Six Use Cases for Earnix Model Accelerator

What Every Carrier Can Do Right Now to Build and Deploy Models Faster

Let’s say for a moment that you’re the head of pricing and underwriting analytics at a forward-looking insurance carrier. One that’s committed to leveraging the latest in AI and machine learning (ML) to personalize customer experiences, grow revenue, and increase profitability.

Your team, over the course of a number of years, has built some very high-performing models, but as we all know, the market doesn’t stand still, and you’re under constant pressure to further improve models, smooth their transition to production, and enhance the team’s performance.

As you ponder how to best accelerate the team’s capabilities, you’re growing increasingly concerned about some shifts you see taking place:

• Since the company has grown through acquisition, you now have a number of modeling environments in house. And the models built in those different environments aren’t portable – you can’t leverage the investments from one toolset to another.

• Some of your smartest analysts have built some incredibly sophisticated models – so sophisticated, in fact, that they can’t actually be placed into production.

• The transformation of features in development and testing has to be replicated again in production, costing valuable time and effort.

• Pressure is mounting from regulatory and compliance, senior management, and customer service to improve the explainability of models and the decisions they render, causing the team additional overhead they feel could be much better spent on “real work.”

• Some of the models that have been trained and are highly productive, giving superior fit for historical data, are not necessarily “usable” for pricing purposes, as they can introduce a wide variety of undesired pricing anomalies.

• The team is spending way too much time working through multiple features and feature engineering settings when they try to create a GLM model.

Getting Further Acquainted with Earnix Model Accelerator

In a recent blog post, we introduced Earnix Model Accelerator, an extension to the Earnix platform that’s designed to streamline the process of building and incorporating advanced models, particularly ML models, in pricing and underwriting. It addresses the challenges outlined above, ones that are faced by many of today’s insurance analytics teams.

Model Accelerator gives those teams newfound freedom to choose whatever tools they prefer for modeling, incorporating virtually any model from other frameworks and formats, including transformations.

When combined with Earnix Price-It^TM and Earnix Underwrite-It^TM, teams can leverage their past investments in models, including those currently in simulation and production, adding to their agility and flexibility, and reducing the time required to react to market changes, including competitive actions.

In this blog post, we’ll look at some specific use cases for Earnix Model Accelerator, each addressing the critical challenges outlined above.

“Build It and They Will Come”

Insurance carriers often find that the models they’re using cannot be ported for use in pricing, because of system limitations imposed by incompatible environments.

This forces them to build a new model for pricing, with the attendant overhead of weeks of preparing data, building out an additional model, and separate testing and approval cycles, all of which may add weeks or months to the process.

Two models are required, costing the team time and resources with every cycle.

With two sets of models in development and deployment in parallel, governance and compliance issues are also likely to occur, as discrepancies between filings and production creep in.

With Model Accelerator, those system limitations are eliminated. The carrier can import its existing decisioning model in minutes, across environments, and start using it for pricing – without recoding the model.

In addition to models generated using Earnix AGLM, a wide variety of modeling environments and tools is supported seamlessly:

Accelerators streamline integration with the most popular analytic frameworks and formats, such as ONNX, H2O, Akur8, DataRobot, Databricks, Python, and others.

Modeling frameworks and techniques supported include sci-kit, TensorFlow, Pytorch, Random Forest, XGBoost, GBM, CatBoost, LightGBM, and Neural Networks.

This saves time in every modeling and deployment cycle going forward, and allows the model to be maintained in a single location, vastly reducing the chances for errors that would cause regulatory and compliance issues.

“Nice Work, But You Say It Can’t Be Implemented?!?”

Sometimes, data scientists are set loose on a critical business problem and come up with an incredibly sophisticated ML model that shows significant promise.

It not only solves the business problem, but it shows promise in beating the competition, who are all struggling with the same issue, meaning the team has found a new source of competitive advantage.

Unfortunately, the model is so sophisticated that can’t actually be put into production – it can’t be coded into the rating engine that’s in use. The ground-breaking work of the team goes to waste, and there’s risk that the competition will actually solve the problem first and gain the upper hand.

With Earnix Model Accelerator and its simple user interface, the team can import even the most sophisticated models and move them into production immediately.

The model will no longer be doomed to sit on the shelf. The team contributes mightily to increasing revenue, reducing expense, or delivering additional profitability – whatever that thorny business problem was that they attacked.

“Transform Once and Re-Use”

A key task for data scientists is to pre-process data to fit into their models. Today, that process has to happen twice – once in the development of the model, and again in production. Especially if the transformations are complex, they will need to be replicated in the rating engine.

That might take a week to code the transformations and compare them with the original calculations. This duplication of effort adds unnecessary time, effort, and money to the process, and is a repetitive task that no one looks forward to.

As with any duplicate effort, the implemented model might also contain errors once it’s put into production, causing unexpected behavior in the market, and risking customer experience (CX) issues, brand erosion, and unwanted regulatory scrutiny.

With Model Accelerator, the model and the pre-processing code are easily imported into the Earnix environment, without having to recreate the transformations.

