Correct me if I'm missing something, but I can mainly see two ways default prediction could be useful. (1) Build a forecast system that customers can use to simulate their chances of defaulting, and provide advice to help them avoid defaulting, and (2) predict if a customer is at risk of defaulting in the future and make decisions based on this (such as refusing certain services based on that).
\n(1) is great, whereas (2) is an open question as there's many unknown variables into play here, but I think that beyond building models that can be used for this, we (the kaggle community) should also discuss about the impact of our work. So here's my question: if we have accurate models that can predict future default, is it ethical to use it to minimize risk at the cost of customers' ability to use financial services provided by companies like American Express?
\nOn one hand, such model could lead to less defaults, and thus less expenses/losses for Amex, and this saving could be passed onto the customer under the form of lower interesting rates, lower annual fees, and higher cashback rewards. Another argument in favor: if we assume that such models are already being used, building more accurate models could result in lower errors (erroneously predicting a custom will default), which is good for customers that would have otherwise been denied certain financial services.
\nOn the other hand, more conservative use of risk modeling could negatively impact the life of certain customers. Let's say a customer just lost their job, and they need to pay for expensive medicine or treatments without insurance. Their financial situation is not great, so they need to use financial services that are decided by such models. What if the model predicts a 25% default chances? Number wise, that is a big risk for the company owning the model, so they would be incentivized in refusing those services. As a result, the customer might have to make difficult decisions that could eventually lead to them defaulting, which becomes a self-fulfilling prophecy, and is added as another data point confirming the model's correct prediction. Of course, that might be less relevant in some countries than others, but remains a reality.
\nWith those two sides presented, how would you answer the question?
", "messages": [ { "id": 1826123, "postDate": "2022-06-20T04:24:28.733Z", "content": "Correct me if I'm missing something, but I can mainly see two ways default prediction could be useful. (1) Build a forecast system that customers can use to simulate their chances of defaulting, and provide advice to help them avoid defaulting, and (2) predict if a customer is at risk of defaulting in the future and make decisions based on this (such as refusing certain services based on that).
\n(1) is great, whereas (2) is an open question as there's many unknown variables into play here, but I think that beyond building models that can be used for this, we (the kaggle community) should also discuss about the impact of our work. So here's my question: if we have accurate models that can predict future default, is it ethical to use it to minimize risk at the cost of customers' ability to use financial services provided by companies like American Express?
\nOn one hand, such model could lead to less defaults, and thus less expenses/losses for Amex, and this saving could be passed onto the customer under the form of lower interesting rates, lower annual fees, and higher cashback rewards. Another argument in favor: if we assume that such models are already being used, building more accurate models could result in lower errors (erroneously predicting a custom will default), which is good for customers that would have otherwise been denied certain financial services.
\nOn the other hand, more conservative use of risk modeling could negatively impact the life of certain customers. Let's say a customer just lost their job, and they need to pay for expensive medicine or treatments without insurance. Their financial situation is not great, so they need to use financial services that are decided by such models. What if the model predicts a 25% default chances? Number wise, that is a big risk for the company owning the model, so they would be incentivized in refusing those services. As a result, the customer might have to make difficult decisions that could eventually lead to them defaulting, which becomes a self-fulfilling prophecy, and is added as another data point confirming the model's correct prediction. Of course, that might be less relevant in some countries than others, but remains a reality.
\nWith those two sides presented, how would you answer the question?
