Most of the activity in the life settlements market today consists of tertiary blocks of business re-trading into increasingly stronger hands at increasingly fairer prices. Fresh capital continues to flow in and, since tertiary brings no new supply, secondary origination ought to be the natural home for that money. For some reason the secondary market falls short. Consumer education and marketing costs are often cited as major headwinds to growth in the secondary market, but there’s a more local, intrinsic effect which tends to be less obvious, especially to those on the buy side.
Basic life settlements probabilistic pricing tells us that cases[1] with a premium stream greater than the death benefit will have a negative present value. If, like me, you were buying policies in the mid-2000’s, negative valuation wasn’t an outlier, it was the norm. Hours were wasted loading case after RSI-inducing case into a spreadsheet, typing in AV and CSV through to age 121, only to bin most of them because the first two LE50s were five years apart.
To be a viable life settlement, a case almost always needs to show a material change in the policyholder’s health since the policy was issued. A second supporting factor is a low premium rate on the policy, maybe due to some aggressive assumptions by the carrier at issuance in order to print business. If neither of these are present, the case will more than likely have a negative present value from an investor’s standpoint. And that’s before fees — and fees are key.
Buy-side pricing was tedious but at least all the information was provided. Secondary market originators face their own triage pain, but they often don’t have the source information. Given that most cases will not ultimately qualify as life settlements, originators must deploy resources carefully, be that dollars to buy LE reports, or time in calling an uncooperative carrier for an illustration. Origination fees only come from cases that actually settle, but this income then has to cover all costs incurred — the successes and the failures. It follows, then, that a relatively high number of failures drives fees up in the aggregate. If we imagine the universe of candidate life settlements, some cases would be successes if not for the fee load. So, a more efficient triage process leads to lower fees which in turn leads to increased supply.
This process is known as pre-qualification (“pre-qual”). If we know most cases will be a No, how do we get to that No as fast and cheaply as possible. There are two distinct unknowns to estimate: the insured’s health status and the policy’s premium profile. These are the two main inputs to a life settlement’s future cashflow profile, so they have a connection in prequal. The more certain we are that one of them is at an extreme, the less we need to know about the other. To see this, take a boundary case. If a policy is paid up, i.e., the future premium stream is zero, then we really don’t need to know the LE to conclude the case will price positively in the market. Similarly, though less extreme, if we know an insured is experiencing, say, some form of cognitive decline then the policy is likely to qualify regardless of the premium structure. Keep in mind that prequal, like life settlements is probabilistic in nature. We’re not looking for perfect accuracy on each case, and nor are we looking for a valuation. It’s more a binary yes/no over whether a case will achieve a certain threshold price.
The realisation that machine learning could be useful in prequal came around 2016 when carriers started to increase cost of insurance (“COI”) rates. COI is the main input a carrier uses to set premiums levels. Retroactive increases in COI were economically devastating for life settlement investors but were more like drone strikes than carpet bombing. And fund managers found themselves suddenly having to explain a risk that never was, and then figure out where the next strikes were going to be. I looked at this prediction issue for a major manager, and it had the feel of a classifier problem in machine learning.
At a non-technical level, we show a model a few thousand life settlement cases. We provide features for each case, e.g., age, sex, smoking status… and the (pre any increase) COI curve. We also tell the model whether the COI on that case was increased or not. The model then tries to separate the two classes – increased and not increased – using a mix of the input features. Setting this up is a delicate balance. Ideally, we’d have millions of cases for model training and a meaningful number of COI increases. We didn’t, so feature selection must be judicious, or the model will lock onto some chance relationship that explains the cases its seen but doesn’t generalise to the future unseen cases – which is the entire purpose of the exercise. There’s also a problem that negative cases aren’t quite a “no”, they could potentially be a “yes but not yet.”
Early results were patchy but there was some hint of predictability. There then followed another wave of increases by other carriers which focussed attention more giving us an expanded training set, more positive cases to train with, and a more corrected target set. Model performance increased dramatically, and John Hancock Performance UL was spat out as a very likely candidate for increase. The next question was, if you can predict where the increases will come, can you also predict how much they’ll be. Yes, we could. The model had learned a solid representation of market-wide COI structure and was effectively detecting cases that were priced far from a fair value level. The gap to fair value gave the magnitude of a likely COI shift, and from there it’s just maths to get to the premium effect the policyholder will face. When looked at this way, the seeming mystery about why increase amounts varied so widely dissolved.
The final step was to realise that with enough training examples, the model could learn the COI structure of the entire market. Suppose you’re looking at a Transamerica Trans-XYZ from 2008 on a 76yr old male preferred, and you can’t get an illustration. At an intuitive level, the model looks at surrounding cases that it does know about. It “knows” the relative cheapness of cases in that 2008 vintage versus others. It knows how Trans priced against, say, AXA at that time; it knows how preferred vs standard risk is priced, etc. It synthesises these weak relationships, and, in the aggregate, they become strong — coalescing on a COI estimate that’s, probably, good enough for prequal.
Matthew Sheridan is a former life settlements portfolio manager and author of ‘The STOLI Worm’.
Footnotes:
[1] I use ‘case’ to mean a particular policy paired with a particular life.
Any views expressed in this article are those of the author(s) and do not necessarily reflect the views of Life Risk News or its publisher, the European Life Settlement Association