DeltaQuote QuotePortal

Up all night
Thank you :)

Inspiration

The DeltaQuote team was exposed to the challenges in insurance underwriting while consulting for Hiscox Ltd. Our customer identified the limitations associated with manual ranking systems currently used by most insurance underwriters in prioritizing quote request queues. Our team realized that an automated artificial intelligence approach can be used to identify missed opportunities by comprehensively address all quote requests.

What it does

The QuotePortal is an automated ranking system for insurance underwriters to comprehensively manage broker quote requests. The QuotePortal is a system that provides a flexible data pipeline through which new information is dynamically integrated to allow underwriters to efficiently fulfill the most profitable requests. A machine learning algorithm uses quote request features such as retention % and gross earned loss ratio to recommend the most profitable insurance quotes. The system seamlessly interfaces with existing digital infrastructure to provide underwriters an easily navigable user interface.

How we built it

First, we listed out the variables and data germane to the problem at hand. This process involved specifying the demands of the customer and end user, planning the ideal qualities of an end product, and developing a vision for how the end product will be used. Then, we planned out what would be possible within 24 hours, and developed an API that a future product could use to link between a separate backend and frontend. The backend was built with Python and Keras, and uses gradient boosting to classify quote requests by the likelihood that an underwriter would process them.

The backend works on a cloud based service where we run our deep learning algorithm. The system uses a Neural Collaborative Filtering Algorithm, which focuses on implicit feedback to indirectly reflects users’ preference through behaviours like accepting, skipping or declining quote requests. Compared to explicit feedback (i.e., ranking each quote request manually), implicit feedback can be tracked automatically and is thus much easier to collect for content providers.

The historical data of quote requests from Hiscox was first cleaned to make it compatible with the model. New features were generated out of existing features based on our research on decision making of underwriters. The new dataset with the generated features is itself used to train the model.

A gradient boosting classifier was used to perform classification and provide a probability of the underwriter accepting the quote request. This gradient boosting classifier model generated feature importance maps and selected the features which the gradient boosting classifier think is important. High performance computation clusters are used to train and deploy our deep learning model. The model uses cross-validation to combine predictions of three different algorithms - Generalized Matrix Factorization, Multi Layer Perceptron and Neural Matrix Factorization. The models updates itself based on the activities of the underwriter, thus the recommendations and ranking of quote requests gets better as the underwriters use them. The front-end performs API calls to the cloud and dynamically updates using clicks from the underwriters.

Challenges we ran into

Our team faced various challenges when deciding on a ranking mechanism for the provided data. We spend ample amounts of time brainstorming different features to add to the existing data and deciding on heuristics for ranking quote requests in a way that would feel natural to an underwriter. Additionally, when analyzing our target market/demographics, we faced challenges regarding the extent to which our project should balance the demands of customers (i.e. managers within insurance companies) and end users (i.e. underwriters). In the end, we decided to satisfy both of these demands through a user interface that still lends control to the underwriter, while still prioritizing quote requests that are predicted to benefit the customer.

Accomplishments that we're proud of

“Developing an end-to-end, deployment-ready deep learning model in 24 hours.” - Srihari

“Same as Srihari” - Michael

“Learning about the inner workings of insurance underwriting and being able to work with it despite having a chemical engineering background” - Brook

“Developing and integrating a frontend for an application within 24 hours” - Naveen

What we learned

“The struggles that underwriters have to go through to in selecting quotes. I roleplayed an underwriter while developing a back-end solution for the final product, and the experience was eye-opening and gave me a lot of insight into the process.” - Srihari

“Same as Srihari” - Michael

“I learned that the range of things that can be insured is limitless” - Brook

“Balancing the demands of customers and end-users is a difficult yet interesting challenge to navigate during B2B product development” - Naveen