GF582 Unit 9 Credit Scoring Model Comparison Example

Course	GF582 Statistical Methods for Decision Making
Unit	Unit 9
Paper type	Model comparison exercise
Length	About 850 words, 6 pages double-spaced
Format	APA 7 student paper
School	Purdue University Global
Program	MS in Finance
Updated	October 2026

Free sample paper for GF582 Unit 9

Seven Predictors or Four? A Unit 9 Model Comparison Exercise on Small-Business Credit Scoring

[Student Name]

Purdue University Global

GF582: Statistical Methods for Decision Making

Unit 9 Model Comparison Exercise

[Instructor Name]

[Date]

The applications, the models and every statistic are composite and illustrative; the articles and textbook cited are real.

What this part is doingNaming the two models, the holdout and the finding in the lead tells the grader where the comparison ends.

The Data and the Outcome

The dataset holds 4,000 approved small-business loans from a composite lender, each followed for eighteen months. The outcome is default, 90 or more days past due, which occurred on about 6 percent of loans. The lender split the loans at random: 3,000 to fit the models and 1,000 held back to test them on data the models never saw.

Model A: Three Predictors

Credit score, per 50 points: coefficient negative 0.87, odds ratio about 0.42. Each 50 points of score multiplies the odds of default by about 0.42.

Debt-to-income ratio, per 10 points: coefficient 0.31, odds ratio about 1.36.

Log of years in business: coefficient negative 0.45, odds ratio about 0.64 per unit, so a business twice as old has odds about 27 percent lower.

All three are strongly significant, with p-values below 0.001.

Model B: Seven Predictors

Model B keeps the three and adds four:

Revolving credit utilization, per 10 points: coefficient 0.18, odds ratio about 1.20, p below 0.001.

Credit inquiries in the past six months: coefficient 0.09, p about 0.21.

Construction-industry flag: coefficient 0.22, p about 0.18.

Collateral loan-to-value, per 10 points: coefficient 0.05, p about 0.44.

Only utilization is clearly significant. The other three have coefficients in plausible directions but standard errors large enough that their true effects could be zero.

The Comparison Table

Fit on the 3,000 training loans. Akaike's information criterion, which rewards fit but penalizes each added parameter (Akaike, 1974), is about 1,412 for A and 1,398 for B. Lower is better, so B fits better even after the penalty.

Discrimination on the 1,000 held-back loans. The area under the ROC curve, the probability that a randomly chosen defaulter is ranked riskier than a randomly chosen non-defaulter, is about 0.741 for A and 0.757 for B.

Model A plus utilization only. A four-predictor model, A with utilization added, has an AIC of about 1,396 and a holdout area of about 0.755, essentially matching B with three fewer variables.

Model B beats Model A on every measure, but nearly all of its advantage comes from revolving utilization; the other three additions bring complexity and data costs while adding almost nothing that holds up on loans the models have not seen.

What this part is doingThe highlighted sentence identifies where the larger model's gain comes from, the judgment graders look for in a model comparison.

An Approval Simulation

To translate the statistics into lending terms, each model was used to rank the 1,000 held-back applications, and the lowest-risk 70 percent were treated as approved. Among loans approved by Model A, about 3.1 percent defaulted; under Model B, about 2.7 percent; under the four-predictor model, about 2.7 percent. On a book of 700 approved loans averaging $85,000, with losses of 40 percent on default, the 0.4-point reduction in defaults saves about $95,000 a year, whether it comes from B or from the four-predictor model.

Why Fewer Variables Is Better Here

Each added variable must be collected, verified and kept current. Inquiries and loan-to-value require extra data purchases and appraisal reviews. Variables that are not significant also tend to make a model unstable when applied to new loans, because their coefficients may reflect noise in the training data. Textbook guidance on logistic regression favors the simplest model that performs well on holdout data (Hosmer et al., 2013), and research on credit scoring finds that modest gains in discrimination often vanish when models are applied to new populations (Hand & Henley, 1997).

Checks

The four-predictor model's coefficients were refit on the holdout sample and kept their signs and approximate sizes, suggesting it is not overfitted. Its predicted default rates in each risk decile were close to the observed rates.

A Note on Fairness Review

Before adoption, the four-predictor model should be reviewed for whether any variable acts as a proxy for characteristics lenders may not consider, such as the business owner's age or location. Utilization and debt-to-income are standard credit measures, but the review should confirm that approval rates do not differ sharply across protected groups without a credit reason.

What Would Change the Recommendation

If the lender expanded into construction lending, where it now has few loans, the construction flag might become significant on a larger sample and earn its place. The model should be refit annually as the portfolio's mix changes.

Calibration and Stability

A model can rank loans well and still predict the wrong level of defaults. Grouping the holdout loans into ten bands by predicted risk, the four-predictor model's predicted default rates matched observed rates within about one point in every band, which matters because the lender prices loans from those predictions. Model B's predictions were similar, so calibration does not favor the larger model either.

Recommendation

Adopt the four-predictor model: credit score, debt-to-income, years in business and revolving utilization. Drop inquiries, the construction flag and loan-to-value from the scoring model, while keeping loan-to-value as a separate underwriting rule for collateral.

