Any of the non-discriminatory solution generates less profit than the max-profit option.
So effectively, a bank is paying to be politically correct i.e. non-discriminatory.
And after all, if an orange person is not given a loan even though he is personally more likely to pay it back than a compared blue person, he only has other orange persons to blame, for dragging down his group score.
While it sucks to be discriminated against because other members of your group dragged down your ‘group score’, it’s also equally unfair to ask other people, who were not responsible in anyway for the underlying reasons for why discrimination is a logical solution, to pay by pretending that those reasons do not exist.
You are not asking for a solution to the cause, only for a clever way to dress up the end result to make it look like there was no cause at all.
Note that in this simulation orange people and blue people are equally likely to pay back. What is different is distribution of credit score. Credit score is approximate proxy. It seems weird for me to blame other people for having bad credit score, when they are not worse paying back! Isn’t the bank to blame for having a bad system of credit score, which does not match actual paying back?
Think of this way: there can be equally profitable alternative system of credit score, where distribution of credit score of orange and blue people are reversed. If you don’t know which system of credit score will be used, it makes sense to advocate the system which uses credit score in a way that tracks underlying reality, instead of proxy. If you know system of credit score used will score you better than reality (“being privileged”), it may not. But I think the later position fails the test of veil of ignorance.
Note that in this simulation orange people and blue people are equally likely to pay back.
Even then, the distribution means that if you only use the credit score to decide who to give a loan to, it is more profitable to give a group more loan than the other. We wouldn’t have the omniscient view that we have here.
it seems weird for me to blame other people for having bad credit score
The problem is that if you are blue, you might end up in a situation, where people with the same credit score as you, are less likely to pay back, compared to orange people. So the problem remains, if you don’t look at blue people as a whole, but blue people in your score group. You can’t blame a bank for not wanting to give you a loan when they know that a blue person with 50 score is 20% to pay back, while an orange person with 50 score is 50% to pay back. To ask for no discrimination here is to tell people to put their head in the sand and to pretend things that are not equal are equal.
Think of this way: there can be equally profitable alternative system of credit score, where distribution of credit score of orange and blue people are reversed.
Isn’t the problem presented in the article not with alternative systems of credit score, but that given a system of credit score, how to set a threshold that will fulfill political requirement i.e. enforced ‘equality’?
Credit scoring model is a machine learning model. What is going on is the model is more predictive for orange people than blue people. In max profit strategy, cost of this imperfection is born by blue people. In equal opportunity strategy, cost is born by bank.
The paper’s argument is that it makes more sense for bank to bear the cost, because bank can better pay credit scoring company to improve its model. In theory, blue people too can raise fund and pay credit scoring company to improve its model for themselves, but due to coordination problem it is impractical. Note that as soon as credit scoring model achieves accuracy parity for demographic groups, equal opportunity strategy is max profit strategy, unrelated to whether which group defaults more, which group gets loan more, etc.
My other comment was written from the point of view of bank customers. So here’s the case for banks.
You are a bank using max profit loan strategy, profit 32400, 12100 from blue people and 20300 from orange people. Your competitor used the same loan strategy in the past, but they switched to equal opportunity loan strategy and are aggresively marketing it. You lose half blue customers to the competitor, profit 26350, 6050 from blue people and 20300 from orange people. Equal opportunity loan strategy had profit 30400, 11700 from blue people and 18700 from orange people. Now it has profit 36250, 17550 from blue people and 18700 from orange people. The end.
The bank is paying 2000 in profit to avoid 6050 in profit from loss of customers. The bank is not paying to be politically correct. This suggests if the chance of such competition is less than 33%, the bank would not implement the change.
Any of the non-discriminatory solution generates less profit than the max-profit option.
So effectively, a bank is paying to be politically correct i.e. non-discriminatory.
And after all, if an orange person is not given a loan even though he is personally more likely to pay it back than a compared blue person, he only has other orange persons to blame, for dragging down his group score.
While it sucks to be discriminated against because other members of your group dragged down your ‘group score’, it’s also equally unfair to ask other people, who were not responsible in anyway for the underlying reasons for why discrimination is a logical solution, to pay by pretending that those reasons do not exist.
You are not asking for a solution to the cause, only for a clever way to dress up the end result to make it look like there was no cause at all.
Note that in this simulation orange people and blue people are equally likely to pay back. What is different is distribution of credit score. Credit score is approximate proxy. It seems weird for me to blame other people for having bad credit score, when they are not worse paying back! Isn’t the bank to blame for having a bad system of credit score, which does not match actual paying back?
Think of this way: there can be equally profitable alternative system of credit score, where distribution of credit score of orange and blue people are reversed. If you don’t know which system of credit score will be used, it makes sense to advocate the system which uses credit score in a way that tracks underlying reality, instead of proxy. If you know system of credit score used will score you better than reality (“being privileged”), it may not. But I think the later position fails the test of veil of ignorance.
Even then, the distribution means that if you only use the credit score to decide who to give a loan to, it is more profitable to give a group more loan than the other. We wouldn’t have the omniscient view that we have here.
The problem is that if you are blue, you might end up in a situation, where people with the same credit score as you, are less likely to pay back, compared to orange people. So the problem remains, if you don’t look at blue people as a whole, but blue people in your score group. You can’t blame a bank for not wanting to give you a loan when they know that a blue person with 50 score is 20% to pay back, while an orange person with 50 score is 50% to pay back. To ask for no discrimination here is to tell people to put their head in the sand and to pretend things that are not equal are equal.
Isn’t the problem presented in the article not with alternative systems of credit score, but that given a system of credit score, how to set a threshold that will fulfill political requirement i.e. enforced ‘equality’?
Credit scoring model is a machine learning model. What is going on is the model is more predictive for orange people than blue people. In max profit strategy, cost of this imperfection is born by blue people. In equal opportunity strategy, cost is born by bank.
The paper’s argument is that it makes more sense for bank to bear the cost, because bank can better pay credit scoring company to improve its model. In theory, blue people too can raise fund and pay credit scoring company to improve its model for themselves, but due to coordination problem it is impractical. Note that as soon as credit scoring model achieves accuracy parity for demographic groups, equal opportunity strategy is max profit strategy, unrelated to whether which group defaults more, which group gets loan more, etc.
My other comment was written from the point of view of bank customers. So here’s the case for banks.
You are a bank using max profit loan strategy, profit 32400, 12100 from blue people and 20300 from orange people. Your competitor used the same loan strategy in the past, but they switched to equal opportunity loan strategy and are aggresively marketing it. You lose half blue customers to the competitor, profit 26350, 6050 from blue people and 20300 from orange people. Equal opportunity loan strategy had profit 30400, 11700 from blue people and 18700 from orange people. Now it has profit 36250, 17550 from blue people and 18700 from orange people. The end.
The bank is paying 2000 in profit to avoid 6050 in profit from loss of customers. The bank is not paying to be politically correct. This suggests if the chance of such competition is less than 33%, the bank would not implement the change.
Perhaps a more apt title would be “Automate discrimination with smarter machine learning.”