The authors acknowledge that house “quality” could be encoding other biases such as race, religion or class. It would have been great if the authors explored this a bit more. How closely did insurance risk correlate with the demographics they mentioned?
How would they obtain this information? The car insurance company (presumably) didn't have it, so the researchers would have to get in touch with each individual in their dataset and have an awkward conversation along the lines of "we're working with your car insurer, please tell us your race and religion". That's not going to go down very well...
Financial situation affects both house appearance and insurance claims. Some people won't bother to file legitimate insurance claims for small things, while other people will file fraudulent claims.
House appearance and car crashes will be related by risk-taking behavior, carelessness, and being too tired. For almost any two things in life, an inability to keep one in good order will most likely predict the other.
The view from within an insurance company: insurance fraud occurs across the spectra of socioeconomic statuses. Whether people file a small claim or make a fraudulent claim is driven mostly by their personality. Often those with more feel entitled to get more back from a claim.
The difference a claimant's personal situation & qualities makes to the claim is often via prejudices of the claims adjuster. Which is why we implement automated fraud detection to help identify those who otherwise appear "honest".
They may be trolling, but it’s a pretty good example of statistical naïveté. What I mean by that is ignoring conditional independence even when it is obvious from common sense. Common sense tells us that living in an old house does not ‘cause’ car accidents, but it also tells us that a poor working elderly person may have a higher risk of both car accidents, and living in an old house. The dependence between the effects disappears when conditioned upon observations closer to the chain of causality.
This also makes me wonder if they house had any relevance at all other than a publicly available show of approximate income. If the researchers had access to all of the incomes of the people they researched would they have been able to build the exact same predictions or were more factors involved like rural/inner city drivers having different results.
They don't claim the effect is independent at all, they claim that an insurance company can decide to not bother collecting correct data about the actual causative variables, and instead just base their model off a picture of your house on GSV.
All good and well, but as far as I know, certain laws are in place that prevent insurance companies from actually using indicators that are "closer to the chain of causality". E.g. poor, elderly (an now recently, male/female). It's then no surprise that they're then actually trying to find other means of determining risk factors for individuals.
Eerily similar to the practice of Banks red-lining districts they wouldn't sell loans or mortgages to.
Legally, this is a sort of "differential impact" discrimination, and iirc it's mostly legal as long as you can justify your reasoning. I think insurers would have to justify their rate adjustments though - a 5% extra premium for comprehensive at the intersection of Skid Row and Mugger's Bend is probably legit, but a 300% extra in Chinatown would land the insurer in legal hot water. Trial lawyers make big money running class-actions for these sorts of things, which hopefully keeps businesses honest.
(IANAL, I picked this up from my mom who does labor/employment defense.)
This is a circle that is going on for very long I think.
People with private locked garage will also have better ratings in general. Wealthier places will also invest more in road maintenance and limiting accident factors (for instance having a ton more protective equipment near school zones, investing in underground parking near shopping areas etc.)
In general I feel that trying to legally screen info from insurance is a lost proposition, as they have all the data in the world and find a proxy for that info anyway.
Interesting. It might depend on the plan, we have a vandalism protection option (in-car theft, window breaking, tire damage etc.) so the insurance company has all the incentive in the world to know where and how it would be parked at night.
you're probably right that it depends on the specific coverage. I recently moved to an apartment with an access controlled garage so I did some research to see how this would affect my premiums. in general I found that it doesn't do much for your overall premiums. I don't think my car insurance offers vandalism protection specifically, although I certainly hope it falls at least partially into some other category! I can certainly imagine that the company would change the price of that specific coverage quite a bit depending on where the vehicle is parked.
> But busier roads also usually correlate with urban density which often correlates with socioeconomic status, race, crime, ...
sometimes, sometimes not. some cities (like mine) vary a lot from neighborhood. I live in a nice one but share a zip code with a poor neighborhood with lots of crime. my premiums are pretty high despite the fact that I keep my car in a locked garage and only use it to hop directly on the freeway and drive out of the city.
My car insurance offers a small tracking device to estimate if you're a risky driver. It got GPS, acceleration sensors and Bluetooth to transmit the data over your smartphone.
I registered for the beta and like it a lot. I'm paying 30% less with no effort - just because I'm not driving as crazy as everyone else.
at least on paper, data protection is fine - is being processed by a separate company with a very restrictive privacy agreement, storing and handing over only a 1-10 integer for several driving skills (speeding, acceleration, hard breaking).
It'd be cool (from a hacker point of view) to spoof the smartphone's GPS/accelerometer data: put a delay of, say 10 seconds. So you can accelerate like mad off the red light, but the blackbox app receives a spoofed acceleration curve which is gentler. I guess it's harder if you like going over the speed limit, e.g. you'd arrive to your destination after driving 90 mph, but the app still has to spoof you being 20 miles away doing 70mph (the highway limit in the UK).
Does that mean it couldn't be used to decline you coverage in the event of an accident where the data showed that, for example, you were doing 33 in a 30?
Exactly, data of incidents never makes it to the insurance company in the first place. They only get like monthly stats of your general driving skills - just had a look in the contract, it's just a single percent value that they get.
Besides, the legal requirements for denying coverage in Germany are pretty high.
If you're crashing with 130 in a 30 zone, you won't need digital evidence.
Sounds like a "bad" thing for them to do, from our perspective. But what if this actually caused those drivers to drive below the speed limit. On some level, this just easily solved a problem that's been plaguing city governments for decades if not a whole century. They've tried fines, tickets, speed-cameras, traffic-stops, police presence on the roads, etc. An entire industry has spawned off to cater for all this "societal fluff" of trying to get people to not do things by metaphorically "spanking" them. And here we have a semi-reasonable solution to the problem that's cheap, self-enforcing, and all it requires is for us to have some level of trust for the tracking devices/data/company.
True, and I believe that accident investigators are very good at accurately deducing a vehicle's speed from the accident scene. My concern here is that as an organisation the insurance company is motivated to find ways to not pay you. Providing this information to them is giving them another thing to turn to, to refuse your claim.
I'm not sure I'd want my insurance company to give me a tracking device, but I wouldn't mind getting some feedback on my driving just for the sake of personal improvement and peace of mind.