", "rawMarkdown": "Correct me if I'm missing something, but I can mainly see two ways default prediction could be useful. (1) Build a forecast system that customers can use to simulate their chances of defaulting, and provide advice to help them avoid defaulting, and (2) predict if a customer is at risk of defaulting in the future and make decisions based on this (such as refusing certain services based on that). \n\n(1) is great, whereas (2) is an open question as there's many unknown variables into play here, but I think that beyond building models that can be used for this, we (the kaggle community) should also discuss about the impact of our work. So here's my question: **if we have accurate models that can predict future default, is it ethical to use it to minimize risk at the cost of customers' ability to use financial services provided by companies like American Express?**\n\nOn one hand, such model could lead to less defaults, and thus less expenses/losses for Amex, and this saving could be passed onto the customer under the form of lower interesting rates, lower annual fees, and higher cashback rewards. Another argument in favor: if we assume that such models are already being used, building more accurate models could result in lower errors (erroneously predicting a custom will default), which is good for customers that would have otherwise been denied certain financial services.\n\nOn the other hand, more conservative use of risk modeling could negatively impact the life of certain customers. Let's say a customer just lost their job, and they need to pay for expensive medicine or treatments without insurance. Their financial situation is not great, so they need to use financial services that are decided by such models. What if the model predicts a 25% default chances? Number wise, that is a big risk for the company owning the model, so they would be incentivized in refusing those services. As a result, the customer might have to make difficult decisions that could eventually lead to them defaulting, which becomes a self-fulfilling prophecy, and is added as another data point confirming the model's correct prediction. Of course, that might be less relevant in some countries than others, but remains a reality.\n\nWith those two sides presented, how would you answer the question?", "votes": 21 }, { "id": 1826852, "postDate": "2022-06-20T16:43:41.740Z", "content": "Economics of Information tries to answer this type of question. I suggest reading the paper The Market for Lemons: Quality Uncertainty and the Market Mechanism, by George Akerlof.
\nBy allowing default customers in, you increase costs for the other \"good\" customers. If you do not discriminate customers, you will be charging an average. Which would be unfair for good customers and cheap for bad customers. This may create a perverse cycle where good customers are driven away by the unfair price, making costs higher, which drives prices up again; until only bad customers stay. Insurance Companies have this dilemma.
\nAn accurate model allows you to reduce prices for good customers. You could offer a different service/price for risky customers. Amex could offer loans to default customers if they pay for a default-insurance or something similar. That way Amex doesn´t lose margin, and all customers get access to their products, at a price that reflects their risk level.
", "rawMarkdown": "Economics of Information tries to answer this type of question. I suggest reading the paper The Market for Lemons: Quality Uncertainty and the Market Mechanism, by George Akerlof.\n\nBy allowing default customers in, you increase costs for the other \"good\" customers. If you do not discriminate customers, you will be charging an average. Which would be unfair for good customers and cheap for bad customers. This may create a perverse cycle where good customers are driven away by the unfair price, making costs higher, which drives prices up again; until only bad customers stay. Insurance Companies have this dilemma.\n\nAn accurate model allows you to reduce prices for good customers. You could offer a different service/price for risky customers. Amex could offer loans to default customers if they pay for a default-insurance or something similar. That way Amex doesn´t lose margin, and all customers get access to their products, at a price that reflects their risk level.", "votes": 8 }, { "id": 1826425, "postDate": "2022-06-20T10:11:41.533Z", "content": "There are different answers… (in bold the two elements that directly answer your ethical concerns):
\n1) We are mostly dealing with existing risk - that is already existing customers. AMEX needs models for those to manage their accounts. For exemple if you evaluate an important risk you can provide a relief program to help the customer (and get more money back). This is generally good for distressed clients.
\n2) Evaluating the Risk for existing customer is important for economical and regulatory purposes too. You need to have an evaluation of you portfolio and set capital aside. The set up of the comp. seems to match this regulatory purpose rather than allowing/disallowing clients demand approach. Having an healthy portfolio has positive impact for everyone involved (basically not defaulting yourself).
\n3) The problem of allowing a demand is more complex. We are working with a binary target in a supervised framework. But in real life the problem of allowing/refusing customer demands is slightly different. This is usually not about refusing a customer but setting an interest rate that match their risk. The problem get more complex as you have to take into account any action you offer (relief program - new loan and even the interest rate you want to set) in evaluating the financial health after the decision.
\n4) For the specific AMEX case, given their business (high end travel) and their business model (high fees, high interest rate - >10%, low default rate) they are not geared towards high risk clients and they are probably not refusing that much clients. (fees act as a barrier for high risk clients). Looking at their disclosure it seems AMEX problem is more about the cost of acquisition (advertising, partnership).
