In the complex world of health insurance underwriting, one enduring challenge is the significant asymmetry of information between insurers and applicants. While insurers must rely on often incomplete and traditional sources of data to price risks and approve policies, applicants frequently possess deeper knowledge of their health status and potential risk factors. This imbalance can lead to adverse selection and inefficient pricing strategies that undermine both the market stability and fairness of insurance systems. However, recent advancements in data science and the proliferation of novel data sources present an opportunity to bridge this gap. A pioneering study published in Risk Sciences explores how integrating alternative data—colloquially known as “big data”—and advanced statistical selection techniques can revolutionize health risk prediction in insurance underwriting, enhancing both accuracy and operational efficiency.
The research team, hailing from Peking University and the University of International Business and Economics in China, undertook a large-scale analysis using proprietary datasets provided by an innovative Chinese InsurTech firm specializing in critical illness insurance. This dataset is distinctive in its multidimensional nature, encompassing not only conventional policyholder demographics and underwriting variables but also granular smartphone-derived data. These smartphone data included device usage patterns, location history, application interactions, and credit inquiry signals, capturing a digital footprint of the insured beyond conventional medical histories. Complementing this, the researchers also incorporated publicly accessible hospital medical-claim records, creating a comprehensive picture of the policyholders’ health-related behavioral and clinical indicators.
Employing modern predictor-selection methods holds the key to effectively extracting actionable insights from these voluminous and heterogeneous data. The authors utilized LASSO (Least Absolute Shrinkage and Selection Operator) and its extended variant, Adaptive Group LASSO, which are penalized regression techniques adept at variable selection in contexts with high-dimensional predictors. These methods impose constraints that shrink irrelevant or less-informative variables’ coefficients toward zero, enabling the model to focus on the most salient features for predictive accuracy while maintaining generalizability on new data. By doing so, the models avoided overfitting and improved out-of-sample prediction performance, a crucial imperative in underwriting where decisions must reliably forecast future claims risks.
One of the study’s pivotal findings concerns the incremental predictive value added by smartphone data. Beyond traditional medical records and demographic factors, behavioral data captured from personal digital devices provided novel indicators correlated with critical illness claim outcomes. This digital information encapsulates subtle health and lifestyle proxies, such as movement patterns, social connectivity, and credit behavior, all of which, collectively, enhance risk profiling in ways classical data sources cannot. This signals a paradigm shift where insurers could leverage everyday digital behaviors as a complementary lens into health risk assessment, potentially streamlining underwriting workflows and reducing reliance on costly medical examinations or historical claims data.
Aware that collecting and processing extensive data across multiple domains presents operational and cost challenges, the study inquiry extended into identifying which categories of data are most valuable—thus, most worth collecting. Using Adaptive Group LASSO, which not only selects individual variables but groups of related features, the researchers pinpointed personal digital devices, recent travel experiences indicated by location data, and applicant credit records as the primary contributors to predictive accuracy. This prioritization offers a pragmatic roadmap for insurers to optimize data collection frameworks by focusing on the most impactful information clusters, balancing prediction gains with practical implementation constraints.
However, the authors emphasize the predictive—not causal—nature of their analysis. While the models can enhance risk stratification and claim prediction, the relationships discovered should not be interpreted as causal effects or direct health determinants. For example, a correlation between travel patterns and risk does not imply that traveling causes illness but rather that these patterns serve as informative proxies for underlying risk factors or behaviors. This nuanced distinction is critical in responsible model deployment, ensuring that insurance decisions remain reflective of statistically robust prediction rather than unsupported causal assumptions.
Beyond methodological rigor, the study openly discusses its contextual limitations. The data stem from a specific Chinese InsurTech environment, characterized by a unique combination of regulatory, cultural, and technological factors that may not translate identically to other markets or insurance lines. Additionally, the dataset’s reliance on applicant-authorized smartphone data presupposes willingness to share sensitive personal information, which raises considerations around privacy, consent, and regulatory compliance critical in real-world adoption. These factors suggest that while the study establishes promising proof-of-concept insights, broader validation and contextual adaptation are necessary before wholesale industry integration.
This research stands at the confluence of data science, behavioral economics, and actuarial science, demonstrating how interdisciplinary approaches can yield transformative improvements in traditionally opaque sectors like health insurance underwriting. The successful integration of digital behavior metrics, combined with advanced, group-aware statistical selection methods, foreshadows a future where insurers harness real-time, multidimensional data streams to finely calibrate risk and personalize policy terms. Such advancements potentially benefit both insurers, through enhanced risk prediction and operational efficiency, and insureds, via fairer pricing and faster underwriting decisions.
Importantly, the study also reflects broader economic and societal trends underpinning the evolving landscape of risk assessment. The rise of ubiquitous smartphones as health-information hubs creates unprecedented opportunities but also demands rigorous frameworks for data security and ethical use. The demonstrated predictive power of credit and digital behaviors elucidates complex interrelations between financial conduct, lifestyle, and health risks, nudging interdisciplinary research forward to decode these nuanced links further. This holistic view may spur novel insurance products and targeted wellness interventions, blending quantitative analytics with human-centered design.
The engagement of major foundations such as the National Natural Science Foundation of China and the National Social Science Foundation of China underscores the strategic priority placed on advancing predictive analytics in insurance economics. By fostering innovation in leveraging big data under stringent methodological standards, such initiatives contribute to building resilient financial systems that can adapt to evolving health risks and demographic trends. This aligns with global challenges, including aging populations and increasing prevalence of chronic diseases, that pressure insurers and policymakers to innovate trustworthy, scalable risk assessment tools.
In summary, the study presented by the Peking University-led team affirms the potential of integrating big data from unconventional digital sources into health insurance underwriting models. Using sophisticated variable selection approaches like Adaptive Group LASSO not only boosts predictive capability but also guides strategic prioritization in data collection. While challenges remain regarding data privacy, causality interpretation, and contextual generalizability, this work marks a substantive step toward more intelligent, data-enriched insurance risk prediction frameworks adaptable to the digital era. Future research expanding on these foundations could accelerate the transformation of underwriting from a largely heuristic process into a scientifically grounded discipline complemented by rich digital insights.
As health insurance sections increasingly intertwine with technological infrastructures, studies like this highlight the crucial intersection of data science, economics, and ethical practice. The integration of digital footprints into risk modeling opens vast horizons, but it also calls for vigilant balancing of innovation with respect for individual rights and transparency. By advancing predictive, transparent methods without overstepping causal claims, this research sets a proactive standard for harnessing big data benefits responsibly, fostering trust and effectiveness in the complex terrain of insurance risk management.
Subject of Research: Not applicable
Article Title: Data-enriched prediction of insurance risk
Web References: http://dx.doi.org/10.1016/j.risk.2025.100028
Image Credits: Shaoran Li, et al
Keywords: Economics, Health care, Insurance, Mathematics, Algorithms
