| Credit scoring: myths, dreams & realities |
 |
 |
 |
| Monday, 11 August 2008 01:13 |
|
This article, written by PIC Solutions, explores some of the myths, dreams and realities in the credit scoring arena. Over time PIC Solutions has provided clients with numerous scoring solutions, from the delivery of generic scorecards, the development of custom scorecards through to the mentorship of internal teams hoping to develop their own scorecards. Clients have ranged from the micro finance fraternity through to credit bureaux, throughout the EMEA region and across all credit product types. Each assignment that PIC Solutions has embarked upon has presented its own set of unique challenges and interesting questions, often requiring the demystification of credit scoring. Some of the myths encountered range from “I don’t need scoring as I know all my customers” through to “All I need is technology to automatically generate my credit scorecards for me.” We will therefore take you through the credit scoring maze, stopping briefly to embrace some of the myths, dreams and realities often encountered. The views expressed are those of the author and may be topical in some instances. Credit Scoring Myth 1: All Customers Are The Same! Worldwide, the consumer credit environment has matured sufficiently to the point where there a few remaining organisations that treat new and repeat credit customers with the same ‘credit brush’. Some credit organisations may only apply simplistic credit policies, whereas others have already honed their ability to differentiate credit risk down to a micron level. Even in private banking circles, the myth that high net worth individuals present the same level of risk is no longer held. On the other end of the credit scale, the micro finance area scoring has proven its ability to sift the chaff from the wheat. Companies have also come to realise that a new customer presents a very different risk to that of an existing customer and now apply the correct scorecard for the correct purpose. The fallacy that a demographic-based application scorecard is applicable to existing customers exhibiting behavioural trends has - hopefully - been put to bed forever. Credit Scoring Myth 2: A Credit Scorecard Is A Credit Scorecard! The fallacy that one scorecard is just like another still exists. Scorecards were originally developed to rank order the risk of customers applying for first time credit and this principle was soon adopted by forward thinking organisations. Since then scorecard developers have quickly realised that existing customers generate a wealth of behavioural information that is more powerful in predicting future risk than a customer’s demographics. As we know by now a scorecard typically comprises of a set of questions (characteristics) for which each answer (attribute) is given a number (weight). The summary of the weights is the final score which should represent a predefined level of risk for the specific credit product that it is being applied to. Application scorecards use the demographic information available from the customer at the point of the credit application. This information typically includes: age, years at employment, marital status, etc. The key issue is that the credit grantor often has to rely on the integrity of the applicant when determining the accuracy of the information supplied. As the bulk of application information is not performance related (i.e. payment behaviour) these scorecards are generally not highly predictive. Behaviour scorecards, on the other hand, use intricate behavioural trends and interactions found on the customers account(s) to predict future payment risk. The integrity of this information is often better as the customer cannot fabricate whether a transaction took place or not (e.g. purchase or payment). The power of behavioural information is that the basic characteristics such as balance, payment, purchases, delinquency, etc. are used to derive up to 500 new characteristics which can be used to pick up most behavioural nuances that can separate a good customer from a bad one. Credit Bureau scorecards use the pooled behavioural information supplied by the contributing members to differentiate between paying and non-paying consumers. A well-built behaviour scorecard for a single product portfolio should always be more powerful than a credit bureau scorecard. The reason behind this is that the bureau scorecard uses pooled information. Dependence on information from many sources is often fraught with many issues. One such example is when the information is stored at the bureau (members may lodge their data later than required affecting the recency of the data collected at the bureau). Another anomaly is the diverse definitions existing for data across members and the variations of information available for each credit product. Organisations should not assume that, with a strong suite of application and behavioural models, the credit bureau scorecards will not add any value. Financial services companies have long realised the value that can be obtained by blending the power of the credit bureau scorecards with that of their own internal scorecards. The credit bureau scorecard provides organisation with a barometer of the customers ‘overall credit health’ whereas the internal scorecards hone in more closely on information known only to the organisation. Some credit bureaux have broken new ground by supplying the behaviour characteristics with which an organisation can build their own internal models and flex out new predictive information. This will certainly become the future trend when using credit bureau information as it provides the ultimate level of flexibility for any organisation. Credit Scoring Myth 3: Credit Scoring Cannot Be Used Across The Entire Consumer Credit Life Cycle! The myth that scoring is not applicable across the entire customer life cycle is sometimes still found. The credit industry long ago already recognised that a customer goes through many phases and coined this the credit life cycle. The first stage is to encourage a prospective customer to respond to a pre-approved credit offering or to encourage them to apply for a specific credit product. The next stage is for the organisation to ascertain whether the applicant falls into their credit approval framework and to either approve or decline the application. Once the customer has been approved the credit life cycle continues with the morphing of the credit product terms and conditions according to the customer’s subsequent behavioural patterns. The reality is that there will be a number of customers who become delinquent and are eventually written-off and a number who decide to close their accounts. Throughout this process there are different credit scores that can be applied, each having their own outcome definition, purpose and shelf-life.  