That seems pretty unobjectionable to me. It's generally good that insurance companies encourage people to behave in ways that result in fewer payouts. Car insurance companies encouraging people to drive in a safer manner, health insurance companies encouraging you to go to the gym, fire insurance companies encouraging sprinkler systems, etc.
One point that confuses me a bit: The study initially used experts (no word on the experts?) and later only selects four experts for the subsample because their kappa is moderate. Is that a common approach?
If there are two raters in a sample of six that disagree with the four others, I would argue that the IRR is an issue. Otherwise one could always run studies with n experts and only select a subsample of those raters that have a high interrater agreement, or not?
It doesn't impact the validity of the results of the study as far as I can see... If they got closer to a perfect rating system, the predictive power of the model could only be better.
You can have Google blur your house on Street View if you are so inclined. No idea in what kind of insurance profile bucket that will shunt you, but it's something to consider for the privacy conscious among us.
Click on 'report a problem' in Street View and centre the viewer you will see on your house, select 'my home' from the options, and add the address for good measure in the comment area.
A portion of the population works in fields like psych where there is standard advice to use unlisted phone numbers to deter stalkers. Given 10 tenants, it is likely one has a good reason to make it harder to picture where they live.
I really don't like this trend of considering every facet of my life when making an insurance offer. All these systems seem to lack the nuance to take into account personal circumstances. What if I bought a "bad" condition house as a fixer-upper? Does that mean I'm a bad driver?
On average yes, because it doesn't consider that facet of your life. You can't have it both ways, I suppose.
I generally agree. The whole point of insurance is to pool risk -- so cutting the population into ever-smaller pools seems missing the point. In the infinite limit we'd get seven billion single-person pools, each of which has that person pay in the exact same amount of money they get back out.
The point of insurance is to replace a high variance distribution of potential losses with a low variance distribution, but with a similar expected loss. In other words, on average you pay a bit more, but you never pay a huge amount.
Now, in a perfect world, everyone would pay exactly what their expected loss is, plus a small premium. Since it's impossible to measure the expected loss exactly, in reality some people pay a bit more and some pay a bit less. Getting a better risk estimate doesn't get you closer to "not having insurance", but rather closer to "perfect insurance" where everyone gets a distribution that exactly matches their risk profile.
How is perfect insurance different from everybody paying their own losses directly? If I correctly understand what you said, then perfect insurance would lead to nobody buying it, so they could avoid paying the useless additional “small premium.”
It's basically "no insurance, plus a credit to pay for the losses". Except that you start paying the rate before you incur the loss itself.
I still think this form of "perfect insurance" is bad - part of what I like about the insurance is exactly the dispersion of cost in the larger population. Should a chronically-ill person pay a very high insurance, or should the society cover the costs? I'm in the second camp, even though I'm not chronically-ill. It's part of the benefits of having an actual society - we don't need to stand alone.
That is indeed part of what people call insurance these days, but in fact it isn't one. It's just a social solidarity scheme, and no one has the balls to call it that way, so it gets mingled into insurance. Without getting into whether it's good or bad, it's just not the true meaning of the term "insurance."
No. Even for vernacular natural language, it only matters what native speakers think. Wéijī will never quite mean "danger and opportunity," no matter how many people view it that way, because native Mandarin speakers don't view it that way.
constantly having the prescriptivism vs descriptivism debate is tiresome and pedantic. what we call "health insurance" clearly works very differently than any other kind of insurance I'm familiar with. it's not wrong to try and maintain the distinction between two different concepts.
There is a benefit in separating the parts that are actually just insurance that replaces low-probabily low frequency costs with a predictable payment (and thus can be efficiently provided by competing capitalist companies) with the parts that are social redistribution schemes.
The latter are necessarily the government's remit, and the political discussion on whether they are desirable can be had in an informed manner if costs are clear and they are not commingled with insurance elements.
healthcare is a bad fit for the concept of insurance. an obvious example is that people use their "insurance" to pay for routine doctors appointments which have highly predictable costs and are scheduled in advance. the chronically ill person in your example enrolls in insurance knowing upfront they will never pay enough in premiums to cover their claims.
what we call "health insurance" is really just a payment planning and collective bargaining service rolled into one. it doesn't really work like most other kinds of insurance.
You never have to pay a large lump sum. So it’s similar to having savings and always covering your own losses. But with the added benefits of covering you even if you have not had time to save sufficiently to cover your losses (or lack the discipline to do so), or are unlucky and suffer multiple large losses in closer proximity than your savings and income can absorb.
It's not useless. You pay premium for the privilege of never having to pay catastrophically large amounts. And that's exactly the difference and the point of insurance: limiting the maximum loss.
The perfect insurance would know that I was going to have to pay catastrophically large amount and would therefore charge me a catastrophically large amount.
There is a difference between estimating a risk with exactitude and predicting with exactitude that an event will or will not happen. Suppose we can accurately say that I have an annual risk of a certain event happening to me as 1/1000 - this is neither a prediction that it will happen to me nor that it will not. If we form an insurance pool of 1000 people all with the same risk, and we each annually put in 1/1000 of the cost of the event to one person, we would have coverage against being unlucky (subject to the usual caveats of liquidity, uncertainty and changing circumstances.)
There's a difference between knowing the risk distribution and knowing the future.
The perfect insurance would know how likely it is that you are going to have to pay a catastrophically large amount, and would charge you amount*probability. If probability=1.0, then sure, they should charge you the whole amount.
The future becomes completely predictable as it approaches, so probability necessarily evaporates in the limit.
As the data gathered increases, the probability eventually reaches 1.0. For instance, if an insurance company could legally gather all telemetry from your car at a high enough frequency and cancel your policy at will, then it could determine that your car was on an unavoidable collision course and eliminate losses.
This may sound like a silly thought experiment, but I think it demonstrates that in principle, there must be some point at which we draw the line on discrimination to have insurance, so if you find yourself defending all discrimination then you've gone wrong somewhere. Either there's a principle that has been missed, or it is valid to just set the rules in society that people are comfortable with.
The approach to the limit is not often smooth, and the cost of insurance is not continuously reevaluated, so this is not an appropriate model for those situations in which insurance is actually the answer.
There are, however, situations where the model is a reasonable fit - notably, some aspects of health care, where the approach to the limit proceeds over multiple billing periods. A society can choose to spread the cost around, but to do this under the guise of insurance leads to the sort of incongruities and inefficiencies that are seen in the US system.
