3 ms·

> In your example that would be preemptive imprisonment if there was a greater than some probability you were a criminal. I don't like the idea of preemptive im

> In your example that would be preemptive imprisonment if there was a greater than some probability you were a criminal. I don't like the idea of preemptive imprisonment, but if that was the case, then I don't see how adding more data to make the predictions more accurate would be a bad thing. That position is utterly untenable. It's so obviously wrong to imprison someone because some signals indicate they might commit a crime based on who they're friends with on Facebook! There's a difference between being buddies with a gaggle of career criminals and actually committing (or conspiring to commit) a crime! > If there was a 99% chance of you becoming addicted to them, then you probably shouldn't be given pain killers. If that person has a history of abusing painkillers, yes. If their friends do? C'mon, you can't possibly justify that position. > There is no rational reason that the world would be better if we just gave loans to everyone at the same interest rate I didn't argue for equal interest rates for all. It makes sense for interest rates are tied to risk, but it's only reasonable to make an assessment of risk based on an individual's performance and ability to pay, not their friends' or peers' performance. > or if we intentionally excluded useful data points from determining the risk of things. I'm not sure whether you've thought this through. Let's say, for example, it's determined by someone's payment history that they're not good at managing their money. They learn from their mistakes, improve their payment behavior, and subsequently lower their risk of default. On the other hand, what if it was determined that Asians are more likely to default on their debt. What can you do to stop being Asian? What if gay people were less likely to default on their debt and then irresponsible people who happen to be gay find it easier to get debt? Your thesis breaks down very quickly and I truly struggle to believe you agree with the position you're making. Neither one of us are making novel arguments now, but thanks for the interesting perspective and debate!

Houshalter 13y ago
>That position is utterly untenable. It's so obviously wrong to imprison someone because some signals indicate they might commit a crime based on who they're friends with on Facebook! There's a difference between being buddies with a gaggle of career criminals and actually committing (or conspiring to commit) a crime! I think you are trying to create a straw-man here. Imprisoning people even though they haven't committed a crime is obviously bad, and I completely agree with that. All I'm saying is if we were going to do it, then we might as well try to make the predictions as accurate as possible. If your predictions are more accurate fewer "innocent" people will end up in prison (innocent as in, wouldn't commit a crime if left out of prison.) In your example, less accurate predictions can literally ruin people's lives. Excluding data in that case would be much worse than just for loans. >> If there was a 99% chance of you becoming addicted to them, then you probably shouldn't be given pain killers. >If that person has a history of abusing painkillers, yes. If their friends do? C'mon, you can't possibly justify that position. I don't see why it matters that the data came from who your friends were. It could just as easily come from blood tests or who your parents are, or all sorts of arbitrary things outside of your control. The point is 99% of people with that feature become addicted to the painkillers and it ruined their lives. If there is a 99% chance of that happening to you, would you take the pain killers? Would you really care where the data came from Would you really ignore information if it was your own life at stake? >I didn't argue for equal interest rates for all. It makes sense for interest rates are tied to risk, but it's only reasonable to make an assessment of risk based on an individual's performance and ability to pay, not their friends' or peers' performance. But assessing risk at all puts some people at a disadvantage. It doesn't matter where the data came from, if the predictions aren't 100% accurate then some percent of people are going to wrongly have a higher interest rate, even if they pay back their loan on time. I think you are thinking of this as an incentive system and not a predicting system. That is it's ok to select for features that people can control, like their past history, but not for things that are outside their control. But the point of interest isn't to reward or punish people for paying or not paying their previous loans back on time. It's just to make up for people that don't pay. The riskier you are, the more it costs to make up for all the other high risk people that didn't pay. Ideally you just wouldn't give out loans to people who wouldn't pay them back, and would give loans with no interest to people who would. But usually it falls somewhere in between 0 and 100% probability. The goal isn't to reward or punish people, it's to predict as accurately as possible.

Houshalter 13y ago

>That position is utterly untenable. It's so obviously wrong to imprison someone because some signals indicate they might commit a crime based on who they're friends with on Facebook! There's a difference between being buddies with a gaggle of career criminals and actually committing (or conspiring to commit) a crime! I think you are trying to create a straw-man here. Imprisoning people even though they haven't committed a crime is obviously bad, and I completely agree with that. All I'm saying is if we were going to do it, then we might as well try to make the predictions as accurate as possible. If your predictions are more accurate fewer "innocent" people will end up in prison (innocent as in, wouldn't commit a crime if left out of prison.) In your example, less accurate predictions can literally ruin people's lives. Excluding data in that case would be much worse than just for loans. >> If there was a 99% chance of you becoming addicted to them, then you probably shouldn't be given pain killers. >If that person has a history of abusing painkillers, yes. If their friends do? C'mon, you can't possibly justify that position. I don't see why it matters that the data came from who your friends were. It could just as easily come from blood tests or who your parents are, or all sorts of arbitrary things outside of your control. The point is 99% of people with that feature become addicted to the painkillers and it ruined their lives. If there is a 99% chance of that happening to you, would you take the pain killers? Would you really care where the data came from Would you really ignore information if it was your own life at stake? >I didn't argue for equal interest rates for all. It makes sense for interest rates are tied to risk, but it's only reasonable to make an assessment of risk based on an individual's performance and ability to pay, not their friends' or peers' performance. But assessing risk at all puts some people at a disadvantage. It doesn't matter where the data came from, if the predictions aren't 100% accurate then some percent of people are going to wrongly have a higher interest rate, even if they pay back their loan on time. I think you are thinking of this as an incentive system and not a predicting system. That is it's ok to select for features that people can control, like their past history, but not for things that are outside their control. But the point of interest isn't to reward or punish people for paying or not paying their previous loans back on time. It's just to make up for people that don't pay. The riskier you are, the more it costs to make up for all the other high risk people that didn't pay. Ideally you just wouldn't give out loans to people who wouldn't pay them back, and would give loans with no interest to people who would. But usually it falls somewhere in between 0 and 100% probability. The goal isn't to reward or punish people, it's to predict as accurately as possible.