An interactive credit-risk analytics workstation developed for the 2026 DataQuest project. The system combines exploratory data analysis, statistical research, interpretable machine learning, logistic regression, credit-scorecard concepts, and business decision support into a unified analytical platform.
The primary objective is not simply to predict credit default, but to understand why borrowers are considered risky, identify meaningful patterns in the data, and use those insights to improve an interpretable logistic regression model.
Credit-risk modelling requires more than achieving a high predictive score. Financial institutions need models that can be:
- Interpreted by analysts
- Explained to decision-makers
- Audited and validated
- Translated into business rules
- Used to support lending decisions
This project therefore follows an exploration → interpretation → modelling → decision-support workflow.
The workstation allows an analyst to:
- Explore borrower characteristics.
- Investigate relationships between variables.
- Identify data-quality problems.
- Research alternative analytical techniques.
- Discover risk patterns and subgroup behaviour.
- Engineer meaningful modelling features.
- Train and evaluate logistic regression.
- Interpret model coefficients and statistical significance.
- Translate model outputs into scorecard-style points.
- Simulate business decisions using different approval thresholds.
The project focuses on four main objectives:
Understand the structure and behaviour of the credit-risk dataset through interactive visualisation and statistical analysis.
Develop an interpretable logistic regression model for binary default prediction.
Use exploratory techniques to identify relationships, transformations, interactions, and subgroup behaviours that may improve the final model.
Translate predictive model outputs into practical lending decisions and demonstrate the trade-off between approval volume and portfolio risk.
The project deliberately separates exploration from final prediction.
Alternative analytical techniques can be useful for discovering structure in the data without necessarily becoming the final production model.
The workflow is therefore:
Raw Credit Data
↓
Data Quality Assessment
↓
Exploratory Data Analysis
↓
Statistical / Multivariate Research
↓
Risk Pattern Discovery
↓
Feature Engineering
↓
Logistic Regression
↓
Model Evaluation
↓
Scorecard Interpretation
↓
Business Policy Simulation
This allows more complex analytical techniques to contribute to the project while keeping the final predictive model interpretable.
The workstation is designed around borrower-level credit-risk information.
Key variables include:
| Variable | Description |
|---|---|
| FICO Score | Numerical representation of borrower creditworthiness |
| Debt-to-Income (DTI) | Proportion of income committed to debt repayments |
| Annual Income | Borrower's annual income |
| Employment Length | Number of years in current employment |
| Defaulted | Binary target indicating whether the borrower defaulted |
The target variable is binary:
0 → No Default
1 → Default
A baseline research framework was created for each analytical technique investigated.
Each research topic follows the same structure:
A conceptual explanation of the method and its purpose.
The underlying mechanics, mathematical formulation, equations, and visual interpretation.
How the technique could be applied to borrower behaviour and default analysis.
Critical evaluation of the method, including interpretability, complexity, assumptions, and limitations.
Evaluation of how the method differs from the final interpretable modelling approach.
Explanation of whether the technique contributed to:
- Feature engineering
- Risk segmentation
- Pattern discovery
- Variable selection
- Interpretation
- Model design
- Business understanding
The final section is particularly important because it demonstrates that the research was integrated into the project rather than simply being theoretical background.
Several methods were considered as part of the exploratory and research process.
Used to investigate non-linear relationships and identify potential decision boundaries or subgroups within borrower populations.
Investigated as an alternative statistical classification approach and compared against logistic regression in terms of assumptions and interpretability.
Considered for exploring relationships between categorical variables and identifying groups of borrowers with similar categorical profiles.
Investigated as a method for understanding associations between categorical dimensions.
Considered as a dimensionality-reduction and visualisation technique for exploring similarity between observations.
These methods were not necessarily included as final production models. Instead, they were evaluated according to their ability to generate interpretable insights that could influence the final logistic regression model.
The workstation provides interactive exploratory modules for examining the dataset before modelling.
Individual variables can be inspected to understand:
- Distribution
- Central tendency
- Spread
- Skewness
- Outliers
- Missing values
- Differences between default and non-default populations
For example, FICO distributions can be compared between:
GOOD → Borrowers who did not default
BAD → Borrowers who defaulted
This provides an initial indication of whether a variable contains useful discriminatory information.