I guess there is a difference between a "infinitely accurate risk profile" and "knowing exactly what will happen in the future" (prescience).
Being prescient about getting the number 4 when you throw a dice means you can assign probability 1 to getting 4 when you throw that dice.
Having an accurate risk profile means knowing the odds you'll get 4 with that dice might be 1/6.4 because you precisely measured the shape and weight distribution of the dice and you can see how biased it is.
Let's assign a cost X for the unlucky event of rolling a 4. You're expected loss is X∙1/6.4.
If the event only happens once, obviously you either pay full cost X or you don't pay anything.
If you pool with a large number (N) of people, each throwing once that same dice, then the expected loss of the pool is N∙X∙1/6.4, that is, if each member of the pool contributes their share, the pool is going to be able to pay off the individual losses.
Unlucky people are better off because they saved X∙(1-1/6.4); lucky people are worse off because they had to pay X∙1/6.4. Pool members agree on those rules, because they don't know if they are going to win or lose, and their contribution to the pool is smaller than the worst case scenario.
So far so good. I deliberately used the "custom" probability number 1/6.4 to show how arbitrary that is.
Other people might have other dice, each biased their own way.
The expected loss for the pool would thus be X∙(P₁+P₂+P₃+....Pn).
Each person contributes to the pool backing proportionally to the contribution to the pooled expected loss, thus the system still makes sense even on the case of 1-sized buckets.
(Disclaimer: I have no idea what I'm talking about, this is not my field, just an average person with everyday math skills. Happy to learn why I'm wrong, in case I am)
That’s not what they said. That small premium is the cost of not paying directly for the large actual loss, you are paying for your predicted risk not the loss.
There isn’t likely a perfect loss prediction to be had, e.g. predicting and understanding risk itself influences the likelihood of avoiding a loss favorably. An ideal scenario might be that people become aware of risky behavior then adjust and therefore change the future prediction. Health insurers do it all the time. Helping at risk customers avoid large losses actually helps the customer and the rest of the less risky customers in the pool (everyone’s premium should go down if the risky become less risky).
So if for example health insurance is based on genetic screening and only those with certain genetic markers get a huge premium would you consider that a step towards perfection?
> Getting a better risk estimate doesn't get you closer to "not having insurance", but rather closer to "perfect insurance" where everyone gets a distribution that exactly matches their risk profile.
You say "Quite the opposite", and proceed to rephrase the same thing as GP said, except for the conclusion, which also is the same thing, yet you claim it isn't.
Not having insurance IS "perfect insurance", in that you pay exactly your own costs, with the added benefit of having no overhead. Most of us don't want that though, in the same way as we don't want children to have to pay their own healthcare, or chronically debilitated people to die to darwinism (whether we succeed here is another point). But a perfect "risk estimate" also happens to be the same thing as prescience, and if we had that we wouldn't need insurance for other reasons.
Not having insurance = paying for yourself = distribution by exact risk profile.
Not having insurance means losses follow your exact risk distribution, with a non-zero risk of a catastrophic loss. Having a perfect insurance means your losses follow a distribution with the same average as your risk profile suggests (+premium), but much lower variance and with zero density on the higher end, meaning you never have a catastrophic loss.
And that right there is where you are hiding all your assumptions. There is no such thing as "your risk profile", there is only your risk profile according to some risk model, which approaches what actually will happen to you individually the more detailed you make it and the better your predictive capabilities get, because as you do that, you are shrinking the pool of individuals that this average is calculcated over, approaching only you as one individual in the limit, at which point the average is the same as the actual thing.
> and with zero density on the higher end, meaning you never have a catastrophic loss.
if I flip a coin, I know exactly the odds of getting heads, but I don't know whether I will get heads unless I can exactly predict the future state of the world. if I had this magical ability, it would still be worth paying me a premium to tell you so you would know how much to save.
> if I flip a coin, I know exactly the odds of getting heads, but I don't know whether I will get heads unless I can exactly predict the future state of the world.
No, that is a fundamental misunderstanding of what odds are. Odds are not a property of a system, odds are a description of your knowledge about that system. There is no such thing as "odds of the system", the odds for the same system can be vastly different for different observers.
If you know how to control the coin, then the odds for you are almost 1:0 for your intended face. For the person next to you who doesn't know you, it's probably 0.5:0.5. For the person that has watched you for a while, it might be 0.6:0.4, if that's the ratio of the faces that you flip (intentionally). All for the exact same system at the exact same moment.
Also, a binary distinction between "having no clue" and "perfect predictions" is not useful. While perfect predictions are the limit of an ever refined model, there are many steps in between a useless model that gives random predictions and a model that makes no mistakes whatsoever.
This is the view of an insurance as the reversed loan, where you pay the loan before you get the principal paid out. It guess the "perfect insurance" could be obtained by reversing the reversed loan? In case of something catastrophically happening, you are simply guaranteed a loan (with fair interest rates).
To me, the aspect of redistribution is important in insurance. The acceptance that the world is chaotic and we can not predict all the bad luck that might happen to us. Hence, I pay for yours bad luck, if you pay for mine.
> Getting a better risk estimate doesn't get you closer to "not having insurance", but rather closer to "perfect insurance" where everyone gets a distribution that exactly matches their risk profile.
That is the same thing if you have produced a perfect risk profile. The perfect risk profile completely predicts reality, and all costs can be accounted for. If it doesn't completely predict reality, it's not perfect.
So with the perfect risk profile, insurance is a zero sum game. Instead of paying a small premium you can put your money in a bank account and instead have them pay you "a small premium" in the form of interest.
And no, I don't believe that there can ever be a perfect risk profile, but as risk assessment gets more fine grained, insurance approaches that zero sum game.
Yes, that's the point. By using more fine grained risk assessment, you attain less useful insurance. If the utility of insurance is to reduce variance, and utility increases as variance is reduced, the insurance with the highest utility is one that completely ignores individual and demographical circumstances.
Perhaps you understand the word "perfect" in a different sense than I do. A better risk estimate to me is (to me) one that is more accurate. A perfect risk estimate is one that can not be more accurate. If a better risk estimate brings you closer to perfect insurance (which the GP concludes), a perfect risk estimate should leave you as close to perfect insurance as is possible.
It would be cool if you could address my reasoning instead because now I have no idea what exactly you disagree with other than the conclusion.
