Actuaries and data science professionals can access more external data products than ever before, due to technology that lowers the barriers to accessing and productionalizing external data and AI platforms (e.g., DataRobot, Sparkbeyond, Dataiku), enabling rapid model development.  

That being said, it is important to continuously monitor the impacts of existing data products and have a strong understanding of how integrating challenger data products into machine-learning models may impact consumer outcomes. State regulators in several states, including Colorado, have taken notice of the need for greater monitoring and transparency.

Consumer Advocate Concerns

Studies conducted by the Consumer Federation of America have identified pricing disparities in auto insurance policies that used non-driving data to calculate premium payments. They found that the use of external data that included zip codes and credit score information, along with other factors that are frequently used as proxies for race, can perpetuate structural racism. 

Consumers have also been wary of insurance companies that claim to use machine learning and AI to improve their processes. One insurer found itself facing substantial online backlash after trying to explain its use of AI for claim decisions — critics warned of unavoidably discriminatory outcomes.

State of Colorado Takes Action

Colorado Senate Bill 21-169 is an example of regulators taking action to address these concerns. The state’s insurance commissioner will develop a framework for insurers to demonstrate that their use of algorithms, predictive models, and external data does not cause any unfair discrimination against protected classes. Under the bill, those classes include race, ethnicity, gender, gender identity and expression, and sexual orientation.

Colorado’s bill was introduced as an attempt to curb unintended discrimination in marketing, underwriting, and pricing not only in auto insurance, but across all business lines for insurers. It requires them to change practices when their use of external data results in unfair discrimination, and the definition of external data includes credit scores, social media habits, and geolocation data.

At a minimum, insurers will be required to provide documentation on how external data is sourced and used in predictive models, maintain a risk management framework to monitor unfair discrimination in the context of external data, and issue regular updates on the results of monitoring their risk management framework.

Insurers who develop new products must be prepared to explain not only the processes behind them, but also the measures they have taken to avoid unintended consequences. These explanations are easier when external data vendors provide detailed information regarding data quality, coverage, and applicable use.

The Silver Lining for Insurers

Rapid innovation is welcome, and it’s absolutely necessary in the context of external data for insurance use cases, but not all data products are equal. Rapid innovation can quickly lead to discriminatory impacts if the underlying data products are inaccurate, have poor coverage, or demonstrate instability over time. 

The onboarding process for Demyst requires external data vendors to provide detailed information regarding data quality, coverage, and applicable use, and we're encouraged that our insurance clients almost always request this information before performing their analysis. With increased access to external data, insurers can now access multiple data products to maximize coverage and verify accuracy across a single attribute (e.g., business sales revenue).

Demyst also works with clients to help them identify and eliminate biases in their predictive models — first identifying potential biases through a comparison of historical underwriting decisions with external data attributes on protected classes (e.g., race, gender, sex), and then either stripping those attributes from the training data or reweighting training samples. Finally, Demyst includes additional variables, either from external data or internal client data, to build an ethically compliant training dataset.

Ultimately, the spirit of the Colorado law is welcome to the extent that it (1) helps governments better understand how insurers are using external data products, (2) provides insurers a framework to innovate their processes while maintaining fairness, and (3) provides consumers with greater access to insurance coverage.