An interactive hurricane damage model for Texas

For most people, hurricane modeling is kind of a black box. Various experts set forth figures on the distribution of losses or statistics derived from those distributions. You pretty much have to take their word on it. I think policy discussions are better when the data is more transparent. So, a few weeks ago I sent a public information request to the Texas Windstorm Insurance Association asking for the raw data that they used to model hurricane losses. TWIA cooperated and send back about 30 megabytes worth of data.

So, I’m now able to create an interactive tool that lets you model the losses suffered by the Texas Windstorm Insurance Association from tropical storms. To run the tool, you will need to get the free CDF plug in from Wolfram Research. Once you have the plugin, you can run any CDF file. CDF is basically like PDF except that it permits interaction.

[WolframCDF source=”http://catrisk.net/wp-content/uploads/2012/12/analyzing-air-2011-data-for-catrisk.cdf” CDFwidth=”630″ CDFheight=”790″ altimage=”http://catrisk.net/wp-content/uploads/2012/12/air-data-static.png”]

Once you have the tool, you can do many things.

You can use the landfall county control to choose a county in which the storm first makes landfall. Notice that some of the counties are outside Texas because storms may first make landfall in, say, Louisiana but then go on to go over Texas and do damage here.

You can restrict the storms under consideration to various strength levels. I’m not sure, honestly, how AIR classifies tropical storms that don’t make hurricane strength. Perhaps they list them as Category 1. Or perhaps — and this would result in an underestimate of damage — they don’t list them at all.
You can also limit yourself to major hurricanes (category 3 or higher) or non-major hurricanes (categories 1 and 2).

You then get some fine control over the method of binning used by the histogram. If you’re not an expert in this area, I’d leave these two controls alone. In the alternative, play with them and I think you will get a feel for what they do. Or you can check out documentation on the Mathematica Histogram command here.

You then decide whether you want the vertical scale to be logarithmic or not. If some of the bin heights are very small, this control helps you see them. If you don’t remember what a logarithm is, you might leave this control alone.

Finally, you choose what kind of a histogram you want to see. Common choices might be a Count or an Exceedance Curve (Survival Function).

The tool then produces the histogram you have requested and generates a number of useful statistics. Here’s a guide to the six rows of data.

Row 1: This is the mean loss from storms meeting your selection criteria.
Row 2: This is the mean annual losses from the types of storms you have selected. This number will be lower than the mean storm loss because Texas (and all of its subdivisions) average less than one storm per year. Many years there are no storms.
Row 3: This is the worst loss from 100 storms. Note again, this is NOT the mean loss in 100 years. Some years have no storms; occasionally some years feature multiple storms.
Row 4: The AIR method for generating storms can be well approximated by a Poisson distribution. Here, we find the member of the Poisson family that best fits the annual frequency data for the selected storms.
Row 5: The AIR method for generating storms can be decently approximated most of the time by a LogNormal distribution. Here, we find the member of the LogNormal family that best fits the loss data for the selected storms.
Row 6: I can create a new distribution that is the product of a draw from the Poisson distribution and the LogNormal distribution. I can then take 10,000 draws from this distribution and find the size of the annual loss that is higher than 99% of all the annual losses. This lets me approximate the 1 in 100 year loss. Notice that this number will move around a bit every time you tinker with the controls. That’s because it is using an approximation method based on random draws. Every time you change a control, new random draws occur. Still, it gives you a feel for that dreaded 1 in 100 year annual loss.

If people have additional features they want added to this tool, please let me know. I may be able to modify it or build a new tool with related capabilities.

It’s (close to) a Weibull — again!

You recall that in my last post, I went through an involved process of showing how one could generate storm losses for individuals over years.  That process, which underlies a project to examine the effect of legal change on the sustainability of a catastrophe insurer, involved the copulas of beta distributions and a parameter mixture distribution in which the underlying distribution was also a beta distribution. It was not for the faint of heart.

