I hereby wish to take part in the SAS Student Ambassador Competition. Please find my essay enclosed in this document.

Transcription

1 Private Bag X6001 North-West University Computer Science and Information Systems Potchefstroom 2531 South Africa Phone: Nina Frauenfeld Academic Initiative Coordinator SAS Europe, Middle East & Africa P.O. Box Neuenheimer Landstr D Heidelberg, Germany Application for SAS Student Ambassador Competition Dear Mrs. Frauenfeld, I hereby wish to take part in the SAS Student Ambassador Competition. Please find my essay enclosed in this document. I feel I should be selected as one of five SAS student ambassadors for the following reasons. SAS forms the foundation of my research which originated from a SAS course. This research was made possible by the successful working relationship between the North-West University in South Africa and SAS Institute which proved beneficial to both industry and academia. With the support from SAS this research culminated into a new modelling technique. In South Africa we were overwhelmed with positive feedback on the research and this made us realize this new modeling technique can be useful to many SAS customers. Consequently, I have a success story to tell the international business community which can directly improve their business practice and increase return on investment. If I am chosen as a student ambassador I can meet business people and discuss new ideas which ultimately benefit the community as a whole. Finally, I want to be a part of and contribute to the winning strategy of SAS Institute. Kind regards, Tiny du Toit

2 The Use of Generalized Additive Neural Networks to Reduce Credit Scorecard Development Time and Increase Return on Investment by Tiny du Toit Credit 1 scoring is a statistical method used to predict the probability that a loan applicant or existing borrower will default on the loan or become seriously delinquent (Mester, 1997). This method was introduced in the 1950s and is now widely used for consumer lending, especially credit cards, and mortgage lending. In business lending there has not been widespread application of this technique, but this is changing. The delay is caused in part by the fact that business loans typically differ substantially across borrowers which makes it harder to develop an accurate method of scoring. With the advent of new methodologies, increased computing power, and increased data availability, such scoring becomes possible. As a result, many banks are beginning to use scoring to evaluate smallbusiness loan applications. Credit scoring is a method that evaluates the credit risk of loan applications. Utilizing historical data and statistical techniques, credit scoring tries to isolate the effects of various applicant characteristics on defaults and delinquencies. The method creates a score that can be used by a bank to rank its loan applicants or borrowers in terms of risk. To built a scoring model, or scorecard, developers analyze historical data on the performance of previously made loans. This helps them to determine which borrower characteristics are useful in predicting whether the loan performed well. A welldesigned model should assign a higher percentage of high scores to borrowers whose loans will perform well and a higher percentage of low scores to borrowers whose loans won t perform well. Unfortunately, no model is perfect, and some bad accounts will receive higher scores than some good accounts. Information on the individuals applying for a loan is obtained from their loan applications and from credit bureaus. Factors such as the borrower s monthly income, outstanding debt, financial assets, how long the borrower has been in the same job, whether the borrower has defaulted or was ever delinquent on a previous loan, whether the borrower owns or rents a home, and the type of bank account the borrower has are all potential characteristics that may relate to loan performance and may end up being used in the scorecard. A higher score in most scoring systems indicates a lower risk, and a lender sets a cutoff score based on the amount of risk it is willing to accept. Strictly adhering to the scorecard, the lender would approve applicants with scores above the cutoff and deny applicants with scores below. In practice, many lenders may take a closer look at applications near the cutoff before making the final credit decision. A good scoring system can t predict with certainty any individual loan s performance, but it should give a fairly accurate prediction of the likelihood that a loan applicant with certain characteristics will default. Developers need sufficient historical data to built a good scoring model. This data must reflect loan performance in periods of both good and bad economic conditions. New information on the credit performance of existing borrowers are frequently obtained and scorecards must be periodically rebuilt to keep the scoring models up to date. Reducing the time it takes to develop credit scorecards is one of the main applications of the research for my Ph.D. thesis which is entitled: Automated Construction of Generalized Additive Neural Networks 1 The word credit is derived from the Latin verb credo, which means I trust or I believe. Taken literally, it means to place trust in an individual's ability to pay.

