Reprinted from the June issue of Carrier Management
Reserving is, on the face of it, one of the simpler elements of actuarial science (once companies get past the nuances of triangular data tables). They look at the way things behaved in the past and assume the same is going to happen in the future, sprinkling in a little industry knowledge and judgment along the way.
However, it’s fair to say that many things need to have happened to get to this stage.
Companies need to collect data consistently for long enough to have sufficient history to understand their claims’ life cycles. They need to have sufficient experience to know what a normal claim is as well as what an extraordinary or outlier claim is. A company’s data needs sufficient and consistent granularity to allow it to dig into line-specific trends.
Once companies have the data, they need to be able to organize it quickly enough into their beloved triangles to give them time to analyze it before they are required to report their findings. They need suitable applications and technology to apply their models confidently, using standard methods where possible and more esoteric methods when required.
Given the importance of the reserving estimate to a company’s balance sheet, once finished with their analysis, they need to demonstrate that appropriate controls were taken, that suitable methods were used and that reasonable judgment was imposed. There are many stakeholders, and they need them all to be confident that things were done correctly.
All of this is to say that reserving teams can’t be “mad scientists.” From a modeling standpoint, they need to do relatively simple things but do them demonstrably well.
Furthermore, reporting and documentation requirements, as well as the need to get a varied audience on board with the deliverables from their reserving analyses, impose constraints that can seem limiting compared to other elements of actuarial practice or data science.
How can reserving teams innovate when they still must be able to explain their results to the CFO in less than 20 minutes? Should reserving teams perpetually exist as actuarial Luddites, slowly explaining loss development factors to a crowd of auditors, as other analytical functions surge ahead with their GBMs (gradient boosting machines) and neural nets?
Of course not.
But it does lay out the challenge that is perhaps unique to the reserving team: The result of your work is massively critical; you only have a certain amount of time to do it; you need a diverse range of stakeholders to have confidence in your analysis; you have more raw materials than ever before to work with; and other people need to be able to not just follow it but recreate it if need be. These pressures are on top of the fact that their primary assumption—that the past is indicative of the future—is perilously close to untenable given things like pandemics, inflation (social or otherwise) and climate change driving claims trends that simply haven’t been observed before.
Automation and innovative uses of AI models have had quite the transformation in reputation over the last few years. Back in 2009, people — analytical teams especially — might bristle at the thought that their jobs could somehow be “industrialized” away. We now have autonomous vehicles from mainstream manufacturers guiding us to our destination and we have ChatGPT to write reports and articles with surprising clarity and insight.
Reserving teams are starting to see the timesaving and life-augmenting benefits of automation as opposed to being bored by the very prospect of “something happening automatically.” But are these examples—ChatGPT and autonomous vehicles—useful analogies for innovation in reserving? Perhaps, but for maybe the wrong reasons. These reasons are the stakes and the liability.
ChatGPT has incredibly low stakes. Ease of entry and a mostly low level of risk means that it can be used with little concern about the consequences of things going wrong.
Autonomous vehicles, conversely, use very sophisticated models, but at the end of the day, the driver needs absolute control over the vehicle at all times. It is their lives, the lives of their passengers and the lives of potentially chaotically behaving third parties that are at risk. Even the slightest error has grave consequences for which the individual remains responsible. With this level of responsibility, even a deviation in the style with which an autonomous vehicle is controlled puts us in an uncanny valley of transportation.
Innovation in reserving can be hindered by these factors: complexity is a barrier to entry, the stakes are high, and reserving opinions are signed by an individual, not a process.
Where can it go wrong?
It’s hard to argue against these noble intentions, but the trick is in the execution. One of the main challenges is resource availability and the cyclical nature of reserving, whereby new analyses are expected quarterly or even monthly. How do you rebuild a railway track when the train is hurtling toward it and your staff is furiously shoveling coal? Why are we deploying overly complicated solutions to relatively simple problems?
One trap that companies often fall into is going after a new approach without giving due consideration to the basics. Oftentimes, the drive for innovation results in building complex structures on very shaky foundations and hoping that sophisticated models will overcome the shortcomings of the data. This type of innovation execution typically fails as it can quickly lose credibility through a lack of integrity. And you can’t “machine learn” yourself out of a mess if you train that machine on rubbish.
Another pitfall is automating inefficiencies. If you have 20 people sewing clothes and are tasked with automating the process, you don’t design and build 20 automatons to sit at the machines and press the peddle. You completely redesign the machine from the ground up. Similarly, macros that automate button clicks and user actions don’t provide the same efficiencies as a complete redesign or transformation.
It’s also important not to automate parts of a human process. Companies must identify where humans add the most value and insert them into an automated process. This isn’t about replacing the most valuable human resources but redeploying them. Macros will only go so far to make our lives more engaging and our output more valuable. Easy as they are, piecemeal solutions won’t provide the lift in the long term that a proper transformation will.
When deciding where in reserving innovation is most profitably deployed, we need to be honest about which parts of the process are at risk from our human interactions or limitations, where perceived assumptions are not being met, and how we can deploy advanced techniques while maintaining a transparent framework and enabling reporting.
Segmentation is an area where we can, perhaps, be a little more creative in our approach to improving reserving without moving away from our standard methods. Rarely are reserving actuaries challenged on their segmentation, and the idea of homogeneity is often just assumed. With richer data assets, machine learning approaches can be leveraged to explore and inform segmentations that increase homogeneity. Once re-segmented, companies could potentially dial down the complexity of their projection methods. They can do the smart work upfront to make the ongoing processes that much simpler and less expensive to run daily.
Within algorithm-based reserving approaches also reside machine-led reserving techniques. These types of deployment could be construed as encroaching upon the actuaries’ area of expertise, but rather than being seen as a supplanter, realistically this is an enabler. Machine-led reserving considers the data available and does its best to produce an “optimized” model using traditional method components. In this way, it could be considered an only lightly experienced reserving analyst. However, it might not be the most reliable actuary—completely reliant as it is on the data that we feed it. This limitation is perhaps its greatest asset, unconstrained as it is by the anchoring bias from prior analyses or outside influences of biases. Further, machine-led reserving techniques are as—if not more—instructive not in what they think they get right now but in what they get wrong and in how they change over time. Machine-led techniques allow us to quickly explore the almost limitless possibilities of forecast estimates, scoring them and reporting back their findings with little of the time and costs burdens associated with human effort.
With all these approaches, it is critical to keep an eye on what your deliverables are. As important as it is that reserving practitioners can reach their goal of a central estimate, they need to keep in mind two things: first, it’s not good enough that they can reach that goal. They need the stakeholders of their reserve estimates to arrive at the same conclusion quickly and confidently. Second, companies are dealing with forecasts and uncertainty. Their ability to deliver assured results lies not just in their ability to prove what something is but in demonstrating what it is not.
Automation and the ability of companies to harness machine-led decision-making support is scary but also has potential. Much like autonomous vehicles, companies can destroy confidence with even the smallest of errors, or they can have the ability to demonstrate that as drivers of this process, they have full control. Deployment of end-to-end automation should be holistic, but intelligent automation needs to be approached far more strategically.
Further duplication without permission is prohibited. All rights reserved.
Reprinted with permission from the June issue of carriermanagement.com.
