We previously introduced a method of deriving large loss claims inflation from a large loss claims bordereaux, and we then spent some time understanding how robust the method is depending on how much data we have, and how volatile the data is. In this post we're finally going to play around with making the method more accurate, rather than just poking holes in it. To do this, we are once again going to simulate data with a baked-in inflation rate (set to 5% here), and then we are going to vary the metric we are using to extract an estimate of the inflation from the data. In particular, we are going to look at using the Nth largest loss by year, where we will vary N from 1 - 20.

Photo by Julian Dik. I was recently in Losbon, so here is a cool photo of the city. Not really related to the blog post, but to be honest it's hard thinking of photos with some link to inflation, so I'm just picking nice photos as this point!

We've been playing around in the last few posts with the 'Nth largest' method of analysing claims inflation. I promised previously that I would look at the effect of increasing the volatility of our severity distribution when using the method, so that's what we are going to look at today. Interestingly it does have an effect, but it's actually quite a subdued one as we'll see.

I'm running out of  ideas for photos relating to inflation, so here's a cool random photo of New York instead. Photo by Ronny Rondon

In the last few posts I’ve been writing about deriving claims inflation using an ‘N-th largest loss’ method. The thought popped into my head after posting, that I’d made use of a normal approximation when thinking about a 95% confidence interval, when actually I already had the full Monte Carlo output, so could have just looked at the percentiles of the estimated inflation values directly. 

Below I amend the code slightly to just output this range directly.

Continuing my inflation theme, here is another cool balloon shot from João Marta Sanfins

In my last couple of post on estimating claims inflation, I’ve been writing about a method of deriving large loss inflation by looking at the median of the top X losses over time. You can read the previous posts here:

Part 1:  www.lewiswalsh.net/blog/backtesting-inflation-modelling-median-of-top-x-losses
Part 2: www.lewiswalsh.net/blog/inflation-modelling-median-of-top-10-losses-under-exposure-growth
​
One issue I alluded to is that the sampling error of the basic version of the method can often be so high as to basically make the method unusable. In this post I explore how this error varies with the number of years in our sample, and try to determine the point at which the method starts to become practical.

Photo by Jøn

I previously wrote a post in which I backtested a method of deriving large loss inflation directly from a large loss bordereux. This post is an extension of that work, so if you haven't already, it's probably worth going back and reading my original post. Relevant link:
www.lewiswalsh.net/blog/backtesting-inflation-modelling-median-of-top-x-losses
​
In the original post I slipped in the caveat that that the method is only unbiased if the underlying exposure doesn’t changed over the time period being analysed. Unfortunately for the basic method, that is quite a common situation, but never fear, there is an extension to deal with the case of changing exposure.

Below I’ve written up my notes on the extended method, which doesn't suffer from this issue. Just to note, the only other reference I’m aware of is from the following, but if I've missed anyone out, apologies! [1]

St Paul's, London. Photo by Anthony DELANOIX

I was reviewing a pricing model recently when an interesting question came up relating to when to apply the policy features when modelling the contract. 

​The Python library SciPy, contains a version of the Lomax distribution which it defines as:
$$f(x,c) = \frac{c}{(a+x)^{(c+1)}}$$
Whereas, the ‘standard’ specification is [1]:
$$f(x,c, \lambda) = \frac{c \lambda ^ c}{(a+x)^{(c+1)}}$$
Which is also the definition in the IFoA core reading [2]:

​​So, how can we use the SciPy version of the Lomax to simulate the standard version, given we are missing the $ \lambda ^c​$ term? ​

I noticed something surprising about the Maximum Likelihood Estimator (MLE) for a uniform distribution yesterday.

Suppose we’re given sample $X’ = {x_1, x_2, … x_n}$ from a uniform distribution $X$ with parameters $a,b$. Then the MLE estimator for $a = min(X’)$, and $b = max(X’)$. [1] All straight forward so far.

However, examining the estimators, we can also say with probability = 1 that $a < min(X’)$, and similarly that $b > max(X’)$. Isn't it strange that the MLE estimators are clearly less/more than the true values? 

So what can we do instead?

I recently received an email from a reader asking a couple of questions :

"I'm trying to understand Net vs Gross Quota shares in reinsurance. Is a 'Net Quota Share' always defining a treaty where the reinsurer will pay ceding commissions on the Net Written Premium? ... Are there some Net Quota Shares where the reinsurer caps certain risks (e.g. catastrophe)?"

It's a reasonable question, and the answer is a little context dependent, full explanation given below.

I wrote a quick script to backtest one particular method of deriving claims inflation from loss data. I first came across the method in 'Pricing in General Insurance' by Pietro Parodi [1], but I'm not sure if the method pre-dates the book or not.

In order to run the method all we require is a large loss bordereaux, which is useful from a data perspective. Unlike many methods which focus on fitting a curve through attritional loss ratios, or looking at ultimate attritional losses per unit of exposure over time, this method can easily produce a *large loss* inflation pick. Which is important as the two can often be materially different.

Source: Willis Building and Lloyd's building, @Colin, https://commons.wikimedia.org/wiki/User:Colin

In which we recreate the previous analysis, but in Python this time. And then add a new submission using Mean rather than median to impute missing values.





Source: https://somewan.design

​

​In which we start a new Kaggle competition, submit a dummy attempt, and then build a very basic Excel model to establish a baseline for future progress.
​




​Source: https://somewan.design

Quota Share contracts generally deal with acquisition costs in one of two ways - premium is either ceded to the reinsurer on a ‘gross of acquisition cost’ basis and the reinsurer then pays a large ceding commission to cover acquisition costs and expenses, or premium is ceded on a ‘net of acquisition’ costs basis, in which case the reinsurer pays a smaller commission to the insurer, referred to as an overriding commission or ‘overrider’, which is intended to just cover internal expenses.

Another way of saying this is that premium is either ceded based on gross gross written premium, or gross net written premium.

I’ve been asked a few times over the years how to convert from a gross commission basis to the equivalent net commission basis, and vice versa. I've written up an explanation with the accompanying formulas  below.

Investopedia defines TBVPS to be:

"Tangible book value per share (TBVPS) is the value of a company’s tangible assets divided by its current outstanding shares." [1]

I'm pretty sure this is incorrect. or at best misleading!

Did you know about this cool tool, which allows you to download data from a Wikipedia table as a csv:

wikitable2csv.ggor.de/