Valen Predictive Analytics gives Pinnacol Assurance underwriters new insight for enhanced tier placement and more effective pricing.

Valen Products & Services Used:

UnderRight^®:  Valen’s predictive analytics application for underwriting enhancement

Pinnacol Assurance Overview

For more than 90 years, Pinnacol Assurance has been Colorado’s assured and trusted source of workers’ compensation insurance for Colorado employers. Through its affiliated agents, Pinnacol provides comprehensive, competitively priced coverage; immediate attention to claims; a highly qualified network of SelectNet medical providers; and proactive safety programs to approximately 60,000 Colorado businesses. Over the past four years, Pinnacol has returned $227 million in general dividends to its policyholders statewide – a mark of the company’s stability and strong financial health.

The Pinnacol Business Opportunity

Pinnacol Assurance’s business model is unique in the industry because they are a political subdivision of the state of Colorado operating as a domestic mutual insurance company and also acts as the carrier of last resort in Colorado.  Pinnacol, like all insurance carriers is highly regulated and required to follow strict guidelines in order to provide pricing consistency and ensure that pricing discrimination does not exist.

While NCCI sets the approved base rates for many states, in Colorado insurers can deviate from these base rates by filing loss cost multipliers to establish company or tier specific pricing.  Insurers often create several wholly owned insurance companies in a state - each with different rating structures.  These different companies enable an insurer more pricing flexibility under their overall corporate umbrella.  However, Pinnacol’s enabling statute does not allow the carrier to create multiple separate companies to refine their rate structures.

The initial challenge Pinnacol faced was to price policies to avoid having low risk businesses subsidize high risk ones.  To achieve this, Pinnacol developed a tiered pricing structure they filed with the Colorado Division of Insurance (DOI).  Pinnacol placed policies into one of four tiers based on several factors including, premium size, three-year loss ratio, NCCI experience modification factor, election of an in-network designated provider and timeliness to report claims.  The factors were used in an additive approach with a few additional constraints to help segment the book of business by expected risk.  Each of the four tiers had a corresponding loss cost multiplier (LCM) filed with the Division of Insurance.

Pinnacol established their tiers by reviewing a limited number of data elements from historical policies and fitting a pricing model that matched their historical experience.  While this pricing model fit what happened in the past very well, Pinnacol felt a more robust method could predict its future loss experience and yield more accurate pricing of individual risks.  Pinnacol realized they needed the modeling capability to modify their current tier placement approach and eliminate systematic pricing redundancies and inadequacies at the individual policy level.

Pinnacol recognized that they had an opportunity to price more accurately however to do so required a better understanding of expected performance for each policy.  By using predictive analytics, Pinnacol believed they would be:

• Better able to price underperforming accounts
• More competitive in the small premium market
• Able to price lower risk accounts more attractively in order to retain existing accounts
• Able to identify new accounts for more aggressive pricing

Looking and Learning before Leaping

While many technology investments are a reaction to a business problem, Pinnacol’s investigation into, and ultimate investment in, predictive analytics was based upon the business opportunity that predictive analytics offered - the ability to use predictive analytics to better segment and thus better price to their exposures. Although Pinnacol believed predictive analytics would provide value, before deploying it they understood they needed more due diligence.

“We knew we had to better understand the technology and predictive analytics before we could intelligently plan our approach to implementation.  Valen was very open to sharing and transferring their knowledge.  Valen knew that we needed a deeper understanding of their solutions before making an investment in predictive analytics.”

Mark Isakson
Associate Vice President and Pricing Committee Chair
Pinnacol Assurance

Therefore, Pinnacol entered into an agreement with Valen designed to expand its knowledge of predictive analytics, create new predictive models, deploy the models and provide a feedback loop for continuous model and process improvement.  The business objectives were to ensure Pinnacol’s financial health by generating adequate premium on a risk-by-risk basis and to eliminate cross subsidies.

“The partnership with Valen proved to be a great complement of skills and knowledge.  We understand our business while Valen knows how to build and deploy predictive models for P&C insurance.”

Mark Isakson

Building the Pinnacol Model

The first step in building a predictive model for Pinnacol was to collect several years of detailed policy and claims data. Pinnacol supplied Valen with nine years of data elements from policy and claims records.  Pinnacol did not supply data to Valen from their most recently completed calendar year as data from that year was to be used in a blind test once Valen completed the predictive model.

As part of the data gathering process, Valen had the option of sourcing even more data from their External Data Warehouse that contains hundreds of firmagraphic (same as demographic but data is specific to a firm) data elements on about 80 percent of U.S. businesses and their ValenNetworks™ which contains policy and claims data on over $10 billion in premium and over 2 million workers’ compensation policies.  During this phase of the project, Pinnacol and Valen worked together to eliminate those data variables that were not appropriate for modeling.  Variable reduction at this early stage of the modeling process is performed to ensure that duplicative, highly collinear, confusing and socially unpalatable (such as race or gender) variables are not used in the modeling process. Once both Pinnacol and Valen agreed on a final list of variables appropriate for modeling, Valen created a statistical grade dataset that would be used by the Valen modelers to build the Pinnacol predictive model.