3 for Predictive Data Mining. I registered for my Ph.D. in Computer Science in 2003 and plan to finish at the end of May, Close collaboration between the Center for Business Mathematics and Informatics at the North-West University in South Africa and SAS Institute made the research possible and this illustrates how both academia and SAS can benefit from such collaboration. The particular way in which this collaboration materialized further illustrated long term benefits that may be derived from all parties involved in such collaboration. Before I discuss the research and explain the benefits of using Generalized Additive Neural Networks in building scorecards, it is worthwhile to start at the beginning where our working relationship with SAS commenced. Where it all started Six years ago, expertise in risk analysis and management was extremely limited in South Africa. In 2000, SAS Institute launched the SAS RiskLab within the Center for Business Mathematics and Informatics (BMI). The latter was established in 1997 as a joint initiative between Absa bank and the university. The Center for BMI is a center of excellence in financial risk/reward management and analysis. It has the mission to train students for professional careers in the area, and conduct research that would benefit the financial services industry and, in particular, the banking industry. Student projects produced at the SAS RiskLab range from implementation of a retail credit portfolio model, on-the-job training of employees on the use of cluster analysis for customer segmentation, and statistical evaluation of the Lifetime Customer Value model. Many of the algorithms developed at the RiskLab are new, and could potentially be marketed as new SAS procedures internationally. The students are so highly prized that banks are using them to help roll-out risk applications within their organizations, and more than half of them are offered jobs before they even qualify. Consequently, the RiskLab is delivering highly qualified risk analysis graduates, alleviating the IT skills crisis, and generating a wealth of risk management knowledge and innovation. Research on Generalized Additive Neural Networks During the past few years the Center for BMI assisted SAS South Africa in presenting SAS Enterprise Miner courses to industry. Enterprise Miner is the most complete and powerful data mining solution on the market today and streamlines the entire data mining process from data access to model deployment by supporting all necessary tasks within a single, integrated solution, all while providing the flexibility for efficient workgroup collaborations (Anonymous, 2005). My advisor, professor André de Waal, presented the Neural Network Course (Potts, 2000) to various members of industry in South Africa. In this course there is a chapter on Generalized Additive Neural Networks (GANNs) that was particularly interesting. Very little research was done on the subject and we decided to investigate it further. We found that GANNs (Potts, 1999) were not available as a basic modeling technique in SAS. Furthermore, this type of neural network is not a black box with respect to interpretation compared to more general artificial neural networks which proved to be a very favorable property when developing credit scorecards. In 2003 and 2004 three research reports followed from this research and talks on the research were given at SAS Forum 2003, SASA 2003 (conference of the South African Statistical Association), SACLA 2003 (conference of the South African Computer Lecturer s Association), as well as SAS Forum During the first half of 2004 we took part in the coveted KDD Cup competition and used GANNs as the modeling technique. We obtained a 7 th place which proved to us that a GANN has predictive power. So much interest were created by the research and the talks that we decided to report back to SAS, Cary, on the latest developments regarding GANNs contained in the research reports. SAS South Africa arranged a visit to Cary and it took place on the 15 th and 16 th of November During the visit we discussed the latest results regarding GANNs, and also the automated construction of GANNs which is similar to the automated neural network construction node that is implemented in Enterprise Miner 5.1. We presented four research talks during the morning of the 15 th of November and the following