The scorecard types that have been developed to-date are too many to describe in this article, however they can be broadly categorized according to the credit life cycle. During the marketing cycle the desired outcome is to model the propensity for the prospect(s) to respond to the marketing collateral sent or pre-approved credit offer made. Prudent organisations pre-screen their marketing list against credit bureau information to ensure that the prospects pass the first credit decisioning hurdle, namely risk. Performing this pre-screening of prospects has the added benefit in that cost savings can be achieved by not processing an applicant that would have been declined in the first place. The challenge with response modelling is to establish the fine balance between the cost of the model development and the desired response required for the marketing campaign to be a success. Response models generally have a short shelf life as they are built on specific information which may not be used in future marketing campaigns.The second stage during the credit life cycle, namely new business assessment, is the one area that application scorecards are fully operational. Over and above the traditional application risk scorecard and the use of credit bureau scores during this cycle, usage or spend propensity scorecards are being applied more frequently. Leading financial services organisations are becoming more adept at not only maximizing their acceptance rates by fine-tuning their risk assessment processes, but also by honing their credit product terms and conditions according to the customer propensity for future usage. Over time this is one area that will receive a lot more attention due to the intense level of competition for the same good quality customer. The third cycle, namely repeat business assessment, really delves into the account management of existing customers. This is the space the traditional behaviour risk scorecards (and credit bureau scores) are applied and the relationship with customers is either maximised or bad performing customers ‘encouraged’ to leave. Spend and attrition scorecards are now playing an increasingly important role in ensuring that an organisation can maximise spend with good quality customers whilst ensuring that proper mechanisms are implemented to slow down the ever increasing attrition problem. There is no silver bullet for the increased level of competition for the same customer, however those organisations that have come to realise that risk is not the only factor in profitability are the financial institutions that are going to continue to sustain future profits. The development and implementation of scoring models that not only focuses on a customer risk is what is going to ensure that organizations maintain their future success. This is reality in today’s competitive world, no matter which country organisations are in or which credit products they may provide. The final two customer cycles can really be dealt with together, namely when a customer is unable to maintain the required payment level and is eventually written off as a bad debt. It is a reality that once an organisation enters the credit world, that there will be bad debt. The art for all organisations is to ensure that the casualties are either kept to a minimum or they can be justified. Collection and recovery agents have feasted, for many years, on those customers who have become credit casualties often leaving a bad stigma for the collections industry. Credit grantors have long dreamt of finding the fine balance between their internal collections costs and that of the ‘easy pickings’ provided to collections agencies. The correct development and application of collections and recovery scorecards has enabled organisations to find this balance and substantially improve profitability. Credit Scoring Myth 4: All I Need Is A Statistician To Build A Credit Model For Me! Scorecard developers have been put on a pedestal and have been deemed to be all encompassing statisticians. The truth is that a good scorecard developer should have a good statistical grounding and ideally have completed the required tertiary education. The reality is that up to 80% of any scorecard development project deals with the client’s data and the desired outcome of the model, all within the constraints of the specific credit portfolio.  It is easy to build a credit model using appropriate development technology, however the challenge is to ensure that this model will serve its intended purpose and is robust enough to stand the test of time. Scorecard developers who do not fully understand the credit business and - just as importantly - do not understand the correct application of the scorecard, have been known to build sub-standard models. This has unfortunately sometimes led to a negative connotation for credit scoring, reducing credibility of a very important aspect of credit. Credit Scoring Myth 5: Technology Is The Deciding Factor When Developing Credit Models! Some scorecard vendors and suppliers of scorecard development software position technology as the deciding factor when building predictive models. The reality is that technology may be a contributing factor for building a specific model, but it should not be the deciding factor. There is a multitude of technology categories – and variations thereof – in existence, applied by organisations in various shapes and forms. Each variation is purported to be better than another and provides that ‘leading edge improvement’. Examining each category – with its variations – falls outside the scope of this article, therefore we will only briefly cover the more traditional and popular categories, namely trees, neural networks and logistic regression.  