Predict. Perform. Prove.

Predict. Perform. Prove.

Satyen Paneri

Guidewire’s approach is to work with customers to deliver analytics in a meaningful way (i.e., at exactly the right moment when they can be valuable to an underwriter or a claims adjuster and do that at a large scale). Guidewire facilitates millions of daily transactions across claims, policies, and billing. We seek to improve the efficiency of each transaction so that we can provide significant value for policyholders and claimants, for the industry, and for the world.

Guidewire Analytics transforms traditional ‘systems of record’ into ‘systems of insight’ where decisions about pricing, risk selection, claims triage, and settlement is all aided by useful data and analytics. The Guidewire Predict application serves as a key enabler for the ‘system of insight’ by taking internal and external data and converting it into intelligent predictions.

blog-image-20221101-prediction

Innovations in Predict with Flaine (pronounced like Flen) make predictions even faster with modeling improvements.

Ask any data scientist what the top 3 things they need to make their jobs easier and more impactful. Their answers will certainly be:

  1. To build the best predictive models using their favorite language and technique.

  2. To be able to explain to all stakeholders how the model will perform.

  3. To prove that the model delivers the desired business outcome.

In short, they want to Predict, Perform, and Prove.

Predict.

Predict supports a BYOM (Bring Your Own Model) approach where predictive models built using third-party tools and languages can be imported by converting them to PMML (Predictive Modeling Markup Language).

blog-image-20221101-predict-r

R and Python are ubiquitous and the most widely used open-source languages for predictive modeling. With Flaine, BYOM can import R and Python models into Predict without having to convert them to PMML. In addition, Guidewire has partnered with Anaconda to use the libraries and packages including R and Python, which it manages on our behalf. This lends more flexibility for data scientists to create models in their preferred packages and not having to worry about operationalizing open-source tools by offloading that review, documentation, and management to Anaconda.

Flaine also adds support for building Generalized Additive Models (GAMs)-based algorithms natively in Predict without requiring the models to be converted to a BYOM-supported format. GAMs provide superior outcomes compared to Generalized Linear Models (GLMs), while still maintaining the ease of understanding and explainability that data scientists and actuaries love with GLMs.

Perform.

Data scientists do not want their models to be black boxes. Explaining how a model works and ensuring its quality is a crucial element for the business and regulators. SHAP is a popular explanation technique; the Shapley value approach attempts to explain why a machine learning (ML) model reports the outputs that it does on an input. It helps understand the contribution of an individual predictor to the overall score.

Flaine adds support with SHAP values for Deep Neural Networks. This helps data scientists explain the inner workings of the model to the business team.

blog-image-20221101-shap-explain-2
blog-image-20221101-shap-explain-1

Prove.

Once a model is put into production, data scientists need ways to prove that it is working and having the desired business impact. Predict offers A/B testing to evaluate model performance by comparing competing models. Two models, A & B, are deployed at once and the percentage of calls that go to Model A and the percentage that goes to Model B are assigned. Once the winner is determined, all traffic is directed to that model and the other model is swapped out. A/B testing also allows for piloting new models by assigning a treated group (scored with the model) and an untreated group (business as usual).

blog-image-20221101-model-impact

In addition, visualizations built with Guidewire Explore, included with the cloud subscription, show the situation before and after model deployment. A/B Testing and Explore visualizations help the data science team prove the business benefits and make quick adjustments where necessary.

Through Predict, Perform, and Prove, Flaine delivers intelligent predictions that improve underwriting and claims performance.

Stay TUNE-ed for even better predictions with Garmish! Why the capital letters you ask, you will need to wait to find out.

Tags