While the commercial property and casualty industry continues to search for signs of a hardening market, several countervailing economic factors are combining to slow any positive development. The deepening recession will continue to reduce premiums in lines based on payroll and sales revenues while generating increased losses in many lines sensitive to the economic downturn.

Carriers that want to emerge stronger from this tough environment must improve underwriting discipline and process efficiency in a cost-effective fashion.

Unfortunately, it is not easy in commercial insurance to reduce underwriting costs while maintaining underwriting control. This is partly because, unlike personal lines, most commercial insurance is still written manually and with minimum use of advanced technology.

There is also a general belief that the sophisticated underwriting judgment that goes into reviewing submissions and applying rating and other pricing tools is too complex to be aided by analytical tools and predictive models.

However, recent innovations in, and applications of, data analytics and predictive modeling in other industries as well as in personal lines and small-market commercial insurance provide us with solutions for dramatically improving underwriting performance in the middle-market commercial insurance sector.

Utilizing analytical tools with a logical structure designed by in-house underwriting experts, these solutions act as an aid to, not a substitute for, the existing underwriting process. The result is a more effective and efficient commercial underwriting program.

Underwriting in commercial insurance tends to include a broad subjective component. Most underwriting teams experience a significant level of variability in rating and pricing decisions, because of the diverse knowledge and experience levels of the staff as well as their disparate abilities in managing negotiations with agents and brokers.

As a result, there is considerable debate about the merits and risks of using analytical tools for standardizing and improving underwriting performance in middle-market commercial insurance. While the perceived benefits are significant, underwriting managers point to the various challenges and hurdles faced in this market, including:

o Lack of Adequate Data:

Because analytical modeling is widely understood to require a large number of records, there is concern that commercial classes may be too broad to provide enough data to develop stable models.

In addition, organizations may not always have the extensive loss and premium history for each line of business needed to develop predictive models of profitability on individual submissions.

o Prohibitive Cost and Disruptive Nature of the Project:

Analytical technologies, when implemented in personal lines and in other industries, have typically involved a significant organizational effort. This has entailed major cost outlays, process changes and infrastructure conversions.

o Highly Subjective Nature of Commercial Insurance:

Perhaps the key objection to using analytical tools is the view that commercial underwriting takes years of experience to perform properly and is too subjective for modeling. By this widely accepted view of commercial underwriting, expert judgment is the key to the decision-making process in delivering the best underwriting outcome.

That outlook is changing. The above concerns, all valid, can now be addressed by modern predictive modeling approaches that include a whole range of techniques–from pure data-driven approaches like cluster analysis and decision trees, to more complex models like multivariate regression and uplift modeling, to expert driven methodologies that are designed by replicating specialists' logic structures, such as belief-based expert systems.

Based on successes in similar applications, a particularly promising approach to improving middle-market commercial insurance underwriting is provided by Bayesian belief-based modeling. This approach relies less on existing data and more on expert judgment.

In effect, the approach models the decision-making thought process of expert underwriters within the organization relying on information available to the underwriter as part of the submission.

All model outcomes can be analyzed in detail, since the causal logical structures are explicit and transparent.

Initially designed based on exhaustive input from in-house underwriting experts (and any relevant internal and external data that may be available), over time the model “learns” from actual submission data and underwriting results. This approach is particularly powerful because it can be designed and tested with a relatively minor investment of time and cost.

Full-scale implementation, with the associated infrastructure investment and process changes, need occur only after the approach has proven itself and after the organization has developed a clear understanding of the benefits and costs of the requisite investment.

The accompanying graphic provides a high-level schematic of a comprehensive approach to improving underwriting performance using Bayesian predictive models. In addition to a technical and performance reporting facility, the solution comprises three key components: a Risk Appetite Gate, Bayesian Risk Quality Predictive Models, and a Bayesian Predictive Methodology for Package Submissions.

The Risk Appetite Gate is a tool managed by a lower-level analyst or technical resource and ensures that only submissions that fall under the organization's specific underwriting risk appetite definitions are evaluated by its underwriters–all others are either declined or returned to the producer for clarification or changes.

Risk appetite rules are absolute criteria that can be based on SIC codes, class, location, line of business, and so on. A standard Business Rules Engine is used for defining and managing the rules.

Existing risk exposure of the organization's book of business is an input into the tool. Consequently, the tool is dynamic, ensuring that underwriting does not review undesired submissions.

There is a Bayesian model for each product line, and all submissions are evaluated on a line-by-line basis. The following are key characteristics of these models:

o Models evaluate Risk Quality (not profitability) of each line of submission.

o Models are designed and developed with critical input from in-house underwriting experts, and “trained” with historical data if available, or with ongoing data for an initial period if not.

o After initial training, the models are “tested” with an additional set of data.

Once fine-tuned with these two groups of data, the models continuously “learn” in a “model learning loop” with live submissions data and underwriting results. Model performance and predictability improves over time.

A model's output is a score evaluating the risk quality probabilities for the respective line of business. Based on thresholds applied because of internal risk quality guidelines and resource constraints, these are divided into specific recommendations including: “Decline,” “Pass to Rating,” or “Refer to Underwriting for Further Review.”

All entries referred to underwriting include guideline indicators, identifying the key reasons for the referral based on the model's underlying analysis.

These models and the associated processes are designed and constructed based on the market segments, underwriting processes and internal data availability specific to each carrier.

After the models have identified a score for each individual line of a package submission, these scores are provided to a Bayesian model that assesses the overall score of the submission.

Similar to the Risk Quality predictive models, this model is designed with expert input to evaluate package submissions based on weights of the relative risk quality scores of each line as would be assessed by internal experts.

In addition, any outside rules–such as a temporary organizational desire to grow the workers' compensation book, for example–can be added to the model to ensure proper treatment of all kinds of package submissions.

Similar to monoline entries, any multiline submission referred to underwriting includes guideline indicators providing reasons why the package was referred and what the underwriter should be focusing on in their evaluation.

What are the advantages of applying these methods?

Internal underwriting profitability models, if available, can provide further enhancement of the process. These models, in combination with the risk quality scores, can help determine a weighted expectation of the profitability of each package submission, thereby helping prioritize submissions further for underwriting.

Carriers looking to dramatically improve underwriting quality by using this solution as a supplement to expert underwriting judgment will realize several advantages, including:

o Increased efficiency of valuable underwriting resources, as they focus only on the most complex and promising submissions.

o Ability to implement the solution even with limited initial data by reliance on models built to replicate expert underwriting judgment.

o Continuous improvement, as the models learn and grow more sophisticated over time.

o Relatively low up-front cost to test the approach before full implementation.

o Rapid acceptance by underwriting, since models are designed with underwriting input and because each model recommendation can be analyzed for root causes.

o Use as a documentation, quality improvement and training tool for inexperienced underwriters.

o Valuable technical/modeling and underwriting performance-related reporting facility for management analysis and decision-making.

Advanced analytical tools have proven successful in many applications where expert judgment is a key element of the decision process.

In today's challenging environment, commercial carriers focused on the middle-market can significantly reduce underwriting expense and maintain underwriting discipline by adopting these tools. Embracing these methods will translate into a sustained competitive advantage through all market conditions.

Aamer Mumtaz is director of business analytics at TNC Management Group based out of Chicago, Ill. He may be reached at amumtaz@tncmanagement.com.

Special Reports /

Predictive Models Can Help Cut Costs Without Sacrificing Underwriting Control

Related Stories

Recommended For You