One purpose of this effort was to generate a histogram that looks like the one below that shows the distribution of scaled claim sizes for non-negligible claims. This histogram was obtained by taking one draw from the copula distribution for each of the [latex]y[/latex] years in the simulation and using it to constrain the distribution of losses suffered by each of the [latex]n[/latex] policyholders in each of those [latex]y[/latex] years.  Thus, although the underlying process created an [latex]y \times n[/latex] matrix, the histogram below is for a single “flattened” [latex]y \times n[/latex] vector of values.

Histogram of individual scaled non-negligible claim sizes

Histogram of individual scaled non-negligible claim sizes

But, if we stare at that histogram for a while, we recognize the possibility that it might be approximated by a simple statistical distribution.  If that were the case, we could simply use the simple statistical distribution rather than the elaborate process for generating individual storm loss distributions. In other words, there might be a computational shortcut that could approximate the elaborate proces.  If that were the case, to get the experience of all [latex]n[/latex] policyholders — including those who did not have a claim at all — we could just upsample random variates drawn from our hypothesized simple distribution and add zeros; alternatively, we could create a mixture distribution in which most of the time one drew from a distribution that was always zero and, when there was a positive claim, one drew from this hypothesized simple distribution.

Continue reading

An idea for future TWIA finance

Although they may thoroughly disagree on the direction in which reform should go, almost everyone agrees has come to agree with what I predicted in 2009:  TWIA finances are in serious need of reform.  This blog entry sketches out one direction in which TWIA might proceed.  The idea here is that TWIA should, in a steady state, have enough cash on hand in its catastrophe reserve fund to pay for insured losses and operating expenses, without having to borrow, with a high probability, say 99%.  Further TWIA should have borrowing capacity to address the rare situations (say 1% of years) in which its reserves would be inadequate. Those borrowings should be repaid by some percentage of TWIA policyholders, persons living on the coast, and Texans generally, perhaps collected through the proxy of insurers doing business in Texas.

Although people can quarrel about the precise parameters in this abstract statement of the goal, I have some hope that people could agree on the concept. Government-sponsored insurance companies that don’t have the right to draw on the government fisc, ought not to be relying heavily on post-event bonding as a way of paying claims; instead they need enough money in their piggy bank just as we require of their private insurer counterparts. But what if TWIA’s catastrophe reserve fund does not meet this lofty goal?  What then?  Especially given the magnitude of the current reserve shortfall and the current economy, matters can not be corrected overnight. There should, I say, be an adjustment period during which premiums are adjusted (either upwards or, at some hypothetical future time, downwards) such that, at the end of the adjustment period, things come into balance and the catastrophe reserve fund meets the goal.

How do we operationalize this idea? Here is the beginning of a statutory draft. I’ve put in dummy statute section numbers for ease of reference. Obviously, the real section numbers would have to be revised by legislative counsel. Also, we’re probably going to have to develop a more comprehensive process for 2210.355A(b)(1) and reconcile this provision with the alternative process currently set form in 2210.355A.

2210.355A

(a) Definitions

(1)  The “Exceedance Function for the catastrophe year” is a function that approximates the probability that insured lossses and operating expenses in the catastrophe year will exceed a specified dollar amount. Insured losses shall be computed on a net basis after consideration of any reinsurance or other sources of recovery.

(2) The term “Loss PDF” means the probability distribution function mathematically associated with the Exceedance Function.

(3) The term “Century Storm Reserve Adequacy” means having a catastrophe reserve fund at the start of each catastrophe year such that this fund would be able, without additional borrowing, to fully pay insured losses and operating expenses in the following catastrophe year with a 99% probability as computed using the Exceedance Function for the catastrophe year.

(4) The term “Reserve Adjustment Period” means ten years.

(b)

(1) The Association shall, prior to the start of each catastrophe year, use the best historical and scientific modeling evidence with considerations of standards in the business of catastrophe insurance, to determine the Exceedance Function and associated Loss PDF for the catastrophe year.”

