Re:The frustration of trying to invest in my hometown
by AnthonyMouse on Monday August 03, 2020 @08:33AM
(#24036879)
> If a specific statistical classification indicates more bankruptcy, then it becomes true.
Fortunately this isn't actually the chicken and egg problem you think it is.
Suppose that for defunct historical reasons the model is predicting a higher default rate for some population, so it says to approve 85% of their loan applications rather than 95%. Then half of those people get into trouble that getting a second loan would save them from, but because of the model, only 85% rather than 95% are given the other loan.
So then their repayment rate will be 90% rather than 95%, because not as many were saved by the loan, but wait -- 90% is more than 85%, because only half of them needed the second loan. So now the model starts predicting that it should expect a 90% repayment rate for that population, which then increases their actual repayment rate, and so on until it stabilizes at 95%.
If you could identify the exact degree to which the prediction is affecting itself then you could take it all out at once, but doing that is hard. If the evidence shows a lower repayment rate in a particular area, is that just the holdover, or is it because that area still has a present-day higher level of drug use and violent crime which negatively affects repayment? If it's the latter and you give them loans at scale under the assumption that it's just an artifact, the bank is going bust.
So unless you can disentangle it all, the data to do which is often not reasonably available, the only realistic option is waiting for it to filter through.
Fortunately this isn't actually the chicken and egg problem you think it is.
Suppose that for defunct historical reasons the model is predicting a higher default rate for some population, so it says to approve 85% of their loan applications rather than 95%. Then half of those people get into trouble that getting a second loan would save them from, but because of the model, only 85% rather than 95% are given the other loan.
So then their repayment rate will be 90% rather than 95%, because not as many were saved by the loan, but wait -- 90% is more than 85%, because only half of them needed the second loan. So now the model starts predicting that it should expect a 90% repayment rate for that population, which then increases their actual repayment rate, and so on until it stabilizes at 95%.
If you could identify the exact degree to which the prediction is affecting itself then you could take it all out at once, but doing that is hard. If the evidence shows a lower repayment rate in a particular area, is that just the holdover, or is it because that area still has a present-day higher level of drug use and violent crime which negatively affects repayment? If it's the latter and you give them loans at scale under the assumption that it's just an artifact, the bank is going bust.
So unless you can disentangle it all, the data to do which is often not reasonably available, the only realistic option is waiting for it to filter through.