\n5) You are right that in general there need to be some regulation to ensure that everyone has minimal access to a bank / minimal financial services. I think in US this is the job of Consumer Financial Protection Bureau. To answer your question more precisely, regulators will usually set a cap for interest rate it is possible to charge (say 15%) thus it is implicitly ethical - even legal - to refuse someone with a risk of default higher than this. Clients that get too much refusals might then ask regulators to designate a bank that will provide minimal services.
\n6) You are also right that more discriminative ML models would remove an important social aspect of banking (sharing of the risk), that can lead to increasing inequalities. The impact should be evaluated in terms of people loosing access to service in a complete framework (with CFPB enforcing the right to have an account). However there is also the ethical aspect of not overcharging healthy clients to compensate for bad ones (sometimes bad on purpose). Lowering interest rates for healthy people / company will generally have a positive impact for the economy (more consumption, more projects, more jobs … etc.) - and more business for you as you get more competitive.
\nSo there is a regulatory framework to refuse clients with too much risk and some very postivie ethical aspects in doing so.
", "rawMarkdown": "There are different answers… (in bold the two elements that directly answer your ethical concerns):\n\n1) We are mostly dealing with existing risk - that is already existing customers. AMEX needs models for those to manage their accounts. For exemple if you evaluate an important risk you can provide a relief program to help the customer (and get more money back). This is generally good for distressed clients.\n\n2) Evaluating the Risk for existing customer is important for economical and regulatory purposes too. You need to have an evaluation of you portfolio and set capital aside. The set up of the comp. seems to match this regulatory purpose rather than allowing/disallowing clients demand approach. Having an healthy portfolio has positive impact for everyone involved (basically not defaulting yourself).\n\n3) The problem of allowing a demand is more complex. We are working with a binary target in a supervised framework. But in real life the problem of allowing/refusing customer demands is slightly different. **This is usually not about refusing a customer but setting an interest rate that match their risk.** The problem get more complex as you have to take into account any action you offer (relief program - new loan and even the interest rate you want to set) in evaluating the financial health after the decision.\n\n4) For the specific AMEX case, given their business (high end travel) and their business model (high fees, high interest rate - >10%, low default rate) they are not geared towards high risk clients and they are probably not refusing that much clients. (fees act as a barrier for high risk clients). Looking at their disclosure it seems AMEX problem is more about the cost of acquisition (advertising, partnership).\n\n5) You are right that in general there need to be some regulation to ensure that everyone has minimal access to a bank / minimal financial services. I think in US this is the job of Consumer Financial Protection Bureau. **To answer your question more precisely, regulators will usually set a cap for interest rate it is possible to charge (say 15%) thus it is implicitly ethical - even legal - to refuse someone with a risk of default higher than this.** Clients that get too much refusals might then ask regulators to designate a bank that will provide minimal services. \n\n6) You are also right that more discriminative ML models would remove an important social aspect of banking (sharing of the risk), that can lead to increasing inequalities. The impact should be evaluated in terms of people loosing access to service in a complete framework (with CFPB enforcing the right to have an account). However there is also the ethical aspect of not overcharging healthy clients to compensate for bad ones (sometimes bad on purpose). Lowering interest rates for healthy people / company will generally have a positive impact for the economy (more consumption, more projects, more jobs … etc.) - and more business for you as you get more competitive.\n\nSo there is a regulatory framework to refuse clients with too much risk and some very postivie ethical aspects in doing so.", "votes": 8, "replies": [ { "id": 1827057, "postDate": "2022-06-20T19:24:59.680Z", "content": "Thank you for the extensive answer Lucas! This was very insightful and I'm glad you took the time to share your expertise in this matter :)
", "rawMarkdown": "Thank you for the extensive answer Lucas! This was very insightful and I'm glad you took the time to share your expertise in this matter :)", "votes": 1 } ] }, { "id": 1826205, "postDate": "2022-06-20T06:36:33.917Z", "content": "Yes, of course it is ethical for a private person or company to refuse service to a customer based on predicted default.
\nIt is ethical because everything else would not be sustainable: If you (as a private person or as a credit card company) too often lend money to customers who don't give it back, sooner or later your money will be used up and you'll be out of business.
\nConsider as well that the bank doesn't only lend its own money to the credit card customers, but other people's money as well (i.e. the money of the bank's creditors). If people put their hard-earned money into a savings account, the bank has the duty to ensure that the money is safe. From this perspective, it would be unethical to lend money to customers with high predicted default risk.