Relationships between variables can be investigated using:
- Scatter plots
- Correlation analysis
- Density plots
- Group comparisons
- Interactive visualisations
The purpose is to identify:
- Correlations
- Potential multicollinearity
- Interactions
- Non-linear relationships
- Risk concentrations
- Subgroup behaviour
These insights can then influence feature engineering and model specification.
Data quality is treated as an important part of the modelling pipeline.
The workstation investigates:
- Missing values
- Invalid values
- Extreme observations
- Skewed distributions
- Potential anomalies
- Outliers
Preprocessing functionality includes approaches such as:
- Outlier capping
- Winsorisation
- Imputation
- Anomaly handling
The goal is to improve model stability without introducing unnecessary bias into the default-rate calculations.
Credit-risk modelling commonly requires interpretable transformations of predictor variables.
The project therefore investigates Weight of Evidence (WoE) and Information Value (IV).
WoE measures the relative strength of evidence associated with a particular variable bin.
Conceptually:
WoE = ln(% Goods / % Bads)
The sign and magnitude provide information about how strongly a particular category or interval is associated with risk.
Information Value measures the overall discriminatory usefulness of a variable across its bins.
Conceptually:
IV = Σ (% Goods - % Bads) × WoE
These concepts help assess whether variables contain useful information for credit-risk modelling.
Logistic regression is the primary predictive modelling technique used in the project.
The model estimates the probability of default:
P(Default = 1 | X)
using the logistic function:
p = 1 / (1 + e^(-z))
where:
z = β₀ + β₁X₁ + β₂X₂ + ... + βₙXₙ
The resulting probability can then be converted into a risk classification or used for ranking borrowers by risk.
The model provides several interpretable statistical outputs.
Each coefficient represents the change in the log-odds of default associated with a one-unit change in the predictor, holding other variables constant.
Represents the baseline log-odds when predictor values are zero.
Measures uncertainty around the estimated coefficient.
Used to assess whether there is statistical evidence that a predictor contributes to the model.
Calculated as:
Odds Ratio = e^β
This provides a more intuitive interpretation of how the odds of default change with a predictor.
The model is evaluated using several complementary metrics.
AUC measures the model's ability to distinguish between defaulting and non-defaulting borrowers across different classification thresholds.
AUC = 0.50 → Random discrimination
AUC = 1.00 → Perfect discrimination
An AUC of approximately 0.892 was achieved in the demonstrated modelling configuration.
The Gini coefficient is related directly to AUC:
Gini = 2 × AUC - 1
For example:
AUC = 0.892
Gini = 2(0.892) - 1
= 0.784
≈ 78.4%
The Gini statistic provides another way of expressing discriminatory power.
The KS statistic measures the maximum separation between the cumulative distributions of goods and bads.
It is particularly useful for understanding how effectively the model separates risky and non-risky populations across score bands.
The project extends logistic regression interpretation into a scorecard-style framework.
Instead of presenting only:
Probability of Default = 0.23
the model can be translated into a points-based representation.
For example:
Higher Risk
↓
Fewer Points
↓
Lower Credit Score
and:
Lower Risk
↓
More Points
↓
Higher Credit Score
This provides a more intuitive representation for users such as credit analysts and underwriters.
Predictive modelling is connected to business decision-making through an interactive policy simulation.
The dashboard allows analysts to investigate how changing a decision threshold affects:
- Approval volume
- Default rate
- Precision
- Recall
- Portfolio risk
- Potential business trade-offs
This demonstrates an important principle:
A model does not make the business decision by itself. The decision threshold determines how model predictions are converted into policy.
For example:
Lower Approval Threshold
↓
More Applications Approved
↓
Higher Potential Volume
↓
Potentially Higher Default Exposure
Whereas:
Higher Approval Threshold
↓
Fewer Applications Approved
↓
Lower Expected Risk
↓
Potentially Lower Portfolio Volume
The dashboard allows this trade-off to be explored interactively.
The final application combines the analytical components into a unified workstation.
Major modules include:
Provides mathematical and conceptual explanations of the analytical methods used throughout the project.