What this part is doingEnding with the variables kept and dropped turns the comparison into a decision a credit team can implement.

References

Akaike, H. (1974). A new look at the statistical model identification. IEEE Transactions on Automatic Control, 19(6), 716-723. https://doi.org/10.1109/TAC.1974.1100705

Hand, D. J., & Henley, W. E. (1997). Statistical classification methods in consumer credit scoring: A review. Journal of the Royal Statistical Society: Series A (Statistics in Society), 160(3), 523-541. https://doi.org/10.1111/j.1467-985X.1997.00078.x

Hosmer, D. W., Lemeshow, S., & Sturdivant, R. X. (2013). Applied logistic regression (3rd ed.). Wiley.

GF582 Unit 9 instructions, in plain terms

The ninth GF582 unit asks you to compare competing models and justify the choice. Describe the data, the outcome and how the sample was split into training and holdout sets. Report each model's coefficients, odds ratios and significance, interpreting them in plain terms. Compare the models with a penalized fit measure on the training data and a discrimination measure on the holdout data. Test whether the larger model's gain comes from all its additions or from a few, for example by fitting an intermediate model. Translate the differences into decision terms, such as default rates among approved loans. Check calibration and stability. Consider costs of collecting each variable and fairness concerns. Recommend a model. Cite research on model selection and credit scoring in APA 7.

How this GF582 Unit 9 example is built

The 4,000 loans, the default definition and the random split open the example. Model A's three coefficients are interpreted as odds ratios, and Model B's four additions are reported with their p-values. The comparison table gives AIC values, citing Akaike, and holdout ROC areas, then shows that a four-predictor model matches B. The highlighted sentence locates B's gain in utilization. An approval simulation converts the models into default rates among approved loans and about $95,000 of annual savings. Sections explain why fewer variables is better, citing Hosmer, Lemeshow and Sturdivant and Hand and Henley, and report checks, calibration by risk band, a fairness review and what would change the recommendation, before the final choice.

Reading the GF582 Unit 9 grading rubric

Grading in this GF582 exercise centers on comparing models fairly and justifying each term. Correct interpretation of coefficients and odds ratios earns the first share. Graders then expect a penalized measure on training data and an out-of-sample measure on held-back data, with the reason each matters. The analytical high point is diagnosing where a larger model's gain comes from, which often reveals that most added terms contribute nothing. Translating model differences into lending outcomes shows applied skill. Instructors in the MS in Finance program value calibration checks, data costs and fairness review. A clear recommendation, readable tables and APA 7 references complete the grade.

Common GF582 Unit 9 mistakes, and how to avoid them

A common GF582 Unit 9 error is choosing the model with the highest training fit, which rewards overfitting. Others compare models only on p-values, or only on AIC without a holdout test. Larger models are accepted whole without asking which added variables matter. Odds ratios are misread as probabilities. Differences in ROC area are reported without saying what they mean for approvals or losses. Calibration and fairness are skipped. Whatever two models the case hands you, the sequence holds: data and split, coefficients, penalized fit, holdout discrimination, a diagnostic intermediate model, decision terms, checks and a recommendation.

Get a GF582 Unit 9 example written to your instructions

Two candidate models, the data or output behind them, and the criteria named for Unit 9 make up the brief, along with the rubric. The custom comparison lines up fit, penalized criteria and holdout evidence and says which model earns its extra terms, returning in 24-48h with the first at no charge. The paper above is an original model document written by our desk, not a submitted student paper and not an official Purdue University Global document.

More GF582 papers and MS in Finance samples

GF582 Unit 9 questions, answered

Why can AIC and BIC choose different models?

They penalize extra parameters differently. AIC adds a fixed penalty per parameter, while BIC's penalty grows with the logarithm of the sample size, so on large datasets BIC favors smaller models more strongly. Neither is wrong; they encode different goals, prediction for AIC and identifying a simpler true model for BIC. A comparison should report both and say which goal the decision serves.

Is a holdout sample better than a penalized criterion?

It tests the thing a scorecard is for, ranking cases it has not seen, so many sections treat it as the stronger evidence. Its weakness is size: a holdout with few defaults gives noisy AUC and error figures. Cross-validation reuses the data to reduce that noise. Reporting holdout results beside AIC or BIC, rather than instead of them, is the usual expectation.

Should non-significant predictors always be dropped?

Not automatically. A predictor may be kept for business or regulatory reasons, or because theory says it belongs, and dropping terms one at a time on p-values alone can mislead. What matters is that the decision is argued. Say why each term stays or goes, and check whether removing it changes the holdout performance the model is judged on.

Where can I find a free GF582 Unit 9 sample paper?

This page holds a full GF582 Unit 9 model comparison exercise on small-business credit scoring, with two logistic models, AIC, holdout ROC areas, an approval simulation and APA 7 references.

What does an odds ratio of 0.42 per 50 points of credit score mean?

Each additional 50 points of score multiplies the odds of default by about 0.42, cutting them by more than half, with the other predictors held constant.