4 decisions on collaboration were taken during the visit. The GANN model selection algorithm presented during the visit will be implemented as an add-on node in SAS Enterprise Miner 5.1 by the Center for BMI. Furthermore, the Center for BMI will test the auto neural node in Enterprise Miner 5.1 on some of the data sets used for research at the Center and provide feedback on the performance on the node to SAS, Cary. In the first quarter of 2005 we seamlessly integrated our automated model selection technique (Du Toit & De Waal, 2003) into SAS Enterprise Miner (Du Toit & De Waal, 2005). We used Base SAS and PROC NEURAL to extend the set of modeling nodes and named it AutoGANN. This node has the potential to reduce model building time dramatically for SAS customers and can be used to evaluate current models performance. We went on a road trip in South Africa to demonstrate the AutoGANN node and received favorable feedback from three banks, a vehicle finance house, and a credit bureau. I chose SAS over other types of software to develop the AutoGANN node for a number of reasons. Firstly, AutoGANN integrates seamlessly into the SEMMA data mining methodology of SAS. Other existing nodes can be used by AutoGANN for pre- and post-processing tasks such as taking a sample from large data sets, partitioning data sets into training, test, and validation sets, and imputing missing values. Secondly, by using the Macro Language of Base SAS, I can design meta programs, that is, programs that create and execute other programs. This powerful capability allows AutoGANN to create GANN source code in real time, execute it, and then assess the predictive power of these models. For other types of software without this ability, the different models would need to be coded by hand. Finally, developing programs in SAS takes much less time compared to other types of software without the powerful data manipulation, analytical techniques, and reporting facilities of SAS. Research on the AutoGANN node were included in the documentation that was submitted by the Center for BMI for the SAS Academic Intelligence Award. SAS recognized the Centre for BMI for outstanding innovation in bridging the gap between industry and academia and this prestigious award was presented to the Center for BMI at SAS Forum International 2005 held in Lisbon. Recently a letter of understanding were signed between SAS and the North-West University. According to this agreement SAS will involve researchers from the data mining group in SAS research and development workshops in the USA and Europe in order to facilitate further research on AutoGANN and development co-operation. Using Generalized Additive Neural Networks in Credit Scoring In the field of credit scoring (Thomas, Edelman & Crook, 2002); (McNab & Wynn, 2000), logistic regression models (Kleinbaum, 1994); (Hosmer & Lemeshow, 1989) occupies a central position as it is relatively well understood and an explicit formula can be derived on which credit decisions may be based. Consequently, it is widely used in industry and has become the standard used by most companies. Despite the fact that artificial neural networks may be more powerful than logistic regression, it is not widely used in credit scoring because it is a black box with respect to interpretation and the absence of reasons why the neural network has reached it decisions may be unacceptable. Obtaining regulatory approval for the use of neural networks to make credit decisions may also be an important issue preventing the acceptance and wide use of neural networks in this environment. In contrast to artificial neural networks and logistic regression, generalized additive models is a compromise between inflexible, but docile linear models and flexible, but troublesome, universal approximators. In an article (De Waal & Du Toit, 2005), we compared the performance of a

5 generalized additive model 2 (GAM) to that of a logistic regression model on a home equity data set, where the aim is to predict whether an applicant will eventually default or be seriously delinquent on a loan. This technique is then used to built a scorecard which is compared to a scorecard built using only logistic regression. By using a GANN model, the effect of each input variable (characteristic) on the fitted model can be interpreted using a graphical method called a partial residual plot. The GANN therefore assists in alleviating the black box perception of artificial neural networks with respect to interpretation. Generalized additive models Predictive modeling is the fundamental data mining task. The generic supervised prediction problem consists of a data set of n cases (observations, examples or instances). Associated with each case is a vector of input variables (predictors, features) x1, x2,, xk and a target variable (response, outcome), y. A predictive model maps the inputs to the expected value of the target and is built on a training data set where the target is known. The purpose is to apply the model to new data where the target is not known. A generalized additive model (Hastie & Tibshirani, 1990) is similar in structure to a linear model and a generalized linear model. It has the form 1 g E( y)) = β + f ( x ) + f ( x ) + L+ f k ( x 0 ( k where the expected value of the target on the link scale is expressed as the sum of individual unspecified univariate functions. The link function, g0-1, is used to constrain the range of the predicted values and is the inverse of the activation function g0. When the expected value of the target is bounded between 0 and 1, such as probability, the logit link function given by g 1 0 E( y) ( E( y)) = ln( ) 1 E( y) is appropriate. Each univariate function can be interpreted as the effect of the corresponding input while holding the other inputs constant. GAMs allow more flexibility than linear models and are easy to interpret graphically. A variety of diagnostic plots have been used for more than half a century to assess nonlinear relationships between the target and input variables in multiple regression models. In this study, the visual diagnostics used to aid the model selection process are plots of the fitted univariate functions overlaid on their partial residuals (Ezekiel, 1924); (Larsen & McCleary, 1972). With partial residuals the effect of the individual inputs, adjusted for the effect of the other inputs, can be investigated. ) 2 A generalized additive model (GAM) implemented as a neural network is called a generalized additive neural network or GANN. In this essay, the terms GAM and GANN are used interchangeably since they refer to the same class of models.

