As a researcher in Stanford’s Human Computer Interaction Group and the institute for Computational Mathematics at Stanford, William Steenbergen, CTO and Co-Founder of Federato has worked on state-of-the-art algorithms in reinforcement learning and dynamic optimization. In 2018, William was Benelux Elsevier Researcher of the Year and has productionized his algorithms at startups and large retail organizations. It was as a graduate student at Stanford that William met Co-Founder and CEO William Ross, and this ‘Tale of Two Williams’ led to the founding of Federato and the development of the insurance industry’s first RiskOps platform that embeds portfolio optimization into the core underwriting workflow. We caught up with William to ask a few questions about how reinforcement learning and Federato’s unique federated data architecture are not only solving difficult problems for commercial and personal lines insurers, but also helping to address massive societal challenges such as climate change.
William, can we start with a definition: in your own words, what is reinforcement learning, and how it is used in Federato’s RiskOps solution to help insurance carriers and MGAs balance their book of business and triage accounts based on appetite, guidelines, and winnability?
Reinforcement learning (RL) is a category of machine learning algorithms that predicts optimal actions using “trial and error.” Reinforcement learning has become a much more popular category in the last couple years. People know it, for example, from the underlying concept behind ChatGPT or the Youtube video recommendation algorithm.
In reinforcement learning, the goal is to train a model (usually referred to as the “agent”) that learns what the best action is to take given a certain environment (also referred to as the “state”). In contrast to other ML categories like supervised and unsupervised learning, RL can generate new data to learn from by interacting with the environment (this is where the “trial and error” comes in). Usually the most difficult part of RL is to define what the “best” in best action really means. The developer of the algorithm defines a good action versus a bad action through defining what is called a “reward function.” The reward function is a way of describing what “reward” the agent should get from taking a certain action.
As an example, let’s think of how this would be applied to what we do at Federato. You can think of the “agent” as the Federato RiskOps platform. The action it can take is to recommend to underwriters that they should look at certain opportunities. The environment would be the portfolio, or book of business, of an insurance company. In the RiskOps platform, executives and portfolio analysts can define what the “reward function” looks like. In other words, executives can define what a good portfolio looks like versus a bad portfolio, and the RiskOps Agent will optimize its actions to achieve the desired portfolio.
People who know you understand that you’re motivated by a genuine belief in the power of technology to make the world a better place. Federato was founded on the idea that helping insurance companies better manage losses related to climate change, like wildfires, hurricanes, and other natural catastrophes, can ultimately close the “protection gap” and make insurance more affordable and equitable for all. Can you talk about this North Star that guides you and the team at Federato?
Growing populations and rising property values, combined with an increase in high-severity catastrophes, have been pushing the insurance protection gap to a critical level. In other words, there have been more and more losses that are not insured. One of the core reasons is that, given the increasing number of catastrophes, it is more difficult to manage accumulations in the portfolio, increasing reinsurance costs and ultimately increasing premiums. Our vision is that through more data-driven portfolio management, we can lower reinsurance costs – and policyholder premiums – to help more people and businesses afford insurance.
In your role, you spend a lot of time with clients who are trying to stitch together complex environments and legacy systems that are often held together with bandaids and chewing gum. From a technical perspective, how does Federato’s data architecture – which lends the company its name – help insurers tackle the issue of legacy tech debt?
One of the hardest things about managing a portfolio at an insurance company is that all the data necessary to do any form of optimization lives in multiple places, and is hard to stitch together. Federato’s federated data architecture allows for ingesting all data into a single system, independent of where that information comes from. Our way of ingesting data allows for a great way of matching different records across a data system into a single source, even if any fuzzy matching of sorts is required.
“As every insurance IT executive should be aware, technology alone cannot solve every problem. Especially with legacy systems, it’s often also a problem of getting the right people together in a room and defining a solid implementation strategy. I think this is why it is so important for technology providers to truly see themselves as partners instead of vendors.”