We probably disagree on definitions, because the term risk can sometimes refer to the expected damage and sometimes to the variance of the damage. So I will try to express myself more clearly:
Let's say you have a house worth $100,000 and statistically it burns down once in 1000 years. This means that the expected damage is $100 per year, but the variance is very high because once the house burns down, the damage is 1000 times greater than the expected damage. With insurance, however, the variance is zero because you pay the same amount every year, whether your house burns down that year or not.
I would argue that with a perfect insurance you would only pay the expected loss, i.e. $100 per year, and not more or less, because ideally the insurance has accurately assessed the expected loss and diversified the variance away.
However, with less ideal risk assessment, you could pay more or less because other houses burn more or less often.
That's not at all the purpose of having insurance. Essentially, in that world, you're just outsourcing the cost of estimating what to set aside to self-insure so you know how much to save and when. Shared risk pools are exactly the original point of insurance. Now that insurance is a lucrative business, the point of insurance as described by those profiting from it will forever morph into whatever is most profitable.
If the intent described were truly the case, explain Medicare and Medicaid.
The problem with pooling risk is if the risks (ie people) are seperable, there's an adverse selection problem. People who can see they are a lower risk will evaporate from the pool, leaving the higher risk people who then would have to pay a higher premium for the pool to break even.
Theoretically yes, practically no. If coverage is universal and mandatory, companies will invent all sorts of ways to market themselves to a low-risk subset of clients -- car insurers marketing themselves as women's insurers, health insurers advertising their online capabilities to attract younger clientele.
The clean way to do it is to separate insurance (can be very efficient if privately provided) from redistribution (will lead to bad distortion unless done by the state).
You're presenting a very cliched argument, but there's something that I don't get about it.
You seem to be saying insurance companies must discriminate unfairly by saying that if insurance companies discriminate unfairly, they will lose customers.
Say I am male, and I pay more for car insurance because of that. Your argument is that if sex discrimination is illegal, then the market will collapse because women will stop buying insurance. But isn't it equally true that men who are low risk should "evaporate from the pool" in the present system? There doesn't seem to me a fundamental distinction between these two examples of discrimination.
Now, there may be a quantitative difference, a difference in economic impact, but if it's only quantitative, then why should one automatically be manageable and the other not, which is the assumption I think people implicitly make.
> You seem to be saying insurance companies must discriminate unfairly by saying that if insurance companies discriminate unfairly, they will lose customers.
It's not actually a value judgement I'm giving. Whether it's unfair as in sex or race based is not part of the incentive mechanism. The mechanism still exists though. If you find out people of one color are bigger risks than another you have an awkward problem.
If you can restrict the companies from doing this you fix the problem, but of course there's games to be played regarding selection, as another commenter notes.
I'm not using any value judgments of my own here; it's a logical problem I'm trying to focus on.
I'm using "unfair" to mean anything someone could call unfair, that is, charging someone as though they are a higher risk than they are, due to some factor.
The point is, it's illogical to say we can't* misprice risk based on ethics or social justice or whatever, when you assume we can misprice risk based on incomplete knowledge. It seems like special pleading.
*can't in the sense of it being a mortal threat to the industry in general
You're taking an ill conditioned and personally owning the responsibility of building it back to market condition. Seems like you're relatively comfortable taking risks relative to the general population.
This doesn't seem less wrong than generalizing based on something like age or gender. I would attack it on privacy grounds, not accuracy.
The whole point of the insurance is to pool the risk - a perfectly operating insurance company is essentially a rent seeking parasite on society, because they will practically never pay out the fees and continue collecting premiums.
As such we really need to reassess the usefulness of such organisations in our society - they're useful because they blunt the monetary pain of rare life events and pool the costs of chronic issues. If they stop covering people who need them the most, they stop being useful for us.
no, insurance is great. it's just not a good match for healthcare.
someone who can't afford their sky high car insurance premiums isn't someone who "needs it the most". they are someone who should be priced out of driving.
Phrased differently, it is just plain discrimination.
If they applied this to "black people are more likely to cost us X" then everybody would be against it. But if they apply it to different discriminators (age, gender, neighborhood), then suddenly it is ok?
This is why privacy and regulating information asymmetries is so important.
Consider ML model that picks 1,000 people for every 10,000 and 99% of those selected are false positives. Insurance fee is $100 and average payment is $50,000.
Rejecting those 1,000 people is $100,000 less revenue and saving $500,000 in payments. It's no brainier to reject people and let there be statistical accidents.
Insurance companies can make profit using models that have huge number of false positives. If your insurance company or bank is allowed to use all data it collects to make model of you, it leads to Kafkaesque world.
"Why my insurance premium jumped 200% and I can't afford it anymore?" Nobody knows that the reason is your dried out flower in garage widow visible to the street.
"Why I can't get mortgage?" You have too many Facebook friends with risk of defaulting and one has criminal record. Select your friends better.
Making such "friend-to-friend" lending a major part of the financing infrastructure in the US, it would clearly lead to strong economic and racial disparities.
You don't have to make that kind of stuff illegal, but you also need to make sure that you don't base important parts of your economic systems on such biased processes.
If you let them be free to do what they want poor people (or at least some people) will never be insured. But I guess that is the basis of the US healthcare system... how is that working out?
Why is it the responsibility of private insurance companies to look after society’s most vulnerable?
I think it is a much more sensible arrangement that insurance companies are free to operate as they see fit and that the government be responsible for the well being of its poorest citizens.
Should grocery and clothing stores be required to feed and clothe the poor, respectively? To me, this feels like a shrinking of government duties onto businesses.
I guess it depends on the market, healthcare is not a free market as you can never say: "Guess I'll die.", no rational human does that. Moreover, the single drug people require is often patented and thus monopolized. But people will say: "Guess I won't buy a Hugo Boss suit."
I guess you can go the route of government sponsored insurance (and everybody know governments hardly negotiate as they can print money) or you make laws and keep tight control over what insurance companies can and can't do.
That doesn't mean you have to let them assess risk based on incidental information that should be private. Assessing vehicle accident risk on your driving record is fine, assessing it based on who your friends are is not.
I think there is an EU ruling that says insurance companies can't discriminate based on genitals.
In reality it didn't work because there were many other variables that are correlated with gender such as profession or types of car driven. In the end the gap between prices has actually increased.