6 Home equity example We built two models for the study on a data set containing loan performance information for 5,960 recent home equity loans (Wielenga, Lucas & Georges, 1999). The binary target variable (BAD) indicates whether an applicant eventually defaulted or was seriously delinquent and occurred in 1,189 cases (approximately 20%). There are 12 input variables: REASON (home improvement or debt consolidation), JOB (six occupational categories), LOAN (loan amount requested), MORTDUE (amount due to existing mortgage), VALUE (value of current property), DEBTINC (debt to income ratio), YOJ (years at present job), DEROG (number of derogatory reports), CLNO (number of trade lines), DELINQ (number of delinquent trade lines), CLAGE (age of oldest trade line in months), and NINQ (number of recent credit enquiries). The data set consists of recent applicants granted credit and is split into training (67%) and validation (33%) data sets. There are a high percentage of missing values for DEBTINC (20%) and consequently some method is needed to handle the missing values in the data set. In this example the standard mean value imputation method for interval inputs is used. Note that the newly created imputed variables have the prefix IMP_. Other variables with missing values are handled in a similar way. To illustrate the main difference between a logistic regression model and a generalized additive model, the usual variable transformation step is deliberately omitted. This transformation would handle nonlinear associations between inputs and the target. The logistic regression model is built using standard modeling practices. As the number of BADs is substantially less than the number of GOODs, over sampling might be considered. This is not done as our main motivation in this study is to compare two modeling techniques on a high level and not to tweak an existing model. The generalized additive model is built using the same modeling practice just described. Logistic regression model A stepwise logistic regression is performed on the training data set with the missing values imputed. Standard significant levels of 0.05 are used. The variables REASON and YOJ are deleted from the model, and JOB is transformed into 5 variables using N-1 dummy coding. In the next table, the events classification information gives an indication of how well the model discriminates between bad and good applicants. Data role Target False positive False negative True positive True negative Train BAD ,101 Validate BAD ,532 Table 1: Event Classification for logistic regression The model has 14 variables plus an intercept and thus 15 degrees of freedom. Generalized additive model The AutoGANN node is used to built a generalized additive model in SAS Enterprise Miner. A specific model that is not too complex is given to highlight some important issues that I will discuss. The resulting model is summarized in the following table.

7 Variable Sub-architecture Interpretation LOAN 1 Linear IMP_JOB 1 Linear IMP_REASON 0 Deleted IMP_CLAGE 1 Linear IMP_CLNO 1 Linear IMP_DEBTINC 4 Nonlinear IMP_DELINQ 1 Linear IMP_DEROG 2 Nonlinear IMP_MORTDUE 1 Linear IMP_NINQ 1 Linear IMP_VALUE 1 Linear IMP_YOJ 0 Deleted Table 2: Generalized Additive Model Results In Table 2, the sub-architecture gives an indication of the complexity of the neural network that is used to approximate the univariate function. The higher the number, the more complex the neural network. Linear indicates that the relationship with the target is linear, Nonlinear indicates that the relationship with the target is nonlinear, and Deleted indicates that the variable is deleted from the model. The events classification information for this GAM model is given in Table 3. Data role Target False False True True positive negative positive negative Train BAD ,101 Validate BAD ,495 Table 3: Event Classification for GAM The resulting GAM model has 23 degrees of freedom. Figures 1 to 9 contain the partial residual plots for the variables in the model. The inputs IMP_REASON and IMP_YOJ are deleted from the model and their partial residual plots are not given. Figure 1: Partial residual plot for LOAN Figure 2: Partial residual plot for IMP_CLAGE

8 Figure 3: Partial residual plot for IMP_CLNO Figure 4: Partial residual plot for IMP_DEBTINC Figure 5: Partial residual plot for IMP_DELINQ Figure 6: Partial residual plot for IMP_DEROG Figure 7: Partial residual plot for IMP_MORTDUE Figure 8: Partial residual plot for IMP_NINQ