Feature engineering is simple to accomplish in both the lab and in production, with the model working exactly as designed in both environments. In addition to the desired decisioning accuracy, errors are less likely to occur between the model development, testing, and deployment phases.

“Could You Explain That, Please?”

We live in a world in which just about everyone has an opinion about artificial intelligence (AI) and its effects, both intended and unintended. Consumers, internal teams, and external watchdogs are all asking more, and more pointed, questions about models and their behavior.

Transparency and explainability, especially for “black box” models, has become a critical ingredient in adoption from a users’ trust and understanding perspective, and to satisfy regulatory requirements. 

When asked to explain how their models work, data scientists might resort to using terms such as “Shapley values” and “permutation importance,” technically accurate terminology but not of much use to the public, regulators, or even other internal staff.

What’s needed is the ability for business users to understand the structure of ML models, so they can be confident and comfortable in using those models. They also need to be able to answer questions about the model that might be posed by internal or external auditors, legislators, and regulators.

Model Accelerator has the unique capability to generate a highly-understandable Explainability Report, which shows which features matter most for the predictions being made, and how the features impact those predictions.

Everyone will be able to understand the model and its business implications, and to defend them if the need ever arises.

“Before We Put That into Production…”

Analytics teams would like the flexibility to use models developed for a particular purpose to do “double duty” and be applied to other problems.

For example (in jurisdictions in which it’s legal to use them), they might want to use a demand model that works well, and then import it into Earnix Price-It, potentially leveraging that model for pricing purposes.

Today, the two models must be developed, implemented, and maintained separately, because it’s difficult to tweak a given model for both purposes and run variations on it until it operates as expected for both needs.

With Model Accelerator, the data scientist can verify that the model works correctly for the new purpose, or if it doesn’t, to easily identify why it’s not delivering the expected results.

Model Accelerator will enable her/him to run variations on the model and how to tweak it until it performs as expected for the new purpose. This shortens the time to delivering business results and makes the modeling effort significantly more efficient and cost-effective.

“Better, Faster, Smarter”

In a non-Earnix world, model building is a more time-intensive process than it needs to be, especially when there are large numbers of features to analyze and massage.

The process relies too heavily on trial and error, especially to ensure that the model’s behavior aligns with business needs and passes regulatory muster. Further, it can be difficult to review and compare models against each other, then modify the proposed model before putting it into production.

The Earnix AGLM tool enables the use of automated GLM capabilities and hybrid modeling techniques.

For example, a pricing analyst can dump all the available features for the model under consideration into the AGLM Model Builder, then set constraints on the model or features in a way that aligns with the business and regulatory environment.

The AGLM Builder automatically optimizes the model under the constraints the user has provided and returns the most accurate model. The user can then be highly confident in the model, and that KPIs are properly calculated and compared against the original results.

The result is that Earnix users can automatically create and compare multiple GLM models with various features and hyperparameters and select those that best fit their needs. This provides a new level of flexibility, ease of use, and accuracy.

Leverage the Tools Your Team Already Knows

Every insurer will have use cases, or variations on the theme, that match those we’ve discussed.

Earnix’s open analytics platform empowers its customers to choose the best analytical approach to fit their needs to determine pricing, rating, underwriting, and personalization.

This gives customers the flexibility to build models within Earnix, build models in an external tool, or employ a mix of both.

Stunning Results

Carriers around the world are putting Earnix-driven machine learning models to use on a daily basis and reporting stunning results. Model Accelerator is just one tool in the Earnix arsenal to help drive these business improvements.

For example, a leading provider of consumer warranty coverage for large appliances has been able to extend their personalized extended warranty offers, with these quantifiable results:

• Throughput – the ability to price up to 40X more warranty plans with this solution

• Model Output was tripled

• Margins on policies doubled

• All while maintaining high conversion rates

Another carrier, part of a large US multi-line insurer, has built out a single solution for all policy quote development and deployment for its small business insurance, using open-source ML models favored by the analytics team – something that many closed/propriety environments simply couldn’t accommodate.

The results:

• It took just 3 months to implement Earnix and move from GLM to advanced ML

• The team can now test and deploy new pricing rules in less than one hour

• The pricing is deployed by business teams, without relying on IT, but within a firm governance framework

• The team reports that this solution has fostered an environment of continuous improvement in agility, governance, and speed

In a third case, in an extremely competitive European market, a carrier that wanted to press the advantages it has gained through the adoption of the Earnix environment and ML modeling:

• Required just 3.5 months to implement Earnix and integrate with its legacy systems

• Resulting in a scalable, end-to-end analytical solution that keeps them ahead of the competition in their national market and keeps them compliant with all applicable regulations - and

• Delivered 29% growth over just the first three years of Earnix production

Summing Up

Earnix Model Accelerator extends the Earnix platform to deliver additional excellence in pricing and underwriting. It offers several key benefits that are unique in the industry, all aimed at moving analytics teams’ efforts faster and more flexibly and delivering real business benefit to insurers.

For More Information

We hope this blog post has helped make some Model Accelerator use cases more concrete.

To learn more, please download our full Use Case Document, which explains them in more detail and will allow you to get a jump on the competition.