Mark Saunders, managing director of risk solutions at Conning, spoke with Insurance Asset Risk about the role of software development in the insurance sector, talking about the importance of transparency and documentation, and the limitations of AI.

A softball question first, given that you have won 'Tech Provider of the Year'. What makes Conning stand out, and puts you head and shoulders above the rest?

Conning has got three business units—asset management, which is the biggest part; insurance research; and software—and we all work closely together. If I'm in the London office and I have a question about the investment side, I can walk to their desk. We're not just technical people building products, we also have knowledge of investments and of insurance. If you look across the whole software team, the head of the software development team used to be a chief actuary for a Canadian insurance company, so he's got decades of experience in insurance. A lot of our developers, too, have worked in the insurance industry.

What, then, would be your approach to developing this technology? Are you continually designing and planning new products, or have you developed one product that you are continually updating and developing?

It's a bit of a mix. At the core of everything we do is a set of stochastic models that use sophisticated mathematical formulae to generate potential future outcomes of the economy—that's things like interest rates, inflation, credit spreads, equity returns, and other financial variables. It generates thousands of potential outcomes. It's those models that are at the core of what we offer. On top of that are the optimisation tools that can help calculate how to adjust a portfolio, so it maximises return for a given level of risk.

How can you factor into your products one-off events like the Covid-19 pandemic. How do you do that?

The thing to remember is that the market movements themselves where similar to those we have seen in the past. If you look at the movements in credit spreads, they were quite similar to those of the Global Financial Crisis. We've also seen worse equity crashes in the last thirty or forty years. When it comes to interest rate movements, the recent jump was the largest in forty years, but we have seen bigger movements in the late 70s and early 80s. So, our calibration already had the ability to generate similar sort of movements to what we see. These weren't completely novel things. The good thing about the stochastic models is that they generate these unusual scenarios already in the tails of the distribution.

Another topic that's been on everyone's mind in the last year or so has been AI. What is your take on it in the sector?

When people talk about AI, they're often thinking of the large language models such as ChatGPT. I saw a recent study that said that over 10% of the responses that ChatGPT came back with were hallucinated. That means the model just made it up. Because of this, I don't see a role for large language models in the financial modelling software that we offer. However, we do have neural networks implemented to model pension liabilities. Compared to a full, traditional model that takes all the policy data, economic data, and runs a step-by-step analysis which could take several hours, a neural network model is a lot faster, but it's not as accurate. That's fine within an SAA model because you're not trying to calculate the liabilities precisely to the nth degree, but rather you want to capture how much those liabilities would change when economic factors emerge. That's a good use case for AI.

What about the use of AI across the industry at large?

AI is obviously good at analysing unstructured data—if you have huge amounts of it, you can use AI to try and pick out patterns and so on. The sticking point is that the insurance and pensions industries are so highly regulated that I don't think you could convince a regulator that you're going with a set of modelling assumptions just because AI told you to do so. The regulator will want to explain why you came to that decision and what assumptions have been used, and why and how you made those assumptions. You'd also need to document that reasoning. We do hundreds of pages of documentation as a tech provider.

We live in interesting times—wars in Europe and the Middle East, China possibly looking sideways at Taiwan, a contentious US election. How are these factored into the tech solutions you provide?

We've got a very suitable product for that. We have the stochastic models that generate thousands of outcomes of what might happen, using these sophisticated mathematical models. We explain to clients what might happen to their portfolios if the markets move so far in one direction or the other, without having to necessarily explain the reason for that movement. And what that does is give them the chance to implement a strategy to protect themselves against these extreme scenarios