– William Steenbergen, CTO & Co-Founder, Federato
If you want to make an implementation a success, it’s important for both parties to truly understand the data and methodology. Well organized implementation docs, state of the art technology with a high degree of flexibility, and experience in doing this over and over again for clients certainly helps with that, of course.
What should insurers keep in mind when planning to implement a data-driven solution like Federato? How can they prepare?
The best way of preparing is to understand who are the right people to get involved, and get them together right from the start. Every database or system has its expert. If you get these experts to communicate effectively, and you get one or two of the end-users in the same room, you’re bound for success. Among the biggest implementation failures that I have seen is when too many people get involved in the decision making process, or when the experts that actually understand the different databases do not get involved early enough.
In working with insurance clients, you are also called upon to weigh in as trusted advisor on broad technical infrastructure issues and investments in new core systems – things that might normally fall outside the scope of Federato. Based on your experience, where are insurers today on their digital transformation journey, and what do you see as the biggest hurdles to overcome?
First, I would challenge the statement that this is outside of the scope of Federato. To my earlier point, I strongly believe vendors should act more like partners.
“As a technology partner, I believe that it is very much in our interest to help and advise on broad technical issues and investments. At the end of the day, an insurer will have to cope with multiple technologies, and these technologies have to talk to each other. If we can help our clients make the right choices that will enable those systems to talk to one another, that is in everyone’s best interest.”
– William Steenbergen, CTO & Co-Founder, Federato
In terms of the biggest hurdles, it is often about the rigidity and flexibility of the software and team. I don’t believe in trying to pick vendors through a long laundry list of features. Every vendor will claim they can do anything, and at the end of the day, features always need to be specifically configured for the individual insurance company to make the implementation successful.
“Don’t ask whether a vendor has a specific feature available, ask them how the feature is configurable to work for the specific customer and need, and to prove it, ask them to demonstrate it configured within a very short time frame. This will give you a real sense of the flexibility of the solution.”
– William Steenbergen, CTO & Co-Founder, Federato
Federato works with a lot of ‘tech-first’ MGAs and MGUs whose business model is based on a proprietary algorithm or data model that allows them to outselect their competitors. But it’s one thing to have a unique model, and another to operationalize it and actually make it a part of the fabric of underwriting and operations. As someone who has commercialized your own algorithms, can you comment on best practices around this?
One of the main things that will make this a lot easier is to have one access point to all data available, including any outputs of any data models or proprietary algorithms. If you create one entry layer that makes things accessible, it encourages any engineers that interact with that layer to explore more and understand the full set of data, versus only interacting with the thing they care about.
Of course, that makes another assumption, namely that internal engineers interact with the same endpoints as any external parties. For me, this is crucial to building a scalable and flexible platform. It encourages engineers to keep documentation up to date, and is a recipe for stability.
It’s about customer-focused thinking, and taking the time to really understand the client’s environment and the daily challenges users face at a deeper level. At Federato, one of our core values is “Businesses don’t have problems, people do.” We go out of our way to solve for the needs of the individuals who work for our insurance customers – front-line underwriters and staff, operations teams, and actuaries.
Shifting gears a bit, one thing you’re passionate about and deeply involved in as a co-founder and leader is the work culture at Federato, and creating an environment where people can learn and grow. Can you share a bit about the culture you and team are building and what makes it unique?
I try to focus strongly on setting explicit learning goals, and share them with the entire team. Not only does this create accountability, it also ensures that everyone is aware of what each other's learning goals are, such that we can assign tasks accordingly. It generates a strong sense of celebrating mistakes, and helping each other learn from them. By talking through learning goals, you also instill a level of trust in each other. Describing one’s weaknesses and providing the right incentives and support to improve those weaknesses is a vulnerable thing to do, which generally creates trust and enhances teamwork.
What’s one thing that people might be surprised to learn about you?
Picking up on the question about learning goals above, one thing people might not know about me is that I completed my entire undergrad without writing a single exam!
Favorite sport and team, and why?
Vitesse Arnhem (Soccer). Growing up in The Netherlands, I had a season pass, and still watch every single game.