The very point of an insurance is making high cost, low probability situations managable through the law of large numbers. Reintroducing individual factors is not needed per se, and also introduces both costs (you have to employ people assessing the risks/ building the models) and discrimination.
As public health care systems in other countries show skipping the risk management part can actually be cheaper for the insuree, overall.
Edit: This is also how the first insurances worked: as mutual insurance companies. In other words: This type of insurance actually was the foundation of the insurance industry.
I don't think that sounds too bad. There's very few things in modern life that bothers me more than insurance. Somehow humans used to get by, helping each other out when times were tough. I guess we're too bought in now to go back, but a healthy, single person can dream of not subsidizing teenagers livestreaming behind the wheel & overpriced medicine and unnecessary medical tests.
> Somehow humans used to get by, helping each other out when times were tough
Not a big fan of history?[0] I've not run the grep on that data, but there aren't many years that didn't have an armed conflict going on (in just Europe). If anything, humans have only recently started helping each other to really any lasting degree.
I'm talking about when communities used to help individuals when they had a fire, a bad harvest, lost someone. Helping each other out on a smaller scale. That along with the church (which is why they had tax exemption) used to be our safety net. People still fell threw the cracks if they were in the wrong community, but a lot could be said for how poorly the system works now.
I mean, the lens of history is firmly nearsighted on the aristocracy, so we don't have a lot of good data on the normal folk. But smaller communities were not any more or less nice than the larger ones.
People are still just people.
One somewhat clear example is the Huguenot riots of France after the Medieval warm period [0]. Towns would routinely tear themselves apart.
Another is actual Socrates. He was killed for teaching the youth 'bad' things, essentially [1].
I'm not saying that people did not help each other out in times of stress. But I am saying that they did fly apart and go all murdery too.
The past was not less rosy than the present, as far as we can tell. That lesson is very important, it means we have have power in our own lives and that our choice, your's and mine, have a lot of effects on those around us.
People used to get by without insurance because robust regulation is required to make an insurance industry viable. Unregulated insurance is a scam by default if there is no way to be sure anyone will pay out in the event.
I briefly worked at a insurance company. I sat through meetings where they actively strategized signing more people like myself up, and obfuscating the process for mothers and the elderly. I never looked at those companies the same way again.
A more realistic scenario is "why is my insurance more expensive than the new lower rate that everyone else is now getting? I used to pay the same amount as everybody else."
If the models improve, you get lower costs provided there is robust competition in the industry. That's why people let insurers drive in their cars now with them via an app to get lower rates.
That said - the main issue that creates the Kafkaesque world you speak of is poor correlation analysis. In the short term the insurer doesn't suffer due to false positives. Over time, however, to grow their market base you would expect that they start to test whether their analysis is actually reducing risk. With ML this presents a problem because they may not actually know why profiles are high at some point.
Denying financing is different than not lowering insurance rates, to be sure, and would also be looked at differently. I'm having trouble deciding where the slope starts getting slippery here as I hate the idea of some bad week of lawn care causing financial stresses for so many reasons - but insurers also already do all of this with a smaller dataset.
I wouldn't assume net rate increases - but I'm not sure if that should impact our overall thinking on the problem.
EDIT: > The paper abstract itself calls out where it gets slippery:
"From this perspective, public availability of house images raises legal and social concerns, as they can be a proxy of ethnicity, religion and other sensitive data."
> If your insurance company or bank is allowed to use all data it collects to make model of you, it leads to Kafkaesque world.
I reflexively agreed with what you wrote then I started thinking more about it, playing some likely scenarios through in my head, and now I'm not sure it necessarily ends in too bad of a place.
My reasoning: Ultimately, risk estimation has to actually reflect reality to be economically sustainable. Insurance is a highly competitive marketplace with low switching costs and relatively low barriers to entry. I've been driving for several decades and have only been involved in a couple of accidents, none of which were remotely attributable to me. I think I'm very low risk, yet I pay what seems to be fairly high premiums. I suspect low-risk drivers like myself are subsidizing higher-risk drivers due to the expense and inaccuracy of more accurately pricing forward risk.
So, taken as a general concept, more accurate risk assessment done more efficiently enables the actual risk component of insurance priced more fairly and higher efficiency actuarial estimation reduces the overhead cost of providing the overall insurance 'product' to a consumer. Broadly speaking, these are net positives. In terms of how it's done, I think that the market will react to pricing changes such that there will be insurance companies who promote the fact that they don't use XYZ risk estimation process, much as some Las Vegas casinos advertise that they only use single decks in blackjack vs six decks, increasing the appeal to card counters.
Counterargument: Imagine insurance company with perfect foresight.
Private insurance makes money by managing risk. But must be a limits of how they manage that risk in consumer markets.
Insurance is pooling the risk so that it's manageable. There is equilibrium. Trying to argue that insurance companies should be allowed to discriminate any way they can make profit is not a good idea.
> Counterargument: Imagine insurance company with perfect foresight.
To put too fine a point on it, imagine insurance is a risk-free business. The only reason to try to buy it is to see if you have a risk. If they reject you, plan your finances accordingly, I guess.
In other words, an image of my house shows run down, no paint, 3 trucks in the driveway - - - highest accident risk ?
But with health insurance, I am a vegetarian for more than 40 years, swim 4500 yards a week for 30+ years, 176 Lbs at 5'10" height, yet I have to pay relatively the same health premiums per age as the fat slobs that eat at fast food every meal . . . in effect I am paying extra for their increased medical needs . . .
Insurance will use whatever data they can get in an attempt to be more accurate. My address is plagued by a fire risk map. Companies will insure my neighbor, but not me. At some point a person or machine drew a live on a map and I crossed it.
A firefighter looked this location and concluded that if either house burns they will both likely burn. If the insurance companies believed him would they cover me or drop my neighbor? It is a crazy industry. Vegas anyone?
133 comments
[ 3.2 ms ] story [ 208 ms ] threadBreak out the champagne!
House appearance and car crashes will be related by risk-taking behavior, carelessness, and being too tired. For almost any two things in life, an inability to keep one in good order will most likely predict the other.
Also, I live in a somewhat decrepit house from around 1890, in a postcode area of no prestige whatsoever.
In 40 years of driving, I have not been in any collision, and have never filed an insurance claim.