9 Figure 9: Partial residual plot for IMP_VALUE It can be seen from Figures 4 and 6 that IMP_DEBTINC and IMP_DEROG exhibit nonlinear relationships with the target. An important aspect that needs consideration is whether the complexity of the nonlinear effects exhibited by the two variables is satisfactory given our domain knowledge of the given problem. The following insights can be gained from closer inspection of the partial residuals and fitted curve in Figure 4: Firstly, there exist at least three extreme points that dramatically influence the complexity of the curve. These points need further investigation as they may be regarded as extreme points. Secondly, the fitted curve is probably too complex and may need simplification so that the model will be able to generalize better on new data. Inspection of Figure 6 provides the following insights: The curve is slightly nonlinear and it is worth investigating whether the added complexity of the nonlinear curve is really needed. It is very unlikely that there can be 0.25 derogatory reports and thus one partial residual seems out of place. The value 0.25 is the result of substituting the missing values with the mean value. This partial residual influences the fitted curve and should be further investigated. The remaining 7 variables exhibit linear or slight nonlinear relationships with the target. These plots should be further investigated and the complexity of the univariate functions adjusted so as to achieve a reasonable fit. Comparison of the logistic regression model and generalized additive model The ROC Chart of Figure 10 reveals that the generalized additive model (indicated by ImportModel in the following charts) does significantly better than the logistic regression model. This result is to be expected as two nonlinear trends have been incorporated into the model. The Score Rankings Chart of Figure 11 also indicates that the generalized additive model is significantly better at discriminating between the BAD and GOOD applicants. The Classification Chart of Figure 12 further shows that the generalized additive model correctly identifies a significantly larger number of BAD customers. Tables 1 and 3 contain further details. We need to further investigate the difference between the two models. From the description of the logistic regression model in a previous section, it is clear that at least one important step was omitted from the modelling process: variable transformation was not done. This step is usually done to model nonlinear trends in a more satisfactory way. From Figure 4 and 6 it is clear that the generalized additive model incorporated these nonlinear trends into the computed model in a transparent way. Furthermore, it is not immediately clear what transformations must be done on the imputed variables to improve the logistic regression results so that similar results to that of the generalized additive model can be obtained. At this stage the modeller as a choice: use the more accurate generalized additive model or search for transformations to improve the logistic regression

10 model. If the latter option is exercised, information gained from the generalized additive model may be used to improve the logistic regression model: 1. Variables deleted from the generalized additive model may be investigated for deletion from the logistic regression model (interestingly, in this specific example the variables excluded from the logistic regression are exactly the same as those excluded from the generalized additive model); 2. Variables having linear relationships with the target may be kept unchanged; and 3. Variables having nonlinear relationships may be candidates for transformation. Suitable transformations should be searched for to model the nonlinear relationships. Figure 10: ROC Chart: BAD Figure 11: Score Rankings: BAD Figure 12: Classification Chart

11 A diligent reader may have realised that quite some effort may be needed to arrive at the given generalized additive model presented earlier. The stated advantages may be negated by this effort as it may have been spent on searching for suitable transformations for the logistic regression model giving similar results. This argument may be valid if the generalized additive model is constructed using an interactive approach relying on the inspection of partial residual plots. However, our AutoGANN node in Enterprise Miner automates the construction of generalized additive models. No user interaction is required except for specifying the maximum complexity of the univariate functions. The AutoGANN node gives substantial time savings while still resulting in an accurate model. The generalized additive model presented in this study was constructed using the AutoGANN node, but with limited time allowed for the construction of the model. The model is therefore not optimal and may be further improved upon. Building a scorecard using a generalized additive model The development of scorecards with logistic regression is well-understood and standard practice in many companies (Anonymous). Given the benefits of constructing a generalized additive model, I will explain the process of building a scorecard with this type of model next. It is possible to get access to the outputs of the univariate functions computed by the generalized additive model. The outputs of the univariate functions become the inputs to a classical scorecard building process that includes variable grouping and logistic regression on the weights of evidence. In building the scorecard however, no variable transformations are considered. The generalized additive model computed the transformations and is now considered a pre-processing step. The building of the scorecard is completed using the outputs of the generalized additive model (univariate functions) computed previously. In Table 4, the scorecard built the traditional way with only logistic regression are given and the new scorecard built utilizing the generalized additive model is given in Table 5. The two scorecards are very similar, except for DEBTINC that now has a wider range of scorecard points. This result is to be expected as DEBTINC was identified as the variable with the most complex nonlinear relationship with the target. The scorecard points have also been changed dramatically for this variable. To compute the split values for the new scorecard, the original split values are runned through the univariate functions giving transformed split values that are then used to built the scorecard. Note that the original variable groupings have been preserved and are transferred to the new scorecard, although the scorecard was actually built on the output of the generalized additive neural network. Keeping the variable groupings unchanged is not optimal and groupings for variables exhibiting nonlinear relationships with the target should be reevaluated. There is a change in scorecard points for the variable LOAN due to the fact that only four decimal places were allowed for split values (the transformed split values are on a completely different scale compared to original split values). Closer correlation of the groupings (and therefore scorecard points) can be achieved by allowing more precision for the transformed split values. In Tables 6 and 7, two extracts from the gains tables are given. The difference in accuracy between the two scorecards is obvious. From the 9 th grouping that is highlighted, the following results can be inferred: With an approval rate of 77%, the bad rate for the improved scorecard is more than 2% lower than that of the original scorecard. The approval rate using the new scorecard can be increased from 77% to 82% giving a similar bad rate to that of the old scorecard. These improvements obtained are significant and may provide huge monetary benefits to companies willing and able to exploit the power of generalized additive models.

