The Taiwan Banker

For my anniversary, I went shopping in a department store. When I swiped my card, I found that my credit limit had been exceeded, and the transaction was rejected. I called my bank’s customer service to request an increase in credit limit. Will they accept the request? If so, how much would they add to the limit? There are many answers to this question, depending on the tools and management practices at hand.Prerequisites for a modelThere are two main considerations for building a predictive model: business needs, and data. In general, it is less necessary to predict the probability of default (PD) for lending with collateral than without collateral, so it is important to regularly estimate the market price of the collateral.For mortgage lending, for an example, as long as the market value of the collateral (house) is greater than the loan balance (exposure) plus the cost of disposing of the collateral (bank auction or foreclosure), the bank will not lose money in a default. When housing prices rise consistently, the realized default rate is also low, because the lender can sell the house or use it to apply for a home equity loan to repay the debt. At this time, a PD model must be built, but there will be insufficient default cases. If real estate prices fall and the market value of the collateral falls below the loan balance, a model will still not necessarily be needed. If the model is built during a period of rising prices, the data will differ from that for future price declines. In addition, the original default sample will be insufficient, and the model will be inaccurate. What then if we used older data from when house prices fell? In fact, timeliness must be considered. Modeling requires six months of data. Usually, it takes another six months to build the model. Credit cards also require predictive models, and data quality and quantity requirements are comparable with those of other credit products, making them the best choice for application of predictive models. The credit card model converts the results of a logit model (between zero and one) into an easy-to-use score, and thus is called a scorecard model.Scorecard proliferationCredit cards can use scorecards in risk control and marketing. Risk control uses so-called AB scorecards (application and behavior), while marketing & sales uses RR scorecards (response and revenue). There are also collection scorecards: models built for late payments which have not yet turned into bad debts. Risk control scorecards (mainly AB) must be stricter and more accurate, because if they are only slightly off, bad debts will soar. A false positive is better than a false negative, and the loan term can be over 5 years. Marketing scorecards have lower requirements for accuracy, with a life cycle of about 2-3 years. Fair Isaac Corporation (FICO), in contrast, uses a so-called “generic scorecard,” which could be an A-card or a B-card, but because it attempts to do everything, its ability is sparse, and the predictive power does not match up to that of individual scorecards. It should only be used as a last resort. The US usually uses FICO scores to measure a borrower’s creditworthiness. A score above 720 is super prime (≥720), while 660-719 is prime. Under 660 is considered defective; 620-659 is classified as near-prime, and the subprime FICO scores which caused the 2008 financial crisis were from 580-619.Citibank has about 30 million credit card holders in the US and widely uses several scorecards and predictive models, the most important of which are the risk-related AB scorecard, application scorecard, and behavior scorecard. In addition, it also uses revenue scorecards, response scorecards, and collection scorecards. Different cards can also be used together. Challenges in application Fig. 1 uses two scorecards to measure risk and profit respectively. Each parameter is divided into high, medium, and low values, forming a nine-square grid. Different strategies can be formulated according for each area (such as areas 1, 2, and 3 as shown in the Figure). Area 1 is a high-risk area. Cardholders in this area have a higher chance of late payment and default, and must be tracked closely. Preventive measures must be taken if incidents occur with other banks. Area 2 is a high-income, medium-risk area – a source of good income, so banks should take care to retain these customers. Such customers are often hard to find, and customer relationship management (CRM) should play a role here. When American Airlines wanted to launch its own co-branded card, for example, it first approached American Express, only to be rejected, and then went to Citibank. The co-branded credit card turned out to be Citibanks’s most profitable card. Because many users of the co-branded cards are businesspeople, the PD is almost zero. They are a low-risk, high-profit group. Area 3 is a low-risk group; many hold cards but do not use them with a revolving balance, instead paying by lump sum each period. By purchasing their information through the Joint Credit Information Center (JCIC), coupled with data mining analysis, customers might be persuaded to complete balance transfers (BTs), bringing their revolving balances in from other banks. BT has been attempted by many major US banks. In 