Man vs. Machine: 10 years ago, I would never have guessed that I would be writing about this topic with such serious concern. Yet, some people are predicting that machine learning technology will produce a jobless future for certain professions, including actuaries. And with news headlines like “Google’s AlphaGo AI beats Lee Sedol again to win Go series 4-1” and “Meet Ross, the IBM Watson-Powered Lawyer”; you have to wonder what the future holds for the actuarial profession and just how much computers can take over the human role within actuarial departments. With the tremendous advancements in machine learning, many financial institutions are already making extensive use of these technologies to do all types of (traditional and new) actuarial work including competitor rating reconstruction and intelligent claims handling.
Let’s take a moment to see what people are thinking about when it comes to the impact of machine learning on the development of their actuarial teams:
Yes, give it time, machine learning will take over!
More and more problems can be addressed by machine learning technology. This technology is able to “learn” and “improve” quicker over time than humans can naturally adapt to new problems. We have already seen this in our daily work; for example, machine learning algorithms can process vast quantities of data produced by financial services institutions and produce accurate rates within minutes.
Unlike humans, predictive methods calculated by computers don’t age or drop in performance over time. With many executives constantly under pressure to cut costs, this could be a welcomed respite. By replacing humans with cheaper machines, this would not only reduce costs but also reduce the dependency on humans; improving reliability and the risk of business disruptions that go along with staff fluctuations in actuarial teams.
No way, machine learning is just a tool for humans!
As much as machine learning has been successful in taking certain tasks away from the human actuary, the current technological development level of machine learning is still at a stage where it requires human governance and intervention.
The level of automation and self-adjustment for most publicly available machine learning technologies is not yet ready for the “ungoverned” use in financial analytics. This is due to the fact that the brain can process far more parallel sets of information and make more highly complex judgments based on past experiences vs. simpler machine learning techniques. Even the more sophisticated techniques such as deep learning have so far only been applied to very clear and confined problems.
Today, working modes for machine learning involve solving partial problems in business (where rules & information are easier and can be coded by data tables). Humans can then plug these partial solutions together using middleware and organized IT interfaces to orchestrate and evaluate the quality of the overall solution using their contextual experiences. In this way, machine learning is simply another tool kit that extends the range of application within the organization that is carried out by the analytic team.
Wait, there’s another option: machine learning can take actuaries to the next level!
While machine learning is a useful tool for actuaries, it is hard to imagine that these computers will be able to think up worst case scenarios, creatively ask new questions or provide the judgement and innovation needed without human intervention. An example of such a question is: “What type of useful problem can I solve as a machine?” Machine learning is also not free of “errors” and cannot validate itself nor its solutions to be predictive in the future. However, there is good evidence that machine learning is predictive in situations that don’t change dramatically from the past and hence don’t excessively violate the assumptions of the methods used to fit the past data. Yet, the final responsibility still lies in the hands of humans as the ultimate decision maker and owner of the process.
Thus, actuaries will still be needed to intervene and interface with machine learning, but to a lesser and more efficient extent. The physical involvement in problems will be reduced because heavy calculations will be outsourced to the machines, and human intervention will be reduced to setting them up and validating the quality of results. The setup of the methods will be more automatic than the iteration mode we have seen with classical methods such as GLM, where the human has to intervene and be very creative during the iterative building of the final result.
This will free up actuaries time and other resources to tackle new phenomena that can be solved with machine learning. Machine learning will allow analytics departments to concentrate on solving previously unsolvable problems, such as market price reconstruction, as well as provide actuaries the opportunity to solve new problems or open up new opportunities for companies, such as clever Pay As You Drive concepts that encourages drivers to take safer trips with their vehicles.
This could end up in a fruitful cycle of “human” innovation using the more powerful and flexible toolsets of machine learning, at least mid-term. Such fields of application could fall into advanced claims handling, clever marketing and allocation of investments, market research and fraud detection.
So let’s not get too concerned just yet. I don’t believe in the foreseeable future that actuaries will be replaced by computers, nor the need for actuarial teams to shrink. However, if it does happen, actuaries will not be the only ones affected, it will affect all financial professions across the board. Make sure to take the time to stay up to date with machine learning techniques, because without it, you run the risk of becoming an actuarial dinosaur!
Recently Earnix conducted a survey to explore how Machine Learning is expected to impact the insurance industry over the next 3-5 years. Download the complete results of the survey.
Editor’s Note: This post was originally published in August 2016 and has been updated since then to include a link to the Earnix Machine Learning Survey.