Analyst: Before the model stuff, can you frame the book for me? What were you actually underwriting?
Expert: Specialty property. Middle market mostly, plus a slice of large account layered and shared where we'd take a piece of somebody else's tower. Gross written across the group was about 1.16 billion when I left, and the genuinely cat-exposed property piece was, call it 640 million of that. No — 640 was two years earlier, before we started cutting. By the end it was nearer 610. Habitational, light industrial, a lot of retail strip, a lot of older commercial roof stock. That last bit is the one that matters. Sorry, it's a boring answer, but everything I'm going to tell you comes back to roofs eventually.
Analyst: And your tenure?
Expert: 2019 to the back end of '25.
Analyst: Let's start there then. When did your internal view of risk actually change? Not the vendor release date. When did the number you priced off move?
Expert: Those are two different dates and you're right to separate them. The vendor's severe convective update landed mid-2023, might have been September, I'd have to check. Our own view had already moved. We changed the loads in the February pricing refresh that year, and we'd been arguing about it internally since the summer of '21, thereabouts.
Expert: So — sequence, because the sequence is the whole thing here. '21 we had a hail year that on paper shouldn't have happened. Wasn't a record industry year. It was just bad for us specifically, which is worse. '22 we had another one. Two in a row and you stop calling it variance. The chief actuary put a deck in front of me, this was the November after, showing modelled against realised on the convective piece, split by year and by state. And the model was low every single year. Not low on average, which you'd expect and can live with. Low in each of them. That's not volatility, that's bias.
Analyst: How low?
Expert: On the hail and straight-line wind component, modelled average annual loss was running around 2.3 points of gross earned on that segment. Realised across '21 to '24 averaged five and a bit. 5.4. Don't hold me to the decimal, that's from the version we took into the board and we restated it twice after.
Analyst: So a two-and-a-bit times miss. Is that the model being wrong, or your exposure data being wrong?
Expert: Both. And the split is the part worth your time, because the exposure half was the bigger one and nobody wants to hear that. It's much less interesting than climate change. Doesn't get you on a panel.