictive accuracy.
  • Balanced bagging ensemble: ensemble of classifiers trained on balanced resamples, objective aligned to expected profit under explicit omission/selection constraints.
  • Lightweight Transformer: compact attention-based architecture engineered to learn richer feature representations from available covariates.
• Cost integration: asymmetric financial cost structure of claims and operational limits were encoded into the training/selection process to prioritize business-relevant outcomes (high-risk identification).
• Evaluation framework:
  • Business metrics focused on expected profit under operational selection limits.
  • Statistical significance and ranking via ANOVA, Friedman tests, and pairwise comparisons to confirm consistent outperformance.
  • Computational and interpretability assessments for deployment suitability.

Implications for AI Economics

• Aligning objectives: explicitly optimizing ML models for profit and operational constraints (rather than global accuracy) materially improves economically relevant outcomes in insurance — demonstrating the value of objective design that reflects firm incentives.
• Decision-theoretic ML in markets: integrating asymmetric costs and selection limits provides a template for using ML to operationalize actuarial portfolio optimization and risk-based customer selection.
• Adoption trade-offs:
  • Simpler ensemble methods can be preferable in regulated markets because they balance performance, interpretability, and efficiency — lowering compliance and audit friction.
  • More expressive models (Transformers) can improve robustness and generalization, which matters for firms facing frequent data shifts or using richer feature sets.
• Market and policy considerations:
  • Better identification of high-risk clients affects adverse selection dynamics, pricing, and risk pooling; insurers should weigh effects on market segmentation and fairness.
  • Explicit profit-driven selection raises questions about transparency and regulatory scrutiny — methods that maintain interpretability will ease oversight.
• Research directions:
  • Broader evaluation on multiple real portfolios and longitudinal deployment studies to measure long-run effects on portfolio risk and market behavior.
  • Integration of fairness constraints, causal analyses, and distributional robustness methods to balance profitability, equity, and regulatory compliance.