The insurance industry is experiencing a fundamental shift in how to define, understand, and quantify risk. Recent technological advancements have led to an explosion of data, demanding new processing and analysis techniques. Consequently, insurers face a new challenge: finding a balance between developing highly accurate models and complying with business and regulatory requirements.
Unconstrained models — those with few limitations — maximize data utility and predictive power by leveraging advanced algorithms. When used strategically, unconstrained models can analyze and enhance traditional models to unlock new insights, even in highly constrained or regulated environments like insurance. For organizations that embrace them, unconstrained models present an opportunity to improve risk management and gain a competitive advantage.
The insurance analytics landscape sees explosive data growth
Data created, captured, copied, and consumed is expected to surpass 180 zettabytes in 2025, as shown in Exhibit 1. For context, just one zettabyte (1 trillion gigabytes) is roughly equivalent to 1 million copies of the entire Netflix catalog. Most of this data is unstructured, meaning it does not have a predefined format and requires the application of techniques like natural language processing (NLP) and large language models (LLMs) to extract value.
The insurance industry has transformed by leveraging this new data. Datasets now include telematics for scoring driving behaviors, Internet of Things (IoT) data for real-time leak detection, satellite imagery, and climate data for a more refined underwriting framework. This new data era presents incredible opportunities to redefine how risk is understood.
The limitations of traditional data analysis models and algorithms
The industry standard for data analysis, modeling, and product development centers on the generalized linear model (GLM) and its variants, which insurance firms have used since the late 1980s. While GLMs are highly useful analytical tools, their development predates modern datasets, which now contain millions of rows and hundreds of columns. Modern GLMs handle larger data but still require additional algorithms to find the most predictive inputs.
The effectiveness of traditional models like GLMs will always be limited by inherent constraints, conditions that restrict the scope of a model’s inputs and outputs to explain a given phenomenon or process. By design, constraints hinder what we seek to understand; however, due to their natural intuitive limits or boundaries, they will always exist as we strive to understand real-world metrics. Having zero constraints is impossible; thus, "unconstrained" is a misnomer and is more accurately described as less constrained.
Model constraints can be grouped into several broad categories. The following critical constraints are commonly encountered in most modeling projects:
- Data. Examples include volume (columns/rows), granularity, quality, and availability.
- Resources. Examples include time, expertise, compute, and storage.
- Regulation and law. Examples include unfair discrimination, redlining, and prescribed calculations.
- Privacy. Examples include the California Consumer Privacy Act (CCPA), Health Insurance Portability and Accountability Act (HIPAA), General Data Protection Regulation (GDPR), and Personal Information Protection Law (PIPL).
- Business. Examples include financial statements, investor reports, and International Financial Reporting Standards (IFRS).
- Values and culture. Examples include brand foundational philosophy and fairness.
From constrained to unconstrained models — unlocking predictive power
Unconstrained models place no limits on data, models, or algorithms. They are designed to maximize predictive power and accuracy by identifying the most relevant data and selecting the optimal analytical framework. It’s important to view these models as strategic tools that complement other models, helping to gain insights, unlock value, and enhance data utility. By employing advanced algorithms, unconstrained models can effectively process large and diverse datasets.
Unconstrained models provide quick insights through gap analysis, where constrained models are compared to a less constrained one that sets a performance ceiling. By examining the resulting performance gap, often visualized in a chart, we can identify the loss of predictive power in the constrained analytical framework. Understanding and minimizing this gap helps to maximize data insights and enhance the value of the model.
Exhibit 2 is a lift chart that illustrates the predictive power of two models: The dark blue model line has minimal constraints, while the light blue model line is constrained. The dark blue line is steeper from left to right than the light blue line, indicating the dark blue model is more powerful.
Alternatively, models are compared by examining the change in the area “under the (lift) curve.” Exhibit 2 illustrates the area difference in the grey region, representing the loss of predictive power. From a business perspective, this may represent loss of segmentation, fairness, profit, and value.
How project constraints can shape model flexibility and predictive power
Statistical modeling is both science and art, blending technical know-how with creative problem-solving. Unconstrained modeling requires an even higher level of skill to ensure the right methods achieve the desired outcomes.
A successful modeling project requires three essential ingredients:
- Data. A large volume of diverse data, combining relevant internal data with potentially relevant external data.
- Talent and collaboration. A multidisciplinary team with appropriate technical skills and subject-matter expertise, including data scientists, data engineers, software engineers, actuaries, product managers, and underwriters.
- Technology and tools. Modern tools, including artificial intelligence, act as force multipliers, making teams more nimble, efficient, and adaptive.
Exhibit 3 shows example projects within a grid where the horizontal axis represents constraints and the vertical axis represents model flexibility. Insurance pricing falls in the bottom left (high constraints, low flexibility), whereas advertisement serving is in the top right (low constraints, high flexibility). Each project’s position on the grid is based on its expected constraints and flexibility and serves as a useful guide when starting a project.
Model governance similarly ensures models are used for their intended purposes, properly maintained, and subject to appropriate change controls. Effective governance requires considering these core principles: fairness, transparency, accountability, security, and safety. While laws and regulations are still emerging, it is important for insurers not to rely solely on regulatory guidelines, but to proactively develop and promote strong self-governance mechanisms that address and mitigate regulatory and business risks.
The power of unconstrained models for insurance innovation
The explosion of data, combined with the limitations of traditional models and the rise of sophisticated algorithms, presents an unparalleled opportunity for insurers. Strategically applying unconstrained models can serve as a powerful catalyst for innovation.
By comparing unconstrained and constrained models using methods like gap analysis, insurers can quantify the performance loss attributed to certain limitations. This insight is essential; it reveals the true cost of constraints and highlights areas where regulatory or business constraints may be unnecessarily hindering segmentation or value creation. Maximizing predictive power does not mean abandoning traditional methods — instead, it means strategically enhancing them with advanced techniques and a robust framework for governance and interpretability.
