Sarah MacDonnell and the Machine Learning in Reserving Working Party discuss how – and why – general insurance reserving actuaries should get on board with machine learning
There is a lot of interest in machine learning (ML) for general insurance reserving and it is a healthy research environment with a plethora of papers, but uptake in practice is slow, and we haven’t yet landed on one optimum solution for using it. This is no surprise, given that the options are endless. It is hard to know which approach is ‘better’; the literature so far has not done much in the way of comparing and evaluating different methods.
For example, if we look at the application of neural networks on individual claims, we see that there are large differences in the approaches used: from a more traditional paid claims developmental approach (using LSTM, or long short-term memory, a technique for sequential data) to Markov chains with five different model steps separating out stages of the claims process. And that’s before we even look at which neural network techniques have been applied, and how they have been applied.
The Machine Learning in Reserving Working Party believes it is worth starting to find that elusive solution now, rather than waiting for someone else to provide it.
WHY IS ML USEFUL?
Isabelle Williams: The world is changing rapidly. Many traditional reserving methods, including the chain-ladder method, assume that development patterns observed in the past will continue to be relevant in the future. More recently, it has been demonstrated that this may no longer be the case. ML models can give us insight into the changing world around us, highlight the likely effect of these changes, and allow us to adjust reserves to better prepare for the challenges inherent in the unknown.
Gráinne McGuire: It’s true that ML offers the potential to gain additional insights into the data, and this is a good reason for using it. But even without that (noting that reserving actuaries are generally good at their jobs and understand their data well), the automated side to ML could generate a more robust and efficient process.
HOW IMPORTANT IS THE DATA?
Isabelle: A model is only as good as the data you put into it. As any data scientist will tell you, 80% of their work is in the data. ML offers us many ways to supplement, clean, check, adjust and correct poor quality data without having to resort to manual cleaning. This saves time and allows actuaries to focus on adding value in other areas of the business. A good quality data process is therefore just as important as pursuing and acquiring good quality data in the first place.
WHY SHOULD ORGANISATIONS OR TEAMS INVEST IN ML SKILLS?
Sarah: From our UK survey (bit.ly/MLWP_findings), we know that reserving actuaries are keen to use ML in reserving, but find that time and resource constraints form barriers. ML is not something you can learn from a course – as with other actuarial skills, it takes time to build up the insight and judgment, and there’s no substitute for applying it in practice.
Jacky Poon: There is no better time to learn about ML, with the wealth of knowledge now available – including the Foundations articles that the working party has established (bit.ly/MLWP_Foundations). The skills are not just about ML – how often has a spreadsheet error impacted a business’s financial results? Investing in a proper replicable data analysis workflow in R or Python can reduce the risk of errors, even if you do not want to use ML.
“By embracing ML, reserving actuaries can provide predictive capabilities and expertise beyond what is currently possible”
John McCarthy: The longer you delay, the greater the chance that your team will be left behind. As older processes become more entrenched, the actuary’s time is wasted on dredging the river, and diverted away from diving for new trends that can help the business improve its decision-making.
Gráinne: The world is moving, and you need to keep up with it. Furthermore, ML and data science are popular – if you don’t give your team the opportunity to develop and use these skills, not only are you missing out on potential insights, but you may also find it hard to attract or retain high-calibre employees.
Isabelle: Four years ago, when I joined the industry, ML skills were a relative rarity and data scientists were a new concept. Now, data scientists have cemented their place in pricing teams throughout the market and many companies have their own data science teams. With the proliferation of ML papers in reserving, particularly relating to the concept of individual loss development models, it seems likely that reserving teams will not be far behind.
By embracing ML, reserving actuaries can ensure their future by providing predictive capabilities and expertise beyond what is currently possible.
WHAT PRACTICAL SUGGESTIONS DO YOU HAVE FOR ORGANISATIONS OR TEAMS WANTING TO IMPROVE ML SKILLS?
Greg Taylor: Gain familiarity with new processes through baby steps. It is better to develop skills via a sequence of small successes than to attempt a grand plan and find it collapses under its own weight. For example, you might choose to substitute a neural network for a single pricing generalised linear model or reserving exercise and try to understand any differences that emerge. Neural networks are data-hungry, so are unlikely to achieve much if applied to a 10x10 yearly triangle – but this triangle could be represented as a 40x40 quarterly triangle, where a neural network would be much more useful. You can also take advantage of a pre-existing model by boosting it with a combined actuarial neural network. Here, the neural network will commence with the given model, and search the data for any improvements that might be made.
Isabelle: Advancing a team’s ML skills is difficult without first freeing up time to do so. Think about the current processes you have in place and ask questions. Which parts of the process could you automate? Which can you afford to leave behind while you develop new processes? Which might be better placed and maintained by non-actuarial teams? Putting an ML model at the end of a faulty process will only increase the time your team spends running it. Before adding ML models to your regular processes, first consider whether the processes themselves will enable you to build and refresh these models in a timely manner. The first step to building ML models is therefore constructing a process that enables them to be used.
Jacky: This can be a serious cultural change if your organisation feels stagnant, so adopt change management techniques. Use small projects to get success stories and promote awareness. Build up knowledge and capability in R and/or Python. Be prepared to invest time in reinforcing this over the long term – don’t let things slip. Recruit allies in the data science and data engineering teams and gain their support – actuaries don’t have to go it alone.
John: Try to find a small band of analysts and actuaries who are keen to rehearse some of the techniques, and point them towards the working party’s Foundations blog, where there are plenty of solid articles to help them get started. You can even apply the methods to triangular data, which is easy to set up and covered in our blog at bit.ly/MLWP_Triangles. Ask your IT team to install R and/or Python, both of which are free.
Gráinne: Start with small projects where your team knows the data and experience very well, so you’ll know whether the ML process is giving you sensible results or not.
Set up a proof-of-concept for one line of business – for example, take an end-to-end process automation from quarter-end data to monitoring to automatic updated valuation, and run this for several quarters so you can build confidence in the process. Make it fun – a good starting point might be to find a suitable competition on kaggle.com (or run your own) and let your team learn new skills as they try to inch their way up the leaderboard.
A USEFUL TOOL
There are many ways in which ML can be used to improve the reserving process. Instead of agreeing a single best-practice approach, actuaries will more likely incorporate ML into their skillset, applying their judgment when it comes to what data to use, which techniques to apply and when to apply them, as well as interpreting the output. We need people with both actuarial and data science skillsets, and we need employers to give actuaries the time to develop these skills.
Sarah MacDonnell is a consultant with LCP and chair of the Machine Learning in Reserving Working Party
John McCarthy is an associate director at WTW
Gráinne McGuire is a director at Taylor Fry and the Foundations Workstream lead for the working party
Jacky Poon is head of finance at nib Travel and convenor of the Young Data Analytics Working Group in Sydney
Greg Taylor is adjunct professor of risk and actuarial studies at the University of New South Wales, Australia
Isabelle Williams is a data scientist with four years actuarial experience
Image credit |istock