(2) If, at any time, the Association finds that its catastrophe reserve fund at the start of a catastrophe year does not achieve Century Storm Reserve Adequacy,  the Association shall adjust the premiums to be charged in the following year either downwards of upwards as appropriate such that, were:


(A) such premiums to be charged for the Reserve Adjustment Period on the base of currently insured properties;

(B) insured losses and operating expenses of the Association to be for the Reserve Adjustment Period at the mean of the Loss PDF for the catastrophe year; and

(C) the Association were to earn on any reserve balances during the Reserve Adjustment Period the amount of interest for reasonably safe investments then available to the Association,

the catastrophe reserve fund at the end of Reserve Adjustment Period would achieve Century Storm Reserve Adequacy.

(c) By way of illustration, if the Exceedance Function takes on a value of 0.01 when the size of insured losses and operating expenses is a equal to 440 million dollars and the mean of the Loss PDF for the catastrophe year is equal to 223 million, the initial balance of the catastrophe reserve fund is 100 million dollars and the amount of interest for safe investments then available to the Association is equal to 2% compounded continuously, then the premiums charged for the following calendar year should be equal to $614,539,421.

And what happens, by the way, if a storm hits that exceeds the size of the catastrophe reserve fund?  Stay tuned.  I’ve got an idea there too.

How do we keep premiums low under this scheme?  Likewise, stay tuned.  Hint: think about coinsurance requirements and lower maximum policy limits.  Think about carrots to get the private insurance industry writing excess policies on the coast with ever lower attachment points.

  • Footnote for math nerds only. Anyone seeing the implicit differential equations in the model and the applications of control theory?
  • Footnote for Mathematica folks only. Here’s the program to compute the premium. Note the use of polymorphic functions.

p[\[Omega]_, \[Mu]_, q_, c_, r_, z_] :=
x /. First@
Solve[Quantile[\[Omega], q] ==
TimeValue[c, EffectiveInterest[r, 0], z] +
TimeValue[Annuity[x – \[Mu], z], EffectiveInterest[r, 0], z],
x];
p[\[Omega]_, q_, c_, r_, z_] :=
With[{m = NExpectation[x, x \[Distributed] \[Omega]]},
p[\[Omega], \[Mu], q, c, r, z]]

  • Footnote for statutory drafters. Note the use of modular drafting such that one can change various parameters in the scheme (such as the 10 year adjustment period) without having to redraft the whole statute.

It’s a Weibull

To understand the premiums charged by the Texas Windstorm Insurance Association and the current legal and financial issues being debated in Austin, you have to get your hands a little dirty with the actuarial science.  You need to have some ability to model the damages likely to be caused by a tropical cyclone on the Texas coast.  Now, to do this really well, it might be thought you need an awful lot of very fine data.  In fact, however, you can do a pretty good job of understanding TWIA’s perspective by just reverse engineering publicly available information.

What I want to show is that the perceived annual exposure to the Texas Windstorm Association can be really well modeled by something known in statistics as a Weibull Distribution. To be fancy, it’s a zero-censored three parameter Weibull Distribution: 

CensoredDistribution[{0, ∞},
 WeibullDistribution[0.418001, 1.26765*10^8, -4.81157*10^8]]

We can plot the results of this distribution against the predictions made by TWIA’ s two consultants: AIR and RMS. The x-axis of the graph are the annual losses to TWIA.  The y-axis of the graph is the probability that the losses will be less than or equal to the corresponding amount on the x-axis. As one can see, it is almost a perfect fit.  For statisticians, the “adjusted R Squared” value is 0.995. 

Image

 

How did I find this function? Part of it is some intuition and some expertise about loss functions.  But a lot of it comes from running a “non-linear regression” on data in the public domain.  Here’s a chart (an “exceedance table”) provided by reinsurance broker Guy Carpenter to TWIA.  It shows the estimates of two consultants, AIR and RMS, about the losses likely to be suffered by TWIA.  Basically, you can use statistics software (I used Mathematica) to run a non-linear regression on this data and assume the underlying model is a censored Weibull distribution of some sort.  And, in less than a second, out pop the parameters to the Weibull distribution that best fit the data. As shown above it fits the AIR and RMS data points really well.  Moreover, it calculates the “AAL” (the mean annual loss to TWIA) pretty well too.

 

Image

In some forthcoming posts, I’ m going to show what the importance of this finding is, but suffice it to say, it explains a lot about the current controversy and suggests some matters to be examined with care.