Machine Learning Models: How can actuaries do more with AI?
Welcome to the hx Machine Learning Blog Series! 🎉 Every month, we'll share an insightful piece discussing Machine Learning (ML) technologies and their impact on businesses, specifically in the Insurance and Insurtech industries. Check out our first blog by Jonathan Bowden, hx Senior Model Developer, discussing the adoption of Machine Learning in the Insurance and Insurtech industries, some challenges actuaries may face, and potential solutions for the future.
People predict that with Machine Learning technologies on the uprise, AI will take all our jobs one day. Now this may be decades away, but Google Trends data shows that interest in Machine Learning began to increase in 2015 and is showing no signs of slowing down. (If you are new to Machine Learning, feel free to familiarise yourself here.) In finance, countless cross-market studies show that most companies are already using Machine Learning techniques, with only a few not even considering it.
However, when I ask around in my network of actuaries, I find that the interest in Machine Learning is limited. There are over 94,000 actuaries worldwide, according to LinkedIn, yet only 2,300 have the term "Machine Learning" on their profile. How has the hype been pervasive in society yet relatively nascent in the actuarial world, with only 2% of actuaries promoting their abilities in this area?
While not everyone can be a competent actuary or Machine Learning specialist, I believe good actuaries can become experts in both fields for two reasons. Firstly, one of the opening lectures in any Machine Learning course is linear regression, a topic which also features in the early actuarial exams. Second, becoming proficient in Machine Learning methodologies requires good coding experience in R or Python. This is another skill set that most actuaries have since most actuarial exams are set as R programming exercises. Even seasoned actuaries will probably have done a bit of Visual Basic for Applications (VBA) programming language, which allows for a relatively easy transition to Python, the most renowned Machine Learning language.
So, if all these similarities and entry points exist, what is causing the number of actuaries embracing Machine Learning to be so low?
Machine Learning in Insurance carries high risk
Good Machine Learning models have several known advantages (more accuracy, better data processing, trend recognition, etc.). Still, they are not always successful, have significant drawbacks, and can even be risky. We are all aware of our professional responsibilities, and with many media stories highlighting unethical uses of AI, we must exercise extreme caution. The last thing we want to do is incorporate bias and discrimination into our model by accident. There are several examples of AI going wrong, and given the media attention it receives, making a mistake in this field is not just terrible for those affected but it can also reflect negatively on our employers, our profession and our industry. As such, we should take extra care when we use Machine Learning methodologies and ensure that we understand what the model is doing and why. There have been a few best practice guides available in IFoA literature, though this is still an emerging field. Another solution to this is to ensure practitioners are making use of Explainable AI (“XAI”) tools, as these can offer meaningful insight into otherwise opaque models.
Actuaries need more training
The path to understanding Machine Learning can feel daunting. But there are countless resources that can take your knowledge from the basic introductory material to the joys of deep learning, reinforcement learning, explainable AI and more. The subject area goes far beyond Scikit-Learn and Caret, and there's an enormous amount of research referencing new and exciting methods that I believe actuaries could harness for innovation. Academic research can sometimes be hard to understand, but many digestible alternatives exist. Google's TensorFlow guidance documentation has grown immensely over the last couple of years, and it's now possible to build your own Recurrent Neural Network just by following a step-by-step guide.
Lack of demand for Machine Learning models
Could it be that we are not learning Machine Learning because there is just no demand for Machine Learning models? All the additional work required could hinder progress when confronted with a wall of complex data, and if employers lack an easy-to-implement technology stack, it's no surprise that there are so few models out there. It's easy to lose motivation if we don't have ideas for additional research, but don't throw that Machine Learning exercise in the bin just yet! Think harder; there is almost certainly something else you could try if you have a bit more time to invest in the exercise, and if you're stuck, there's probably some deeper training you could do. Talk about it with other experts. You may just stumble across a gold mine if you hunt for ideas because there are likely to be a range of competitive advantages from finding a model that is more accurate, more efficient, and makes better use of your data or delivers a better experience.
Slow pace of innovation in Insurance industry
On the other hand, it may not be down to a lack of demand or training. Innovation in the Insurance industry has always been a bit slower than in other areas, so maybe as a profession, we're just not that innovative. This is a theory I refuse to accept as many actuaries are pioneering new methods and technologies, and the trend is growing. The scope of Machine Learning is broad, and some of the best AI models out there appear to be far beyond most of our knowledge. But the actuary's responsibility is equally vast; we do more than calculating premiums.
Actuaries apply business principles, historical experiences, maths and statistics to various scenarios, including past, present, and future considerations. I don't believe AI will be able to take our jobs in their entirety, but I do think we are not bringing in enough support from our robotic colleagues. Machine Learning algorithms, with the correct level of supervision, may detect patterns and trends in complex datasets that a human would never hope to identify. For now, a more traditional, statistical model communicated well by an actuary certainly does seem to work, but for how long will this be the case?
With all of this in mind, I am in no doubt that we as actuaries must press ahead and continue our learning and innovation in emerging technologies like Machine Learning and AI. Actuaries are ideally placed to take advantage of the AI revolution. We have the skills and data, and most of us enjoy the process. There are competitive business advantages, and if anything, it's fascinating to watch what robots do in different situations, especially when they are so naive about the human world. Who knows, maybe one day, an AI will replace most of my job, and perhaps they'll write an article about encouraging greater use of humans.
Interested in finding out more about hx?
Pricing effectively within an ever-changing reinsurance landscape
Insurtech, Trends2 Mins
See the key drivers for digital transformations within the reinsurance sector and why pricing has established itself as a key initiative for successful reinsurers.
Is Python the right programming language for insurance?
In this blog, Dan Johnson, hx Head of Learning, explores various programming languages and the peculiarities of the insurance industry to highlight why Python is the right programming language for forward-thinking insurers.
Rapid reinsurance dealmaking is here with hyperexponential and Supercede partnership
Insurtech, Partners2 Mins
hyperexponential and Supercede have brought their SaaS platforms together to reimagine and reinvigorate reinsurance dealmaking, enabling reinsuers to create, validate, broker, and quote treaty submissions in 29 minutes instead of 12 weeks.