2016, Goldman Sachs moved into consumer finance, buying out the entire credit card business with GE Capital and converting its portfolio into personal loans. After spending a lot of money, it finally ended its personal credit business last year. In fact, despite great efforts, nobody has made much money through BT, mainly because they have failed at risk control, resulting in bad debts. It would still be worth a try by local banks, if they can do a good job controlling risks. Model application is another challenge. In traditional commercial banks in Taiwan, credit skills are passed down from master to apprentice. Due to changes in customer behavior over time, as well as a lack of systematic integration, banks may miss the forest for the trees; senior employees may even resist introduction of more accurate models. Furthermore, different senior employees may even have conflicting experiences. Statistical models are not sacrosanct. If divorced from reality and not integrated with operations, they are more like academic research, and will have a negative overall impact. The strengths of younger analysts should also be integrated, paying more attention to overall operating status, including credit collection, customer feedback and complaints, and late payment collection. Models are not the only basisWhether in daily operations or formulation of mid- or long-term strategies, models are an important basis for decision-making, but not the only such basis. The more accurate the model, the higher the proportion of the basis. Some models are inherently more accurate, while some are less so. It is important to understand the limitations of a model being used. Returning to credit cards, the process of approving a card usually requires passing some criteria, which may come from compilation of senior employees’ years of experience in credit underwriting, which is ideally also be coupled with evidence-based statistical analysis. One of the conditions may be a score cut-off, directly rejecting scores below the threshold. Assuming the scorecard is well-designed, then other conditions can be relaxed, allowing it to bear the burden of the decision. All scorecards however eventually become invalid, and other conditions may need to be tightened before a new scorecard is established.It is better to carry out regular criteria analysis and model validation, and make prompt adjustments in response to any changes. Verification is an important part of model construction and subsequent application, and is a rigorous science. Citibank set some conditions for direct card rejection (however the approval process must still be completed, and the results are retained and stored for future analysis). These criteria include bankruptcy within the past 7 years, major crimes in the past few years, or a FICO score below 720. Some of these conditions can be adjusted, and must be tested regularly.How should the regulator conduct audits, and what should they audit? They should first confirm that the bank uses a model, and require documentation or data for verification. Banks which do not use models for credit risk control should explain their reasoning in writing. They should first receive a written warning and deadline for improvement, and then a fine and six-month license suspension. Finally, their license can be suspended indefinitely, forcing it to sell its business, and helping it resolve its over-banking problem.Major international banks have complete systems and resources to train talent. After a model is built, before it goes online, senior colleagues from the modeling unit will usually review its documentation. Local banks are unable to replicate this approach. Most advanced models are built by outside consultants, who usually only provide black boxes, due to commercial considerations. The documentation shows how to use and verify it, but does not include the processes, parameters and methodology.What if a local bank builds its own forecasting model? Taiwan currently does not emphasize intellectual property rights. Will it look good if the code is leaked, but nobody can understand the complete model? Or who is responsible if the code is modified by someone who does not understand it as well as they think, and something goes wrong? Transfer of know-how is best done through systems to train and retain talent, rather than through source code or documentation. Copying is not the same as innovating, and makes it hard to find answers to future problems. Credit risk management is not a new concept, but to implement it, local banks must undergo cultural transformation. Creating the conditions to train and retain talent will test the wisdom of executives.Tien-wang Tsaur is Chairman of the Chung-Hua Institution for Economic Research, honorary professor at Soochow University, and chair professor of law and business at Soochow University. Chung-ying Lee has a Ph.D. in economics from Texas A&M University in the United States. He worked in the credit risk department of Citibank's New York headquarters, and later worked in Taiwan, assisting in building credit scorecards, calculating economic capital, designing stress tests, and calculating stress losses in many banks. He is an expert in bank credit risk control and a part-time assistant university professor.