Estimating social inflation by analysing jury verdicts

Let me give you a thought experiment to motivate why we might care about social inflation. Let’s say you’re the CUO of an insurance company with a decent sized US excess casualty book. For 2018 year of account, the book is currently booked to a 140% ULR. Obviously, this was not your original business plan loss ratio, and most likely writing the business at these prices has made a lot of people very angry and been widely regarded as a bad move[1]. You have probably not fully paid out 2018 yet, there’s still outstanding and IBNR in that number, but 2018 is reasonably mature, so you think 140% is probably a pretty credible estimate of where it will end up.

Let’s estimate where the same casualty book sits today in terms of profitability compared to the 2018 YOA. Since 2018, let's say you’ve pushed through an average of 10% rate per annum (defined here as rate change above economic inflation). That’s good! Now, all else equal you’re writing to a 65% ULR. If that’s the end of the story, you should probably go out and write as much business as you can at these terms.

But what about social inflation? We need to subtract that back out because our definition of rate was just ‘rate above economic inflation’.

@ commons.wikimedia.org/wiki/User:Jovianeye

If social inflation has been say 3%, you’re actually writing now to an 83%, all else equal. Maybe that’s still okay once you account for PV, but we’d want to be more cautious where we end up once we add on brokerage and internal expenses.

If social inflation has been 6% instead, you’re now writing to over 100% ULR. We’re creeping into the territory of making people angry again.

And if social inflation has been say 10%, the social inflation and rate cancel each other out and you’re back to writing to a 140% ULR. Which is the same as the very bad year you had in 2018… and let’s not even talk about what social inflation of >10% looks like.

Three of the above numbers we’ll have a pretty decent view on. We’ll know our ULR for 2018, we’ll be able to calculate the risk adjusted rate change we’ve achieved since then, and we can look up economic indices for economic inflation. Therefore, putting a number on the other piece of the puzzle, social inflation, is going to be the hard but valuable part.

How to estimate social inflation in the US

As we touched on above, the definition of social inflation is something like ‘the level of inflation over and above general economic inflation’. Meaning we can’t just use standard economic indicators, as social inflation is by definition everything not already captured by these indexes. Economic inflation gives us a floor for our total inflation value, but doesn’t help with how much to add on top. Instead, the only real way to come at it, is to approach it as an empirical question, ‘how much have losses increased by in practice, and is this likely to get better or worse in the future?’. From which we can then subtract out economic inflation.

To analyse the empirical increase, we’re obviously going to need some data. If you’ve got a nice big dataset of individual incurred losses over time, great, that should definitely be your starting point. We’d expect different inflation dynamics in say US trucking liability compared to SME retail general liability, so a dataset of losses relevant to that class is going to be the main thing you would base your analysis on.

There are however reasons that it might be nice to have an additional benchmark number. Firstly, it would be great to know, in general, are these trends getting better or worse over time? And also, as with any estimation problem, it’s always helpful to have a more generic benchmark estimate, based on a bigger or independent dataset which we can use to credibility weight our more specific number which might be based on less data.

The dataset

The dataset we’re going to use in this analysis is based on the top 100 jury awards in the US from 2012-2025, as published by topverdict.com.[2]

And before we dive into the analysis, just a brief note on verdicts vs. settlements. A vast majority of insurance claims are resolved as settlements, but a small minority can’t be resolved this way and do go to trial. The dataset we are working with is just the subset of claims that went to trial and for which there was a ‘verdict’. Given these are a small minority, is it still worthwhile analysing verdicts? Firstly, it’s the public dataset we have, as most settlements are not in the public domain. Arguably a settlement database would be more relevant, but we're simply not likely to be able to build one based on public data. And also in practice, the expected value of a verdict should the claim go to trial is what informs any settlement offer or negotiation. Meaning analysing verdicts should give us a steer on what is happening with settlements. It won’t be exact, but it should be directionally correct.

The analysis

In all the below, we’re going to be analysing inflation by looking at the growth in the Nth largest claim over time. An explanation of this method can be found below, and I won’t spend too much time on it in this post.

www.lewiswalsh.net/blog/backtesting-inflation-modelling-median-of-top-x-losses

Step 1 – download data

A year ago, collecting the data would have been a laborious task of copy and pasting, but I just pointed Claude Code at the website and told it what I would like, and it managed to download it within minutes and output in a nice csv format. I did a few spot checks, and it all looks good.

I then stripped out any verdicts which are not relevant to excess casualty insurance. Things like patent disputes or med-mal claims. This took out around half of the verdicts and left us with a dataset of around 700 datapoints rather than the original 1400. Its a shame to lose this much data, but this should make our analysis a much better proxy for US excess casualty inflation.

