When Close Enough Isn’t Good Enough

Dec 10, 2025

insurance industry ai hero image

When “Close Enough” Isn’t Good Enough: A New Prediction Method That Could Transform Insurance Underwriting

Underwriting has always been an exercise in prediction. Every time an underwriter evaluates a risk, they’re essentially trying to tell the future: Will this policyholder file a claim? How much will it cost? What’s the likelihood of loss?

For decades, the insurance industry has relied on predictive models that focus on minimizing average error. These models work reasonably well, but they share a fundamental limitation: being “close on average” doesn’t necessarily mean your predictions align with reality.

A breakthrough from researchers at Lehigh University might change that. Their new prediction method, called the Maximum Agreement Linear Predictor (MALP), takes a different approach by prioritizing how well predictions match actual outcomes rather than simply reducing mistakes. For an industry built on accurately assessing risk, this distinction could be transformative.

The Problem with “Good Enough” Predictions

Traditional predictive models in insurance, including the widely used least-squares method, aim to minimize the average difference between predicted and actual values. This approach has served actuaries and underwriters well for generations. But there’s a catch: a model can have low average error while still producing predictions that don’t closely match real-world results.

 

Think of it this way: if you’re predicting claim costs, a model might be wrong by $1,000 on half the claims and wrong by $1,000 in the opposite direction on the other half. The average error looks great at zero, but none of your individual predictions were actually accurate.

According to Capgemini’s World Property and Casualty Insurance Report 2024, 83% of insurance executives believe predictive models are very critical for underwriting’s future. Yet only 27% of insurers currently possess the advanced technology needed to leverage predictive analytics effectively in their underwriting models. The gap between aspiration and capability remains substantial.

How MALP Changes the Game

MALP works differently. Instead of minimizing error, it maximizes what statisticians call the Concordance Correlation Coefficient (CCC). This metric evaluates how closely data points align with a perfect 45-degree line on a scatter plot, reflecting both precision (how tightly points cluster) and accuracy (how close they are to the ideal line).

“Sometimes, we don’t just want our predictions to be close, we want them to have the highest agreement with the real values,” explains Taeho Kim, the Lehigh University statistician who led the research. “If the scatter plot shows strong alignment with this 45-degree line, then we could say there is a good level of agreement.”

The researchers tested MALP using both simulated data and real-world measurements. In studies involving medical imaging data and body composition assessments, MALP consistently delivered predictions that aligned more closely with actual values than traditional methods, even when those traditional methods had slightly lower average errors.

What This Means for Insurance Underwriting

The implications for insurance underwriting are significant. Consider these scenarios:

Premium Pricing Accuracy: When setting premiums, insurers need predictions that closely match actual loss costs, not just predictions that are “close enough on average.” MALP’s focus on agreement could help carriers price policies more accurately at the individual risk level. According to Tricontech research, companies implementing advanced predictive modeling have reduced premium leakage by approximately $14 million per billion dollars of written premium.

Claims Cost Forecasting: Underwriters regularly forecast expected claim costs to determine appropriate reserves and pricing. A prediction method that prioritizes alignment with actual outcomes could reduce the variance between projected and realized losses. This matters enormously when you consider that insurance fraud alone costs the industry billions annually, and accurate predictions help identify anomalies earlier.

Risk Classification: MALP could improve how insurers segment risks into rating tiers. When predictions more closely match reality, the distinction between low-risk and high-risk insureds becomes sharper, leading to better risk segmentation and fairer pricing. Guidewire notes that by utilizing vast data sets, insurers can achieve pricing accuracy that has been unattainable with traditional actuarial methods.

Catastrophe Modeling: For catastrophe-exposed lines like property insurance, predicting losses from low-frequency, high-severity events is notoriously difficult. A method that prioritizes strong alignment between predictions and actual values could enhance cat model accuracy, helping insurers better prepare for extreme events.

The Tradeoff: When to Use Which Approach

The research reveals an important nuance: MALP isn’t always the best choice. Traditional methods like least-squares still outperform MALP when the primary goal is minimizing overall error across a large portfolio. The researchers observed that when reducing average error matters most, established methods remain effective.

This creates a strategic decision point for insurance executives: What are you actually optimizing for?

For portfolio-level decisions, where aggregate accuracy matters more than individual precision, traditional methods may suffice. But for decisions requiring close alignment with individual outcomes, such as:

  • Pricing individual policies in competitive markets
  • Evaluating specific high-value risks
  • Making underwriting decisions on complex commercial accounts
  • Assessing medical stop-loss or individual disability claims

MALP offers a promising alternative. The distinction matters because different lines of business and different underwriting decisions have different tolerance for variance. A homeowners policy with a premium of $1,200 versus a predicted $1,150 might not matter much at the portfolio level. But a $5 million directors and officers liability policy priced at $250,000 when it should be $350,000 could materially impact profitability.

