1. Homepage
  2. -
  3. Insights
  4. -
  5. A Machine Learning Use Case...
A Machine Learning Use Case for Insurers
May 22, 2019 TIA , Article , machine learning

With great enthusiasm insurers in Denmark, Norway, Sweden, UK, Germany and many other countries move to data driven business models. But the majority prefers to ignore it. It seems to be too much of a challenge to achieve tangible results. Together with an insurer which operates on TIA we can demonstrate that it is not that difficult.

For a long time now, the challenge for financial institutions has not been about getting enough data. Currently the amount of data we produce every day is truly mind-boggling. There are 2.5 quintillion bytes of data created each day at our current pace, but that pace is only accelerating with the growth of the Internet of Things (IoT). Over the last two years alone 90 percent of the data in the world was generated.[1] The Harvard Business Review found that

  • less than half of structured data is actively used in decision making
  • less than 1% of unstructured data is analysed or used at all [2]

The challenge is rather acting on all that data in order to get actionable insights and to reach reasonable business results. Working with such quantities requires non-traditional approaches, hence the emergence of machine learning and deep learning (artificial intelligence). Microsoft’s Principal Solutions Architect, Scott Seely, defines 8 business use cases for the insurance industry:

  1. Lapse management, helping prevent losing existing customers,
  2. Recommendation engine, defining the right set of coverages and insurance sums for each customer,
  3. Assessor assistant, automating many manual processes performed during claim regulation,
  4. Property analysis, especially for mass events,
  5. Fraud detection, applying automated mechanisms of pro-active prevention,
  6. Personalised offers, improving the customer experience by bundling coverages from different lines of business and additional services,
  7. Experience studies, helping predict claims activity,
  8. Scaled training, training models using GPUs or thousands of CPU cores.[3]

The extent, to which each of these use cases is already used by insurers varies significantly.

Machine Learning, for all its clear advantages, is employed in Insurance much less often than it could. Most companies consider it a vastly complex technology, and therefore opt against any activities, which would employ such methods in any way – in fact, this reason is quoted by 53% of Insurance decision makers as a main reason for not employing ML or AI.[4] It is not surprising – reasonable utilization requires technological know-how, mathematical expertise, and industry knowledge, which should guide the entire process towards business-oriented goals. But the fact is that reasonable employment of machine learning is much easier than many decision makers believe. There are many Data Science Platforms on the market, some of them acting on an open source principle (e.g. h2o.ai).

Sollers is currently exploring and developing a machine learning solution for the insurance business. A lot of consumer data in core systems,  gives numerous possibilities to support the business with valuable insights. Conclusions based on machine learning analysis may, for example, increase effectiveness of insurers cross-sell campaigns. This can be achieved through answering questions such as:

  1. Which customers who have product A (e.g. credit card) are more willing to buy product B (e.g. travel insurance)?
  2. What is the best time of the year to offer a specific product to customers e.g. travel insurance?
  3. What is the best time to offer a product, e.g. travel insurance (one day before the trip or earlier)?

We chose an approach with h2o.ai, one of the leading Machine Learning platforms. It is widely adopted and operates on an open source basis. It also requires a relatively simple technological setup.

Of course, industry knowledge is required to make sense of the data, select the most important data points and guide machine learning solution towards analysis, which has a chance of providing useful and actionable insights. Creation of appropriate algorithms can include very complicated calculations, but even relatively simple mathematical actions can help achieve significant results.

The approach taken was to integrate the insurance core system TIA with the h2o.ai engine and to train the engine with customer data applying machine learning to the dataset. The input data needs to contain structured and relevant information, including e.g. personal data, policy data, claims data and transaction data. The scheme below presents what architecture can be applied to put h2o.ai into action.

A Machine Learning model organised in a way  described on the graphic and trained on a file containing 1.5 million data sets with 100 columns, was able to select the best model with 180 decision trees. As a result, the model was able to select 1% of consumers with a 10-times higher likelihood of purchasing a given product at a given time. For the top 20% of customers, this likelihood was nearly 3x higher.

The Preparation of such a model requires only a few weeks and a small team. This proves that no big investments and months-long projects are required to reap early benefits and evaluate usefulness of Machine Learning for a company. And the confirmation of the usefulness is of the utmost importance. Wolfgang Wahlster, founder and former CEO of German Research Center for Artificial Intelligence, admitted in an interview: “I have already seen three hype cycles for AI – it’s very important, as I tell our researchers, not to over-promise and under-perform.”[5].

It seems to be easy to get involved in trendy, flashy technologies and devote significant resources to projects which might be interesting but are unlikely to deliver any tangible results which would justify the investment. As defined by research of BCG, navigating an unpredictable field of technology such as Machine Learning and Artificial Intelligence, an adaptive strategy is required.[6] It is therefore advisable to initiate a series of small, low-cost Proof-of-Concepts, targeted at validating different business use cases. The ones that deliver the most promising results can be scaled up.

Piotr Pastuszka, Senior Manager at Sollers Consulting


[1] https://bit.ly/2Iz6Hbv

[2] http://bit.ly/2P7poDd

[3] http://bit.ly/2P5x48V

[4] http://bit.ly/2P4Adpz

[5] http://bit.ly/2P9ICYI

[6] https://on.bcg.com/2JajfWB