Valen first segmented the information by using data from 1999 through 2003 for model development and data from 2004 for model testing.  The next step in the model building process was to build a predictive model. After determining the baseline performance of Pinnacol’s current tier pricing structure, Valen worked collaboratively with Pinnacol to on-level premiums (bring past premium to current using current rating algorithms) and trend losses.  Once completed, Valen utilized their predictive analytics platform to winnow the number of variables from several hundred to the set of variables that would produce the most predictive model possible.  Valen then used 2004 data to test the model before sending it to Pinnacol for the final blind test.  Pinnacol then ran the model against 2005 policy data (without claims information as an underwriter would not actually see it at quote time) to simulate model performance in actual production.  Valen uses this blind test to compare what actually happened in 2005 to what the model predicted would happen.

In the blind test, the Valen model significantly outperformed Pinnacol’s previous tier structure by more finely segmenting Pinnacol’s portfolio of risks.

“We view the model developed with Valen to be a vast improvement over our previous approach and a performance differentiator in our market.”

Mark Isakson

Once the model was built, Pinnacol completed an internal, independent evaluation and validation of the model to make absolutely certain the model would perform in production.  Convinced of the model’s performance during this validation phase, Pinnacol began the implementation process.

The Implementation

When Valen developed the model, it was designed to integrate directly with Pinnacol’s existing internally developed information systems.  As Pinnacol’s IS team develops, maintains and supports all of its own information technology, it was critical that the in-house IS staff be able to understand and support the Valen predictive model.

While Pinnacol utilized internal IS resources to implement the model, it was essentially a pain free process.

“The investment in predictive analytics was worthless if we weren’t able to integrate the model into our existing applications and processes.  We worked with Valen from the initial design through final sign-off to ensure we could implement the end product.  The time was well spent on the front-end as we were able to seamlessly integrate the model on the back-end.”

Mark Isakson

In addition to the need for predictive analytics expertise, Pinnacol also acknowledged an insurer cannot just take an off-the-shelf statistical application.  They chose Valen in large part due to Valen’s combination of technology and expertise in developing a highly customized predictive model for Pinnacol’s workers’ compensation business.

“Ironically, the biggest implementation challenge was not the technology - Valen’s approach to building a customized model to our needs negated that concern.  We were able to focus our attention on change management strategies – on-boarding our underwriters and agent partners to understand how and why we enhanced the pricing model.”

Mark Isakson

Pricing model changes can be difficult for underwriters and agents.  Pinnacol is always thorough in these situations and works through all the various contingencies when they introduce significant change.  As a result, Pinnacol outlined a comprehensive strategy to implement the new pricing methodology.

First, they had to ensure that Colorado’s Division of Insurance approved the new pricing model.  Pinnacol, with Valen’s assistance, prepared the necessary rate filings and explanatory documentation so the DOI could clearly understand the new structure. Second, Pinnacol embarked on a process to educate their underwriters and agent partners on the forthcoming pricing changes.  It was important to Pinnacol that their underwriters and agents recognized they were not employing ‘black box’ pricing, but rather providing new insight into policy pricing decision making.  In communicating the value of the new approach to underwriters and agents, Pinnacol, focused their messaging on the business benefit rather than the sophistication of the model. Third, Pinnacol’s IS staff added functionality to their existing underwriting desktop that enabled underwriters to see key tier placement drivers for individual policies.  The system displays the suggested tier and provides the tier characteristics related to the underlying risk factors.

Pinnacol then trained their underwriters on the use of the new functionality in the underwriting desktop and the new pricing structure before deploying it.  The key drivers allowed Pinnacol to give underwriters the information necessary to understand and explain individual policy tier placement to agents and policyholders, while still protecting the propriety mechanics of the model. Lastly, the design of the model enables Pinnacol to make their own ongoing adjustments with respect to premium distribution and other factors.  This capability allows for continuous model refinement as Pinnacol discovers even more about individual risk factors and policy performance through the use of predictive analytics.

The Results

As a not-for-profit, Pinnacol’s ultimate goal for predictive analytics was not to generate increased profitability.  Rather, Pinnacol desired to refine their existing tiering approach to more effectively price individual exposures based on their expected future performance and to improve Pinnacol’s ability to compete in the marketplace through reducing pricing redundancies.  Pinnacol is well on its way to realizing both objectives.

Pinnacol’s underwriters are utilizing the new pricing structure to determine initial tier placement, which then allows the underwriter to take a closer look at account specifics to arrive at a final price.  The model complements the underwriters’ efforts, enabling them to focus on the underwriting aspects that rely most heavily on their experience. From the agent perspective, the tier placement process is and should continue to be transparent. As an added benefit, Pinnacol has a new ability to target sales efforts on the most profitable policies while adequately pricing less attractive risks.