12 sinput Scorecard Attribute Name Points Clage.-> Clage 120-> Clage 180-> Clage 240-> 74 Clno.->9 34 Clno 9->13 60 Clno 13->18 59 Clno 18->24 55 Clno 24->28 54 Clno 28->. 49 Debtinc.->30 96 Debtinc 30->35 29 Debtinc 35->40 82 Debtinc 40->. 54 Delinq.-> Delinq 0.9-> Delinq 1.9->. 4 Derog.-> Derog 1.9->. 15 Job Sales 35 Job Self 44 Job Mgr 49 Job Other 50 Job Profexe 58 Job Office 62 Mortdue.-> Mortdue > Mortdue >. 58 Ninq.-> Ninq 0.9-> Ninq 3.9->. 29 Value.-> Value > Value >. 49 Loan.-> Loan 5999-> Loan 9999-> Loan > Loan > Loan > Loan >. 56 Input Scorecard Attribute Name Points Clage.-> Clage 120-> Clage 180-> Clage 240-> 74 Clno.->9 34 Clno 9->13 60 Clno 13->18 59 Clno 18->24 55 Clno 24->28 54 Clno 28->. 49 Debtinc.-> Debtinc 30->35 95 Debtinc 35->40 84 Debtinc 40->. 22 Delinq.-> Delinq 0.9-> Delinq 1.9->. 4 Derog.-> Derog 1.9->. 17 Job Sales 36 Job Self 44 Job Mgr 49 Job Other 50 Job Profexe 58 Job Office 61 Mortdue.-> Mortdue > Mortdue >. 60 Ninq.-> Ninq 0.9-> Ninq 3.9->. 32 Value.-> Value > Value >. 48 Loan.-> Loan 5999-> Loan 9999-> Loan > Loan > Loan > Loan >. 55 Table 4: Scorecard Built with Logistic Regression Table 5: Scorecard Built with GANN Conclusions As nonlinear models are better understood and become available in statistical and data mining systems, the move from linear models to nonlinear models is inevitable (progress waits for no one). There are however external constraints, such as the need for regulatory approval, that may hinder or temporarily delay the replacement of linear models by yet unproven, but potentially more powerful, nonlinear models such as generalized linear models and neural networks. In this essay, a way forward is sketched. The standard theory of scorecard building is not tampered with, but a pre-processing step is introduced to arrive at a more accurate scorecard that discriminates better between good and bad applicants. The pre-processing step exploits generalized additive