Step 2 – EDA

Now we have the data downloaded, let’s have a quick look at it. I created a chart showing the Nth largest verdict by year for N = 1, 10, 20, and 35. ​
​​

A couple of points jump out straight away.

• There does appear to be a general upwards trend over time, which is consistent with a positive inflation rate. So far so good.
• 2020 looks unusually low. This is due to Covid causing a drop in economic activity, but also courts being closed
• 2021 for N = 1 looks very high, but then for N=10,20,35 still looks marginally lower than expected. This may be due to courts partially reopening in 2021, or may just be volatility.

My read is that we’re going to have to take N = 1 with a pinch of salt as it looks pretty volatile, even leaving aside 2021.

Step 3 – modelling
​
Based on our EDA, I’m going to drop N = 1 completely, and I’m going to run 3 versions – one with all the years, one excluding 2020, and another excluding 2020 and 2021.

Here’s a graph showing the fitted trend lines.

Let’s talk about what we are seeing.

• The lines don’t visually look massively different, but there is around a 1-2% difference between them, which compounded across 13 years, is pretty material.
• Once we ignore 2020 and 2021, we get a pretty good visual fit. Given the y axis is a log scale, a straight line is actually a geometric increase not a linear increase, which is consistent with modelling inflation as something like (1+i)^n.
• Also, if we look at the R^2 values in the table, there’s a clear improvement in fit when dropping 2020 as expected.
• Based on the EDA and the fact we know Covid likely had a strong distorting effect, I’m just going to take the ‘excluding 2020 and 2021’ view forward and drop the other two views.

Let's discuss the actual inflation numbers for the excl. 2020/2021 view. We’ve arrived at a value of between 11% - 19% for the years 2012-2025 depending on which N you look at. That’s a fairly chunky inflation number, and to be honest, the upper end is probably higher than my prior range would have stretched to.

Step 4 – changes over time

The inflation range derived above is one take away from the analysis, but it’s not the end of the story. We said earlier that another thing we were interested in was how social inflation had changed over time. To look at this, let’s look at the inflation estimate for a rolling 7 year window instead of using all the years at once. This way we can see if the inflation was say lower at the start and then picked up at some point. For this trailing analysis, we're just using the excl. 2020/2021 view.

Interestingly, we have a general upward trend in inflation rate over time. Our 11%-19% range was really more like  7-12% in earlier windows, and 12-23% in 2017-2025. Which is broadly consistent with market commentary that social inflation really starting to appear in loss data in earnest in the 2017 -2018 AY onwards.
 
Conclusion

We saw that relevant US jury verdicts, which we are taking as our proxy for US excess casualty inflation (social and economic component combined) has been in the range 11-19% in the window 2012-2025. The rate of inflation has also increased over that period, and in more recent years is running at more like 12-20%. Returning to our original motivating example, if economic inflation has been 3-4% in the US in that period then social inflation has been in the range of 7-15%. Comparing against the 10% rate we said we achieved since 2018, our ULR is going to be more like 112-200% than the 65% we initially hoped for.


Caveats and other thoughts

• I haven't investigated to what extent TopVerdict is simply capturing more of the verdicts over time. An increase in coverage for 2025 vs. 2012 would come through as an increase in trend. In fact, what we are measuring here is really a rolled up frequency and severity trend.
• TopVerdict only shows initial verdicts, and does not reflect appeals and other mechanisms that might reduce the actual amount paid for these cases.
• The size of losses we are analysing are in the \$40m - \$100m range for N=35, and \$100m-\$500m for the N=10/20 range. A majority of US GL claims will be much lower than this, and it’s unclear to what extent these estimates would apply to those smaller claims. Therefore the number we are deriving is more applicable to excess casualty.
• To arrive at an aggregate inflation number across all claim sizes, I think one reasonable approach would be to use a weighted average of heads of damage by claim size. Something like losses >\$100m will be driven by 80% social inflation dynamics and 20% cost of indemnity component. Whereas losses of ~\$5m might be more like 20/80 the other way.
• Just because verdicts are increasing by X% a year, doesn't mean that settlements are. I think there are arguments that the fear of large verdicts may actually increase settlements by more than this, or conversely, its possible that confidential settlements are getting done at a lower effective inflation rate.

 
All the code I used in the analysis can be found here [3]

[1] to quote Douglas Adams
[2] https://topverdict.com/lists/2025/
[3] https://github.com/Lewis-Walsh/USJuryVerdictSocialInflation/