The difference a claimant's personal situation & qualities makes to the claim is often via prejudices of the claims adjuster. Which is why we implement automated fraud detection to help identify those who otherwise appear "honest".
The usual explanation is that some areas have busier roads and busier roads are conducive to more accidents.
But busier roads also usually correlate with urban density which often correlates with socioeconomic status, race, crime, ...
Legally, this is a sort of "differential impact" discrimination, and iirc it's mostly legal as long as you can justify your reasoning. I think insurers would have to justify their rate adjustments though - a 5% extra premium for comprehensive at the intersection of Skid Row and Mugger's Bend is probably legit, but a 300% extra in Chinatown would land the insurer in legal hot water. Trial lawyers make big money running class-actions for these sorts of things, which hopefully keeps businesses honest.
(IANAL, I picked this up from my mom who does labor/employment defense.)
People with private locked garage will also have better ratings in general. Wealthier places will also invest more in road maintenance and limiting accident factors (for instance having a ton more protective equipment near school zones, investing in underground parking near shopping areas etc.)
In general I feel that trying to legally screen info from insurance is a lost proposition, as they have all the data in the world and find a proxy for that info anyway.
this doesn't help much in my experience. my insurance company doesn't even have a field to tell them whether you keep your car in a garage.
sometimes, sometimes not. some cities (like mine) vary a lot from neighborhood. I live in a nice one but share a zip code with a poor neighborhood with lots of crime. my premiums are pretty high despite the fact that I keep my car in a locked garage and only use it to hop directly on the freeway and drive out of the city.
I registered for the beta and like it a lot. I'm paying 30% less with no effort - just because I'm not driving as crazy as everyone else.
at least on paper, data protection is fine - is being processed by a separate company with a very restrictive privacy agreement, storing and handing over only a 1-10 integer for several driving skills (speeding, acceleration, hard breaking).
I had to drive 400km to get a score. Of course, I drove like my grandma was about to have a neck operation to get that score.
It'd be cool (from a hacker point of view) to spoof the smartphone's GPS/accelerometer data: put a delay of, say 10 seconds. So you can accelerate like mad off the red light, but the blackbox app receives a spoofed acceleration curve which is gentler. I guess it's harder if you like going over the speed limit, e.g. you'd arrive to your destination after driving 90 mph, but the app still has to spoof you being 20 miles away doing 70mph (the highway limit in the UK).
Besides, the legal requirements for denying coverage in Germany are pretty high.
If you're crashing with 130 in a 30 zone, you won't need digital evidence.
If there are two raters in a sample of six that disagree with the four others, I would argue that the IRR is an issue. Otherwise one could always run studies with n experts and only select a subsample of those raters that have a high interrater agreement, or not?
The Fleiss kappas are already relatively low..
Click on 'report a problem' in Street View and centre the viewer you will see on your house, select 'my home' from the options, and add the address for good measure in the comment area.
It's the 21st century equivalent of hiding under the blanket to protect from the evil spirits.
I generally agree. The whole point of insurance is to pool risk -- so cutting the population into ever-smaller pools seems missing the point. In the infinite limit we'd get seven billion single-person pools, each of which has that person pay in the exact same amount of money they get back out.
This is also known as "not having insurance".
The point of insurance is to replace a high variance distribution of potential losses with a low variance distribution, but with a similar expected loss. In other words, on average you pay a bit more, but you never pay a huge amount.
Now, in a perfect world, everyone would pay exactly what their expected loss is, plus a small premium. Since it's impossible to measure the expected loss exactly, in reality some people pay a bit more and some pay a bit less. Getting a better risk estimate doesn't get you closer to "not having insurance", but rather closer to "perfect insurance" where everyone gets a distribution that exactly matches their risk profile.
I still think this form of "perfect insurance" is bad - part of what I like about the insurance is exactly the dispersion of cost in the larger population. Should a chronically-ill person pay a very high insurance, or should the society cover the costs? I'm in the second camp, even though I'm not chronically-ill. It's part of the benefits of having an actual society - we don't need to stand alone.
The latter are necessarily the government's remit, and the political discussion on whether they are desirable can be had in an informed manner if costs are clear and they are not commingled with insurance elements.
what we call "health insurance" is really just a payment planning and collective bargaining service rolled into one. it doesn't really work like most other kinds of insurance.
The perfect insurance would know how likely it is that you are going to have to pay a catastrophically large amount, and would charge you amount*probability. If probability=1.0, then sure, they should charge you the whole amount.
As the data gathered increases, the probability eventually reaches 1.0. For instance, if an insurance company could legally gather all telemetry from your car at a high enough frequency and cancel your policy at will, then it could determine that your car was on an unavoidable collision course and eliminate losses.
This may sound like a silly thought experiment, but I think it demonstrates that in principle, there must be some point at which we draw the line on discrimination to have insurance, so if you find yourself defending all discrimination then you've gone wrong somewhere. Either there's a principle that has been missed, or it is valid to just set the rules in society that people are comfortable with.
There are, however, situations where the model is a reasonable fit - notably, some aspects of health care, where the approach to the limit proceeds over multiple billing periods. A society can choose to spread the cost around, but to do this under the guise of insurance leads to the sort of incongruities and inefficiencies that are seen in the US system.
Being prescient about getting the number 4 when you throw a dice means you can assign probability 1 to getting 4 when you throw that dice.
Having an accurate risk profile means knowing the odds you'll get 4 with that dice might be 1/6.4 because you precisely measured the shape and weight distribution of the dice and you can see how biased it is.
Let's assign a cost X for the unlucky event of rolling a 4. You're expected loss is X∙1/6.4. If the event only happens once, obviously you either pay full cost X or you don't pay anything.
If you pool with a large number (N) of people, each throwing once that same dice, then the expected loss of the pool is N∙X∙1/6.4, that is, if each member of the pool contributes their share, the pool is going to be able to pay off the individual losses. Unlucky people are better off because they saved X∙(1-1/6.4); lucky people are worse off because they had to pay X∙1/6.4. Pool members agree on those rules, because they don't know if they are going to win or lose, and their contribution to the pool is smaller than the worst case scenario.
So far so good. I deliberately used the "custom" probability number 1/6.4 to show how arbitrary that is.
Other people might have other dice, each biased their own way.
The expected loss for the pool would thus be X∙(P₁+P₂+P₃+....Pn).