", "rawMarkdown": "Yes, of course it is ethical for a private person or company to refuse service to a customer based on predicted default.\n\nIt is ethical because everything else would not be sustainable: If you (as a private person or as a credit card company) too often lend money to customers who don't give it back, sooner or later your money will be used up and you'll be out of business.\n\nConsider as well that the bank doesn't only lend its own money to the credit card customers, but other people's money as well (i.e. the money of the bank's creditors). If people put their hard-earned money into a savings account, the bank has the duty to ensure that the money is safe. From this perspective, it would be unethical to lend money to customers with high predicted default risk.", "votes": 6 }, { "id": 1826948, "postDate": "2022-06-20T17:46:15.793Z", "content": "Banks also care about the risk of default in aggregate across all customers so they can adequately reserve against future losses and tighten/loosen their lending standards as that risk changes.
", "rawMarkdown": "Banks also care about the risk of default in aggregate across all customers so they can adequately reserve against future losses and tighten/loosen their lending standards as that risk changes. \n", "votes": 1 }, { "id": 1826928, "postDate": "2022-06-20T17:30:17.567Z", "rawMarkdown": "", "votes": 3, "isDeleted": true }, { "id": 1826444, "postDate": "2022-06-20T10:53:45.887Z", "rawMarkdown": "", "isDeleted": true }, { "id": 1826383, "postDate": "2022-06-20T09:21:13.957Z", "rawMarkdown": "", "isDeleted": true } ], "comments": [ { "id": 1826852, "author_name": "Alberto Lanata", "author_url": "", "post_date": "2022-06-20T16:43:41.740000", "content": "Economics of Information tries to answer this type of question. I suggest reading the paper The Market for Lemons: Quality Uncertainty and the Market Mechanism, by George Akerlof.
\nBy allowing default customers in, you increase costs for the other \"good\" customers. If you do not discriminate customers, you will be charging an average. Which would be unfair for good customers and cheap for bad customers. This may create a perverse cycle where good customers are driven away by the unfair price, making costs higher, which drives prices up again; until only bad customers stay. Insurance Companies have this dilemma.
\nAn accurate model allows you to reduce prices for good customers. You could offer a different service/price for risky customers. Amex could offer loans to default customers if they pay for a default-insurance or something similar. That way Amex doesn´t lose margin, and all customers get access to their products, at a price that reflects their risk level.
", "votes": 8, "replies": [] }, { "id": 1826425, "author_name": "Lucas Morin", "author_url": "", "post_date": "2022-06-20T10:11:41.533000", "content": "There are different answers… (in bold the two elements that directly answer your ethical concerns):
\n1) We are mostly dealing with existing risk - that is already existing customers. AMEX needs models for those to manage their accounts. For exemple if you evaluate an important risk you can provide a relief program to help the customer (and get more money back). This is generally good for distressed clients.
\n2) Evaluating the Risk for existing customer is important for economical and regulatory purposes too. You need to have an evaluation of you portfolio and set capital aside. The set up of the comp. seems to match this regulatory purpose rather than allowing/disallowing clients demand approach. Having an healthy portfolio has positive impact for everyone involved (basically not defaulting yourself).
\n3) The problem of allowing a demand is more complex. We are working with a binary target in a supervised framework. But in real life the problem of allowing/refusing customer demands is slightly different. This is usually not about refusing a customer but setting an interest rate that match their risk. The problem get more complex as you have to take into account any action you offer (relief program - new loan and even the interest rate you want to set) in evaluating the financial health after the decision.
\n4) For the specific AMEX case, given their business (high end travel) and their business model (high fees, high interest rate - >10%, low default rate) they are not geared towards high risk clients and they are probably not refusing that much clients. (fees act as a barrier for high risk clients). Looking at their disclosure it seems AMEX problem is more about the cost of acquisition (advertising, partnership).
\n5) You are right that in general there need to be some regulation to ensure that everyone has minimal access to a bank / minimal financial services. I think in US this is the job of Consumer Financial Protection Bureau. To answer your question more precisely, regulators will usually set a cap for interest rate it is possible to charge (say 15%) thus it is implicitly ethical - even legal - to refuse someone with a risk of default higher than this. Clients that get too much refusals might then ask regulators to designate a bank that will provide minimal services.