Decision trees (interaction term models) have been extensively used and are readily available through inexpensive software packages. As trees are visually descriptive, they are easy to comprehend, the results are easy to represent and interpret. Some of the major negatives of trees are that they require huge samples of data and are often difficult to implement. The popularity of trees for robust credit models has long gone, however they are still being used in the marketing arena for customer analysis and sometimes for quick and dirty response models. Neural networks model the interaction between variables very well and are superb for handling unknown and unexpected interactions. They have found extensive application in the fraud environment where the models have to be developed on very scant information. Neural networks are ‘learning models’ which can adapt to new circumstances in the data more readily than any other modelling technique. These models do however suffer from many issues, namely the ‘black box’ syndrome as the user of the model does not understand how the interactions have been defined. Human nature dictates that the neophyte wants to understand how something works before applying it with confidence. This often leads to a difficult buy-in from the decision makers in a business for implementing neural technology. Logistic regression remains an old favourite amongst the credit grantors as the scores can be easily interpreted and easily implemented. Software to develop logistic regression models is nowadays fairly inexpensive and readily obtainable. Logistic regression is the COBOL of the scorecard development fraternity, has been around for a long time, has been ‘replaced’ with many differing technology offerings, but is still the workhorse in its particular field. Saying that, logistic regression has some serious constraints where omitted variables can result in a bias in the coefficients and multi co-linearity increases the standard errors for the coefficients. The inclusion of irrelevant variables can result in a model that does not fit the data sample very well. As it is still an art to developing programs in COBOL, it is an art for scorecard developers to apply logistic regression correctly. The reality is that when comparing technologies data analysis is fundamental to the final strength and future application of any credit scorecard. No one modelling technique is superior and produces models with a longer shelf life than another. The reality is that a well constructed model, by an experienced analyst, can stand the test of time and that the strength of a well thought out model seems to be independent of technology. Credit Scoring Myth 6: Now That I Believe In Credit Scoring, I Want To Build My Own! Once an organisation embraces a new business principle the natural tendency is to cost-justify owning the new process. The dream of many established credit grantors has been to remove the substantial costs of outsourcing scorecard development so that the financial resources can be freed up internally. The reality is that this depends on the stage that an organisation is in during the scorecard development lifecycle. Typically all scorecard development is still outsourced for small and start-up credit grantors who have low volumes and often limited account performance information. As these organisations will be new to predictive techniques, it does not make business sense for them to tackle scorecard development themselves. Once a financial institution starts building up sufficient account volumes and account performance, the development of in-house models becomes a reality. Often these organisations will start dabbling with low cost, low impact developments such as response models. When an organisation has reached a level of maturity and has high account volumes to warrant the development of a number of models, the cost justification of grooming an in-house development becomes feasible. Before the dream of establishing an in-house development team can be turned into reality, there are a number of issues to consider. Some of the major benefits of having an in-house team are that the organisation has control over the deliverables and can now use this new-found skill to perform regular maintenance of existing scorecards. Provided that there is a sufficient need for new scorecards, the cost benefit of this adventure is generally realised. Another spin-off is that the newly established team can be successfully applied in other areas such as economic forecasting, bad debt provisioning and treasury requirements. There are some issues associated with the creation of an in-house team, namely the long initial development timeframes whilst the team is being trained and the cost of purchasing new software. This new software often requires the recruitment of new specialised staff who may, in turn, embark on a steep learning curve. Once these new skills are on board, they need to be retained and challenged enough to ensure that they make a success of the new business function. As this new function is normally not a core competency of a credit grantor, one often finds that the decision makers in the business do not buy into their credit models very easily. Sadly it is often easier for an external party, such as a vendor, to get the undivided attention of decision makers in a business. This can result in an in-house team developing effective predictive models that gather dust on the shelf. There are some tangible benefits for an established credit grantor to outsource the development of their credit models. Scorecard development vendors generally have access to expert knowledge in the industry and have access to experienced specialised resources. As said before, it is often easier for a vendor to obtain business buy-in to ensure that a recently developed model will be implemented and applied correctly. Another benefit is that the organisation maintains focus on its core functions, namely the granting of credit. A scorecard development vendor should always have access to leading-edge technology and techniques to ensure that they have value to pass on to their clients, keeping both clients and vendor competitive in their respective fields. However, there are some pitfalls when outsourcing scorecard development as the client has limited control over the deliverables and the direct costs are normally very high. There is often an associated consulting cost for the effective implementation of the model. It should be remembered that the outsourcing of the development of response models is generally not viable as these have a short shelf life. Summary There are many myths, dreams and realities in the scoring industry. One truth is that credit scoring is here to stay and that only the companies that fully embrace its potential are going to attain profitability. Hopefully this short article has touched on a few of the more common issues on credit scoring and has provided insight into some of these topics. About the Contributor PIC Solutions is the leading specialist credit risk solutions company in the EMEA region. With offices in Cape Town, Dubai, Johannesburg, Manama and Nairobi, we deliver integrated analytics, consulting and software solutions to over 150 companies in 30+ countries. We work worldwide with organisations to improve performance, drive strategies and enhance profitability. Analytics | Consulting | Software |
| Last Updated on Friday, 21 May 2010 13:21 |