Each person contributes to the pool backing proportionally to the contribution to the pooled expected loss, thus the system still makes sense even on the case of 1-sized buckets.
(Disclaimer: I have no idea what I'm talking about, this is not my field, just an average person with everyday math skills. Happy to learn why I'm wrong, in case I am)
(EDIT: formatting and typos)
There isn’t likely a perfect loss prediction to be had, e.g. predicting and understanding risk itself influences the likelihood of avoiding a loss favorably. An ideal scenario might be that people become aware of risky behavior then adjust and therefore change the future prediction. Health insurers do it all the time. Helping at risk customers avoid large losses actually helps the customer and the rest of the less risky customers in the pool (everyone’s premium should go down if the risky become less risky).
You say "Quite the opposite", and proceed to rephrase the same thing as GP said, except for the conclusion, which also is the same thing, yet you claim it isn't.
Not having insurance IS "perfect insurance", in that you pay exactly your own costs, with the added benefit of having no overhead. Most of us don't want that though, in the same way as we don't want children to have to pay their own healthcare, or chronically debilitated people to die to darwinism (whether we succeed here is another point). But a perfect "risk estimate" also happens to be the same thing as prescience, and if we had that we wouldn't need insurance for other reasons.
Not having insurance = paying for yourself = distribution by exact risk profile.
Not having insurance means losses follow your exact risk distribution, with a non-zero risk of a catastrophic loss. Having a perfect insurance means your losses follow a distribution with the same average as your risk profile suggests (+premium), but much lower variance and with zero density on the higher end, meaning you never have a catastrophic loss.
Quite the contrary.
> average
And that right there is where you are hiding all your assumptions. There is no such thing as "your risk profile", there is only your risk profile according to some risk model, which approaches what actually will happen to you individually the more detailed you make it and the better your predictive capabilities get, because as you do that, you are shrinking the pool of individuals that this average is calculcated over, approaching only you as one individual in the limit, at which point the average is the same as the actual thing.
> and with zero density on the higher end, meaning you never have a catastrophic loss.
Also, that's not a thing.
No, that is a fundamental misunderstanding of what odds are. Odds are not a property of a system, odds are a description of your knowledge about that system. There is no such thing as "odds of the system", the odds for the same system can be vastly different for different observers.
If you know how to control the coin, then the odds for you are almost 1:0 for your intended face. For the person next to you who doesn't know you, it's probably 0.5:0.5. For the person that has watched you for a while, it might be 0.6:0.4, if that's the ratio of the faces that you flip (intentionally). All for the exact same system at the exact same moment.
Also, a binary distinction between "having no clue" and "perfect predictions" is not useful. While perfect predictions are the limit of an ever refined model, there are many steps in between a useless model that gives random predictions and a model that makes no mistakes whatsoever.
To me, the aspect of redistribution is important in insurance. The acceptance that the world is chaotic and we can not predict all the bad luck that might happen to us. Hence, I pay for yours bad luck, if you pay for mine.
That is the same thing if you have produced a perfect risk profile. The perfect risk profile completely predicts reality, and all costs can be accounted for. If it doesn't completely predict reality, it's not perfect.
So with the perfect risk profile, insurance is a zero sum game. Instead of paying a small premium you can put your money in a bank account and instead have them pay you "a small premium" in the form of interest.
And no, I don't believe that there can ever be a perfect risk profile, but as risk assessment gets more fine grained, insurance approaches that zero sum game.
The utility of insurance is to reduce variance, it is not a zero-sum game.
Perhaps you understand the word "perfect" in a different sense than I do. A better risk estimate to me is (to me) one that is more accurate. A perfect risk estimate is one that can not be more accurate. If a better risk estimate brings you closer to perfect insurance (which the GP concludes), a perfect risk estimate should leave you as close to perfect insurance as is possible.
It would be cool if you could address my reasoning instead because now I have no idea what exactly you disagree with other than the conclusion.
Let's say you have a house worth $100,000 and statistically it burns down once in 1000 years. This means that the expected damage is $100 per year, but the variance is very high because once the house burns down, the damage is 1000 times greater than the expected damage. With insurance, however, the variance is zero because you pay the same amount every year, whether your house burns down that year or not.
I would argue that with a perfect insurance you would only pay the expected loss, i.e. $100 per year, and not more or less, because ideally the insurance has accurately assessed the expected loss and diversified the variance away. However, with less ideal risk assessment, you could pay more or less because other houses burn more or less often.
If the intent described were truly the case, explain Medicare and Medicaid.
The clean way to do it is to separate insurance (can be very efficient if privately provided) from redistribution (will lead to bad distortion unless done by the state).
You seem to be saying insurance companies must discriminate unfairly by saying that if insurance companies discriminate unfairly, they will lose customers.
Say I am male, and I pay more for car insurance because of that. Your argument is that if sex discrimination is illegal, then the market will collapse because women will stop buying insurance. But isn't it equally true that men who are low risk should "evaporate from the pool" in the present system? There doesn't seem to me a fundamental distinction between these two examples of discrimination.
Now, there may be a quantitative difference, a difference in economic impact, but if it's only quantitative, then why should one automatically be manageable and the other not, which is the assumption I think people implicitly make.
It's not actually a value judgement I'm giving. Whether it's unfair as in sex or race based is not part of the incentive mechanism. The mechanism still exists though. If you find out people of one color are bigger risks than another you have an awkward problem.
If you can restrict the companies from doing this you fix the problem, but of course there's games to be played regarding selection, as another commenter notes.
I'm using "unfair" to mean anything someone could call unfair, that is, charging someone as though they are a higher risk than they are, due to some factor.
The point is, it's illogical to say we can't* misprice risk based on ethics or social justice or whatever, when you assume we can misprice risk based on incomplete knowledge. It seems like special pleading.
*can't in the sense of it being a mortal threat to the industry in general
This doesn't seem less wrong than generalizing based on something like age or gender. I would attack it on privacy grounds, not accuracy.
As such we really need to reassess the usefulness of such organisations in our society - they're useful because they blunt the monetary pain of rare life events and pool the costs of chronic issues. If they stop covering people who need them the most, they stop being useful for us.
someone who can't afford their sky high car insurance premiums isn't someone who "needs it the most". they are someone who should be priced out of driving.