\n6) You are also right that more discriminative ML models would remove an important social aspect of banking (sharing of the risk), that can lead to increasing inequalities. The impact should be evaluated in terms of people loosing access to service in a complete framework (with CFPB enforcing the right to have an account). However there is also the ethical aspect of not overcharging healthy clients to compensate for bad ones (sometimes bad on purpose). Lowering interest rates for healthy people / company will generally have a positive impact for the economy (more consumption, more projects, more jobs … etc.) - and more business for you as you get more competitive.
\nSo there is a regulatory framework to refuse clients with too much risk and some very postivie ethical aspects in doing so.
", "votes": 8, "replies": [ { "id": 1827057, "author_name": "xhlulu", "author_url": "", "post_date": "2022-06-20T19:24:59.680000", "content": "Thank you for the extensive answer Lucas! This was very insightful and I'm glad you took the time to share your expertise in this matter :)
", "votes": 1, "replies": [] } ] }, { "id": 1826205, "author_name": "AmbrosM", "author_url": "", "post_date": "2022-06-20T06:36:33.917000", "content": "Yes, of course it is ethical for a private person or company to refuse service to a customer based on predicted default.
\nIt is ethical because everything else would not be sustainable: If you (as a private person or as a credit card company) too often lend money to customers who don't give it back, sooner or later your money will be used up and you'll be out of business.
\nConsider as well that the bank doesn't only lend its own money to the credit card customers, but other people's money as well (i.e. the money of the bank's creditors). If people put their hard-earned money into a savings account, the bank has the duty to ensure that the money is safe. From this perspective, it would be unethical to lend money to customers with high predicted default risk.
", "votes": 6, "replies": [] }, { "id": 1826948, "author_name": "Erik Duus", "author_url": "", "post_date": "2022-06-20T17:46:15.793000", "content": "Banks also care about the risk of default in aggregate across all customers so they can adequately reserve against future losses and tighten/loosen their lending standards as that risk changes.
", "votes": 1, "replies": [] }, { "id": 1826928, "author_name": "", "author_url": "", "post_date": "2022-06-20T17:30:17.567000", "content": "", "votes": 3, "replies": [] }, { "id": 1826444, "author_name": "", "author_url": "", "post_date": "2022-06-20T10:53:45.887000", "content": "", "votes": 0, "replies": [] }, { "id": 1826383, "author_name": "", "author_url": "", "post_date": "2022-06-20T09:21:13.957000", "content": "", "votes": 0, "replies": [] } ], "raw_markdown_by_id": { "1826123": "Correct me if I'm missing something, but I can mainly see two ways default prediction could be useful. (1) Build a forecast system that customers can use to simulate their chances of defaulting, and provide advice to help them avoid defaulting, and (2) predict if a customer is at risk of defaulting in the future and make decisions based on this (such as refusing certain services based on that). \n\n(1) is great, whereas (2) is an open question as there's many unknown variables into play here, but I think that beyond building models that can be used for this, we (the kaggle community) should also discuss about the impact of our work. So here's my question: **if we have accurate models that can predict future default, is it ethical to use it to minimize risk at the cost of customers' ability to use financial services provided by companies like American Express?**\n\nOn one hand, such model could lead to less defaults, and thus less expenses/losses for Amex, and this saving could be passed onto the customer under the form of lower interesting rates, lower annual fees, and higher cashback rewards. Another argument in favor: if we assume that such models are already being used, building more accurate models could result in lower errors (erroneously predicting a custom will default), which is good for customers that would have otherwise been denied certain financial services.\n\nOn the other hand, more conservative use of risk modeling could negatively impact the life of certain customers. Let's say a customer just lost their job, and they need to pay for expensive medicine or treatments without insurance. Their financial situation is not great, so they need to use financial services that are decided by such models. What if the model predicts a 25% default chances? Number wise, that is a big risk for the company owning the model, so they would be incentivized in refusing those services. As a result, the customer might have to make difficult decisions that could eventually lead to them defaulting, which becomes a self-fulfilling prophecy, and is added as another data point confirming the model's correct prediction. Of course, that might be less relevant in some countries than others, but remains a reality.