DreamQuark

The insurance sector is undergoing a transformation, but its main asset remains the amounts of data accumulated for years from policyholders. However, the way data-analysis is done has changed tremendously over the past years. ‘Business intelligence’ was replaced by data-science and artificial intelligence, and its subfield of machine learning. These are the key elements of the data-science revolution.

A lot of actors have indeed started to offer data-science related services. DreamQuark is one of these actors. DreamQuark is a start-up born from the experience of its founder in analyzing Big-Data in particle physics. The initial technology behind DreamQuark consists of algorithms derived from neural-networks, a class of algorithms belonging to the very popular deep-learning family, which have been built several years before it become such a buzzword. These technologies are now constantly being refined.

Deep-learning is a true revolution, even though its fundamental principles are more than 30 years old. The development of new hardware, the build-up of huge datasets, and the increase in the number of experiments over the past few years have given a much better understanding of the underlying principles of this technology, and have helped reach outstanding performances in computer vision, language recognition or data-analysis. These performances can reach, and sometimes even surpass, human capabilities. These algorithms are still far from reasoning like humans, usually referred to as strong intelligence, but they are able to achieve great performances on single specific tasks, making them a wonderful tool to automate a lot of business tasks.

This revolution comes with some challenges. First, it is still complex to implement these algorithms at a larger scale, because their design and implementation requires a lot of experience and allocated resources. Then, these algorithms remain intricate mathematical structures, delivering hard to read results: it is the ‘Black box’ effect. Finally, implementing these technologies is very time-consuming: building a pipeline, assembling the various building blocks, the open-source software to treat Big-Data and the algorithms that are used to create value. Altogether, these factors reduce the capability of firms to fully take advantage of this technology, specifically in highly regulated sectors such as financial services or healthcare.

One issue remains: while insurance companies have gathered some Big-Data, yet probably less than Google or Facebook, it is generally not exploitable by algorithms without a large pre-processing phase. These firms’ Big-Data is successively made up of several smaller databases, or is unbalanced, or is an ‘unlabelled database’.

Let’s illustrate this last example in building fraud detection models for life insurance: some databases only identify few fraudsters over a large history, giving little information on which to train algorithms to recognise new fraudsters. With most machine learning techniques, insurance companies would have to test a large number of potential fraudsters to enrich the database before being able to achieve results in new fraud identification.

These various challenges have caused insurance companies to take a step back from the ‘AI hype’  after a few costly and time-consuming proof-of-concepts.

But the opportunities for improvement with adequate technology remain intact.

DreamQuark answers all these challenges with a single, intuitive, user-friendly platform bringing novel solutions to business users.

The main feature of this software ‘Brain’ is the ability to turn the complex mathematical computations into a set of features closer to human reasoning that can be understood by a market expert.

It only takes a few steps to use the powerful technology on the user’s data. A user identifies the information they want to use to build a model, and the software automatically provides a set of transformations to clean the data before it is fed to the algorithms. The platform provides three different approaches to model training, depending on whether the user has labelled data, and the type of results to achieve: prediction or segmentation.

The algorithms learn from the data, and show the results and the performances of the models through a series of visualisations. The platform automatically trains several models and allows the user to compare and choose among these models, before applying a selected one to new data.

DreamQuark technologies apply to structured data, time-series, text, voice and images and the product has found applications in fraud reduction, AML, churn analytics, asset management and process automation with impressive results.

The main challenge to AI applications now remains related to the access and quality of databases: information, when it exists, is not often well defined, is highly biased, or is not easy to extract. This is one of the last remaining challenges on which DreamQuark is pushing its research, as it is the last limit to the scalability of data-science in insurance…