If they applied this to "black people are more likely to cost us X" then everybody would be against it. But if they apply it to different discriminators (age, gender, neighborhood), then suddenly it is ok?
Would it be better if the insurance agents visited the house themselves?
Consider ML model that picks 1,000 people for every 10,000 and 99% of those selected are false positives. Insurance fee is $100 and average payment is $50,000. Rejecting those 1,000 people is $100,000 less revenue and saving $500,000 in payments. It's no brainier to reject people and let there be statistical accidents.
Insurance companies can make profit using models that have huge number of false positives. If your insurance company or bank is allowed to use all data it collects to make model of you, it leads to Kafkaesque world.
"Why my insurance premium jumped 200% and I can't afford it anymore?" Nobody knows that the reason is your dried out flower in garage widow visible to the street.
"Why I can't get mortgage?" You have too many Facebook friends with risk of defaulting and one has criminal record. Select your friends better.
Would you lend a close friend $1k? Maybe, if they really needed it.
This is an example of you discriminating based on what you know about me/them. Should such discrimination be made illegal?
You don't have to make that kind of stuff illegal, but you also need to make sure that you don't base important parts of your economic systems on such biased processes.
I think it is a much more sensible arrangement that insurance companies are free to operate as they see fit and that the government be responsible for the well being of its poorest citizens.
Should grocery and clothing stores be required to feed and clothe the poor, respectively? To me, this feels like a shrinking of government duties onto businesses.
I guess you can go the route of government sponsored insurance (and everybody know governments hardly negotiate as they can print money) or you make laws and keep tight control over what insurance companies can and can't do.
Because drivers are required by law, in many jurisdictions, to have insurance.
This sounds like an argument for nationalized insurance, which may have been the point you were trying to make.
Was that the intention?
In reality it didn't work because there were many other variables that are correlated with gender such as profession or types of car driven. In the end the gap between prices has actually increased.
The very point of an insurance is making high cost, low probability situations managable through the law of large numbers. Reintroducing individual factors is not needed per se, and also introduces both costs (you have to employ people assessing the risks/ building the models) and discrimination.
As public health care systems in other countries show skipping the risk management part can actually be cheaper for the insuree, overall.
Edit: This is also how the first insurances worked: as mutual insurance companies. In other words: This type of insurance actually was the foundation of the insurance industry.
Not a big fan of history?[0] I've not run the grep on that data, but there aren't many years that didn't have an armed conflict going on (in just Europe). If anything, humans have only recently started helping each other to really any lasting degree.
[0] https://en.wikipedia.org/wiki/List_of_conflicts_in_Europe
People are still just people.
One somewhat clear example is the Huguenot riots of France after the Medieval warm period [0]. Towns would routinely tear themselves apart.
Another is actual Socrates. He was killed for teaching the youth 'bad' things, essentially [1].
I'm not saying that people did not help each other out in times of stress. But I am saying that they did fly apart and go all murdery too.
The past was not less rosy than the present, as far as we can tell. That lesson is very important, it means we have have power in our own lives and that our choice, your's and mine, have a lot of effects on those around us.
[0] https://en.wikipedia.org/wiki/Huguenot_rebellions
[1] https://en.wikipedia.org/wiki/Trial_of_Socrates
If the models improve, you get lower costs provided there is robust competition in the industry. That's why people let insurers drive in their cars now with them via an app to get lower rates.
That said - the main issue that creates the Kafkaesque world you speak of is poor correlation analysis. In the short term the insurer doesn't suffer due to false positives. Over time, however, to grow their market base you would expect that they start to test whether their analysis is actually reducing risk. With ML this presents a problem because they may not actually know why profiles are high at some point.
Denying financing is different than not lowering insurance rates, to be sure, and would also be looked at differently. I'm having trouble deciding where the slope starts getting slippery here as I hate the idea of some bad week of lawn care causing financial stresses for so many reasons - but insurers also already do all of this with a smaller dataset.
I wouldn't assume net rate increases - but I'm not sure if that should impact our overall thinking on the problem.
EDIT: > The paper abstract itself calls out where it gets slippery: "From this perspective, public availability of house images raises legal and social concerns, as they can be a proxy of ethnicity, religion and other sensitive data."
I reflexively agreed with what you wrote then I started thinking more about it, playing some likely scenarios through in my head, and now I'm not sure it necessarily ends in too bad of a place.
My reasoning: Ultimately, risk estimation has to actually reflect reality to be economically sustainable. Insurance is a highly competitive marketplace with low switching costs and relatively low barriers to entry. I've been driving for several decades and have only been involved in a couple of accidents, none of which were remotely attributable to me. I think I'm very low risk, yet I pay what seems to be fairly high premiums. I suspect low-risk drivers like myself are subsidizing higher-risk drivers due to the expense and inaccuracy of more accurately pricing forward risk.
So, taken as a general concept, more accurate risk assessment done more efficiently enables the actual risk component of insurance priced more fairly and higher efficiency actuarial estimation reduces the overhead cost of providing the overall insurance 'product' to a consumer. Broadly speaking, these are net positives. In terms of how it's done, I think that the market will react to pricing changes such that there will be insurance companies who promote the fact that they don't use XYZ risk estimation process, much as some Las Vegas casinos advertise that they only use single decks in blackjack vs six decks, increasing the appeal to card counters.
Private insurance makes money by managing risk. But must be a limits of how they manage that risk in consumer markets.
Insurance is pooling the risk so that it's manageable. There is equilibrium. Trying to argue that insurance companies should be allowed to discriminate any way they can make profit is not a good idea.
To put too fine a point on it, imagine insurance is a risk-free business. The only reason to try to buy it is to see if you have a risk. If they reject you, plan your finances accordingly, I guess.
But with health insurance, I am a vegetarian for more than 40 years, swim 4500 yards a week for 30+ years, 176 Lbs at 5'10" height, yet I have to pay relatively the same health premiums per age as the fat slobs that eat at fast food every meal . . . in effect I am paying extra for their increased medical needs . . .
A firefighter looked this location and concluded that if either house burns they will both likely burn. If the insurance companies believed him would they cover me or drop my neighbor? It is a crazy industry. Vegas anyone?
In US, home insurance companies already have such info in detail. Most also sell auto coverage.
State regulators might require a "Chinese wall" between the two operations, but effective enforcement would be another matter.