\n\nWith those two sides presented, how would you answer the question?", "1826852": "Economics of Information tries to answer this type of question. I suggest reading the paper The Market for Lemons: Quality Uncertainty and the Market Mechanism, by George Akerlof.\n\nBy allowing default customers in, you increase costs for the other \"good\" customers. If you do not discriminate customers, you will be charging an average. Which would be unfair for good customers and cheap for bad customers. This may create a perverse cycle where good customers are driven away by the unfair price, making costs higher, which drives prices up again; until only bad customers stay. Insurance Companies have this dilemma.\n\nAn accurate model allows you to reduce prices for good customers. You could offer a different service/price for risky customers. Amex could offer loans to default customers if they pay for a default-insurance or something similar. That way Amex doesn´t lose margin, and all customers get access to their products, at a price that reflects their risk level.", "1826425": "There are different answers… (in bold the two elements that directly answer your ethical concerns):\n\n1) We are mostly dealing with existing risk - that is already existing customers. AMEX needs models for those to manage their accounts. For exemple if you evaluate an important risk you can provide a relief program to help the customer (and get more money back). This is generally good for distressed clients.\n\n2) Evaluating the Risk for existing customer is important for economical and regulatory purposes too. You need to have an evaluation of you portfolio and set capital aside. The set up of the comp. seems to match this regulatory purpose rather than allowing/disallowing clients demand approach. Having an healthy portfolio has positive impact for everyone involved (basically not defaulting yourself).\n\n3) The problem of allowing a demand is more complex. We are working with a binary target in a supervised framework. But in real life the problem of allowing/refusing customer demands is slightly different. **This is usually not about refusing a customer but setting an interest rate that match their risk.** The problem get more complex as you have to take into account any action you offer (relief program - new loan and even the interest rate you want to set) in evaluating the financial health after the decision.\n\n4) For the specific AMEX case, given their business (high end travel) and their business model (high fees, high interest rate - >10%, low default rate) they are not geared towards high risk clients and they are probably not refusing that much clients. (fees act as a barrier for high risk clients). Looking at their disclosure it seems AMEX problem is more about the cost of acquisition (advertising, partnership).\n\n5) You are right that in general there need to be some regulation to ensure that everyone has minimal access to a bank / minimal financial services. I think in US this is the job of Consumer Financial Protection Bureau. **To answer your question more precisely, regulators will usually set a cap for interest rate it is possible to charge (say 15%) thus it is implicitly ethical - even legal - to refuse someone with a risk of default higher than this.** Clients that get too much refusals might then ask regulators to designate a bank that will provide minimal services. \n\n6) You are also right that more discriminative ML models would remove an important social aspect of banking (sharing of the risk), that can lead to increasing inequalities. The impact should be evaluated in terms of people loosing access to service in a complete framework (with CFPB enforcing the right to have an account). However there is also the ethical aspect of not overcharging healthy clients to compensate for bad ones (sometimes bad on purpose). Lowering interest rates for healthy people / company will generally have a positive impact for the economy (more consumption, more projects, more jobs … etc.) - and more business for you as you get more competitive.\n\nSo there is a regulatory framework to refuse clients with too much risk and some very postivie ethical aspects in doing so.", "1826205": "Yes, of course it is ethical for a private person or company to refuse service to a customer based on predicted default.\n\nIt is ethical because everything else would not be sustainable: If you (as a private person or as a credit card company) too often lend money to customers who don't give it back, sooner or later your money will be used up and you'll be out of business.\n\nConsider as well that the bank doesn't only lend its own money to the credit card customers, but other people's money as well (i.e. the money of the bank's creditors). If people put their hard-earned money into a savings account, the bank has the duty to ensure that the money is safe. From this perspective, it would be unethical to lend money to customers with high predicted default risk.", "1826948": "Banks also care about the risk of default in aggregate across all customers so they can adequately reserve against future losses and tighten/loosen their lending standards as that risk changes. \n", "1826928": "", "1826444": "", "1826383": "" } }