Practical Implementation Challenges

Despite its promise, implementing MALP in insurance operations would face several hurdles:

Data Quality: Like all predictive models, MALP requires clean, comprehensive data. According to McKinsey research, only 15-20% of insurers believe they have the necessary data to support their underwriting decisions effectively. Incomplete or outdated claims data undermines any prediction method.

Integration with Existing Systems: Most insurers have legacy systems built around traditional actuarial methods. Introducing a new prediction framework requires careful testing and validation before deployment. Industry research indicates that real-time monitoring and continuous assessment are becoming the new standard, but legacy infrastructure often struggles to support these capabilities.

Regulatory Considerations: Insurance regulators scrutinize rating algorithms to ensure they’re actuarially sound and don’t create unfair discrimination. Any new prediction method would need to meet these regulatory standards, which vary by state and line of business.

Underwriter Adoption: Perhaps the biggest challenge isn’t technical but human. Underwriters may be skeptical of new predictive models, particularly those that produce different results than familiar methods. Capgemini research emphasizes that carriers need to meet underwriters where they are by showing them the value of new models and helping them communicate modeled decisions.

The Broader Context: AI and Prediction in Insurance

MALP emerges at a moment when the insurance industry is grappling with how to effectively leverage artificial intelligence and machine learning. Nationwide notes that while insurers have used data to forecast risks for decades, they can now leverage forecasting more meaningfully by analyzing large datasets and harnessing advanced computing technology backed by machine learning and AI.

The fundamental challenge remains constant: prediction accuracy directly impacts profitability, competitiveness, and the ability to serve policyholders fairly. MALP represents a refinement in how we think about prediction accuracy by explicitly prioritizing agreement between predicted and actual values.

This matters because insurance operates on thin margins. A seemingly small improvement in prediction accuracy can translate to millions in improved underwriting results. When predictions align more closely with reality, insurers can:

  • Price policies more competitively while maintaining profitability
  • Reduce adverse selection by more accurately identifying risk
  • Allocate capital more efficiently based on reliable loss forecasts
  • Build more stable, predictable book performance

Looking Ahead: From Linear to General Predictors

The researchers acknowledge that MALP currently operates within the class of linear predictors, which is “large enough to be practically used in various fields, but still restricted mathematically speaking.” They’re working to extend the approach to remove the linear constraint, creating what they call the Maximum Agreement Predictor.

This evolution could further enhance the method’s applicability to insurance, where relationships between risk factors and outcomes are rarely strictly linear. Telematics data, IoT sensor readings, and behavioral patterns often exhibit complex, non-linear relationships that challenge traditional modeling approaches.

Key Takeaways for Insurance Executives

As insurers evaluate their underwriting analytics strategies, several principles emerge from the MALP research:

Match the Tool to the Task: Not all prediction problems require the same approach. Portfolio-level aggregate accuracy and individual policy-level precision demand different optimization strategies.

Agreement Matters: For high-stakes individual underwriting decisions, how closely predictions match actual outcomes may matter more than average error across the portfolio.

Prepare for Evolution: As prediction methods advance, insurers need flexible analytics infrastructure that can incorporate new approaches without wholesale system replacement.

Invest in Data Quality: No prediction method, no matter how sophisticated, can overcome poor data quality. The foundation of accurate prediction remains comprehensive, clean, timely data.

Focus on Implementation: The gap between theoretical model performance and practical business impact depends heavily on change management, underwriter training, and systematic deployment.

Conclusion

MALP won’t replace traditional actuarial methods overnight, nor should it. But it represents an important evolution in how we think about prediction in insurance. By explicitly optimizing for alignment between predictions and actual outcomes rather than simply minimizing average error, it offers a tool better suited to certain underwriting decisions.

For an industry where the difference between profit and loss often comes down to prediction accuracy, that matters. As the researchers note, improved prediction tools could benefit many scientific areas, but few industries depend as heavily on accurate forecasting as insurance. Every underwriting decision is a bet on the future. MALP offers a way to make those bets more precisely aligned with reality.

The question for insurance executives isn’t whether to adopt MALP immediately, but rather to understand the principle it represents: different prediction problems may require different optimization strategies. As carriers continue investing heavily in predictive analytics and AI, recognizing when “close on average” isn’t good enough could provide a meaningful competitive edge.

Sources

Subscribe to AI Insights Newsletter

AI Disclaimer: This blog post was created with assistance from artificial intelligence technology. While the content is based on factual information from the source material, readers should verify all details, pricing, and features directly with the respective AI tool providers before making business decisions. AI-generated content may not reflect the most current information, and individual results may vary. Always conduct your own research and due diligence before relying on information contained on this site. 

Related Posts

468