13 models (implemented as generalized additive neural networks) to achieve significant reductions in marginal and cumulative bad rates. Table 6: Gains Table for Logistic Regression Table 7: Gains Table for GANN The AutoGANN node we developed, automatically creates generalized additive models and no user interaction is required. This ensures that models are selected objectively and the problems centered around the choosing of good models is largely solved. When logistic regression models are built, the modeler must search for suitable variable transformations to capture nonlinear relationships. This step is automatically handled by the AutoGANN node. By using this node the time it takes to built scorecards can be drastically reduced. Furthermore, the more powerful generalized additive models can increase return on investment with more accurate scorecards. In future, these types of automated model building techniques will become more common to analyze the data deluge. This data glut is the result of the prevalence of automatic data collection, electronic instrumentation, and online transactional processing (OLTP). We as humans can t process all the data and need automated methods to assist us in finding interesting patterns in the data. In this essay I discussed the successful collaboration between the Center for BMI and SAS Institute. This working relationship proved to be beneficial to both industry and academia. My research on generalized additive neural networks originated from a SAS course and with the support we received from SAS this research culminated into a new modeling node in Enterprise Miner. In South Africa we were overwhelmed with positive feedback and this made us realize the AutoGANN node can be useful to many SAS customers. SAS Forum International will provide me with the opportunity to introduce this modeling node to an international business audience. Also, here I can meet business people and discuss new ideas which ultimately benefit the community as a whole. Thousands of delegates from around the world attend this annual conference which serves as a platform where people get to know each other, share knowledge and experience, and learn new business processes. Each year, a wide number of papers are offered by people having practical as well as theoretical experience. Before 2004 this event was known as SEUGI which stood for SAS European Users' Group International. The name change coincides with a shift in emphasis for the conference, which now prides itself as the premier international event for Enterprise Intelligence. Finally, I want to be a part of and contribute to the winning strategy of SAS Institute.

14 Acknowledgements I wish to thank SAS Institute for providing us with Base SAS and SAS Enterprise Miner software used in computing all the results presented in this essay. This work forms part of the research done at the North-West University within the TELKOM CoE research programme, funded by TELKOM, GRINTEK TELECOM and THRIP. Bibliography Anonymous, Building Consumer Credit Scoring Models with Enterprise Miner White Paper. SAS Institute Inc., Cary, NC, United States of America. Anonymous (2005), SAS Enterprise Miner 5.1 Fact Sheet. [Web:] [Date of access: 1 November 2005]. De Waal, D. A. and Du Toit, J. V. (2005), An Investigation into the Use of Generalized Additive Neural Networks in Credit Scoring, In Proceedings of Credit Scoring & Credit Control IX, University of Edinburgh, Scotland. Du Toit, J. V. and De Waal, D. A. (2003), Automated construction of generalized additive neural networks for predictive data mining. In Proceedings of the 33rd annual conference of the South African Computer lecturer s Association, J. Mende and I. Sanders, eds. Du Toit, J. V. and De Waal, D. A. (2005), The AutoGANN node in SAS Enterprise Miner (in preparation). Ezekiel, M. (1924), A method for handling curvilinear correlation for any number of variables. Journal of the American Statistical Association, 19(148): Hastie, T. J. and Tibshirani, R. J. (1990), Generalized Additive Models, Vol. 43 of Monographs on Statistics and Applied Probability, Chapman and Hall, London. Hosmer, D. W. & Lemeshow, S. (1989), Applied logistic regression, Wiley, New York. Kleinbaum, D. G. (1994), Logistic regression: a self-learning text, Springer, New York. Larsen, W. A. and McCleary, S. J. (1972), The use of partial residual plots in regression analysis. Technometrics, 14(3): McNab, H. & Wynn, A. (2000), Principles and Practice of Consumer Credit Risk Management, Financial World Publishing, Canterbury. Mester, L. J. (1997), What s the Point of Credit Scoring?, Business Review, Federal Reserve Bank of Philadelphia. Potts, W. J. E. (1999), Generalized additive neural networks. In Proceedings of the Fifth ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, Potts, W. J. E. (2000), Neural Network Modeling Course Notes, SAS Institute Inc., Cary, NC, United States of America. Thomas, L. C., Edelman, D. B. & Crook, J. N. (2002), Credit Scoring and Its Applications, SIAM, Philadelphia.

15 Wielenga, D. and Lucas, B. & Georges, J. (1999), Enterprise Miner: Applying Data Mining Techniques Course Notes. SAS Institute Inc., Cary, NC, United States of America.