I don't see nearly enough complexity in this analysis to justify the claim of having found any insurance fraud.
Firstly, there's no account for correlation between the features identified. The article mentions VINs which have several single-vehicle accidents, for example, but someone who has one single-vehicle accident is probably more likely to have another. Switching coverage is another of those potentially-correlated features; if you claim and it bumps your premium, aren't you likely to shop around as a result?
Secondly, there's no attempt to account for the law of large numbers. It's incredibly unlikely that someone has three single vehicle accidents in a year, but because the probability of that is nonzero, we know that with enough vehicles on the road then someone is going to do it.
The article covers itself by acknowledging this, of course, but if you title your blog post "We Found Insurance Fraud in Our Crash Data" then you should actually do that.
Also, anyone who lives in an area where there are deer (or, kangaroos in australia) will tell you - strange sounding single car crashes late at night or early in the morning.. aren't that strange, that's when the deer (or kangaroos) are out and about. Where I grew up if someone had hit a kangaroo 3 times in a year it would be strange sounding for sure, but not that strange.
My Mother got a letter saying "if you hit another deer we will cancel your insurance", because she hit 3 deer in one year. Those 3 are the only deer she has hit in > 60 years of driving. Random events cluster.
Or just snow. I'm pretty sure every doordash driver hits a dozen things sliding down hills in the winter around here. If you get lucky it doesn't really matter because the tree or rock isn't gonna care and the damage is cosmetic. If you aren't lucky whoever's mailbox or fence you hit files a claim.
yeah i didn't even think about doordash or uber drivers. A whole extra group of cars/people who will be wild outliers in driving stats even if they have the exact same skills.
> The article mentions VINs which have several single-vehicle accidents, for example, but someone who has one single-vehicle accident is probably more likely to have another.
I haven't followed this story to know whether she's still driving, but one driver was involved in 7 crashes in 4 years, including two fatalities.
>Anderson received the maximum sentence by Judge Charles Miller of three years, laid out in the plea agreement.
>Under the plea agreement, Anderson’s license will be suspended for six years, the maximum under a Level 5 felony, and she will also be named a habitual traffic offender, which will suspend her license for 10 years. The license suspensions will run concurrently.
This is a marketing blog for some AI startup trying to sell AI voodoo to the insurance industry, so it just needs to sound like it might be useful for finding fraud.
I remember when some game developers started restricting installations of their game. There was an outcry, but then they said it was something like 100 installs in a certain period. It became clear their outliers were probably reasonable to restrict.
Maybe just investigating the outliers of the crashes could legitimately put them onto something.
Now it's time for the devil's advocate-- if the title was "some people get in a lot of accidents", would the top comment be "actually it's more likely that insurance fraud is the answer"
That doesn't really matter. I mean, sure, I could look at this data and say "hmm, looks like insurance fraud to me."
But the point being, I really don't have evidence one way or the other. Implying insurance fraud is saying you think a criminal act occurred, and when you say something like that the burden should be on you to provide some evidence to that effect.
Random internet commenters bullshit all the time. But this blog post provides specific data and an unambiguous conclusion that fraud occurred (the title does not hedge, it is "We Found Insurance Fraud in Our Crash Data"), so if you do something like that you need to provide evidence for it. I appreciate the parent commenter for pointing out that, despite all of their data, they didn't really provide much evidence for their conclusion.
Exactly this. When I open a post like this I get excited because I think I'm either about to learn something interesting about data analysis, or I'm about to see a neat application of a technique that I already know about and it's going to make me happy.
When the post actually boils down to "hey if you look at a large dataset then you'll find unusual events in it, now visit our AI startup!" then that irritates me enough to whine about it in the comments.
i have absolutely no problem with that, i suspect that no matter the title, whoever has a contrarian opinion to it will be the one to speak up while the rest remain silent
Firstly, there's no account for correlation between the features identified. The article mentions VINs which have several single-vehicle accidents, for example, but someone who has one single-vehicle accident is probably more likely to have another. Switching coverage is another of those potentially-correlated features; if you claim and it bumps your premium, aren't you likely to shop around as a result?
Secondly, there's no attempt to account for the law of large numbers. It's incredibly unlikely that someone has three single vehicle accidents in a year, but because the probability of that is nonzero, we know that with enough vehicles on the road then someone is going to do it.
The article covers itself by acknowledging this, of course, but if you title your blog post "We Found Insurance Fraud in Our Crash Data" then you should actually do that.