Allows individual variables and their relationship with default behaviour to be explored.
Investigates relationships between variables and potential interaction patterns.
Provides a structured overview of missing values, outliers, anomalies, and preprocessing decisions.
Displays model performance and statistical parameters.
Converts model information into an interpretable points-based representation.
Allows lending-policy scenarios and risk/volume trade-offs to be explored.
Documents how artificial intelligence was used during the development and reasoning process.
AI was used as an analytical and reasoning assistant, rather than as a replacement for project implementation or decision-making.
AI assistance was primarily used to:
- Interpret the project specification
- Clarify modelling requirements
- Explain statistical concepts
- Compare analytical techniques
- Validate architectural decisions
- Explore possible modelling approaches
- Discuss feature-engineering strategies
- Review conceptual interpretations
- Assist with presentation planning
The final analytical decisions remained subject to human evaluation.
The AI reflection process specifically considered:
AI Suggestion
↓
Human Evaluation
↓
Accept / Modify / Reject
↓
Reason Documented
↓
Project Decision
This approach ensures that AI-generated suggestions were treated as inputs into the reasoning process rather than automatically accepted as correct.
The project's AI reflection follows several key themes.
AI was used to break down the DataQuest requirements and clarify the relationship between:
- EDA
- Feature engineering
- Interpretable machine learning
- Logistic regression
- Business decision support
AI was used to evaluate whether the proposed architecture and analytical workflow aligned with the project requirements.
Different modelling and exploratory techniques were discussed to understand their advantages, disadvantages, assumptions, and relevance to credit-risk analysis.
The project distinguishes between:
AI-assisted reasoning
and:
Final project decisions
The latter were determined through evaluation of the project's requirements, data, analytical objectives, and practical constraints.
The application can be demonstrated through the following workflow:
Demonstrate the data portal and show how a credit-risk dataset is imported.
Identify missing values, anomalies, skewness, and outliers.
Use univariate and bivariate analysis to understand borrower behaviour.
Explain why different techniques were considered and how they contributed to the project.
Train and evaluate the final interpretable predictive model.
Review:
- Coefficients
- Standard errors
- P-values
- Odds ratios
- AUC
- Gini
- KS
Translate model behaviour into a more intuitive points-based framework.
Change decision thresholds and observe the effect on portfolio volume and risk.
Demonstrate how AI contributed to the development process and how decisions were evaluated.
At a high level, the system follows:
┌─────────────────────┐
│ Credit Dataset │
└──────────┬──────────┘
↓
┌─────────────────────┐
│ Data Validation │
│ & Quality Checks │
└──────────┬──────────┘
↓
┌─────────────────────┐
│ EDA │
│ Univariate/Bivariate│
└──────────┬──────────┘
↓
┌─────────────────────┐
│ Analytical Research │
│ Trees / MCA / DA / │
│ CA / MDS / WoE / IV │
└──────────┬──────────┘
↓
┌─────────────────────┐
│ Feature Engineering │
└──────────┬──────────┘
↓
┌─────────────────────┐
│ Logistic Regression │
└──────────┬──────────┘
↓
┌───────────────┼───────────────┐
↓ ↓ ↓
Model Metrics Scorecard Interpretation
AUC/Gini/KS Points Coefficients
│ │ │
└───────────────┼───────────────┘
↓
┌─────────────────────┐
│ Business Dashboard │
│ Risk vs. Volume │
└─────────────────────┘
The project incorporates concepts from:
- Exploratory Data Analysis
- Binary Classification
- Logistic Regression
- Maximum Likelihood Estimation
- Odds and Log-Odds
- Weight of Evidence
- Information Value
- Credit Scorecards
- ROC Curves
- AUC
- Gini Coefficient
- KS Statistic
- Statistical Significance
- Feature Engineering
- Outlier Treatment
- Multicollinearity
- Classification Trees
- Discriminant Analysis
- MCA
- Correspondence Analysis
- Multidimensional Scaling
- Business Threshold Optimisation
- Model Interpretability
- AI-Assisted Development
The final outcome is an interactive credit-risk analytics workstation that connects statistical theory with practical credit-risk decision-making.
Rather than treating machine learning as a black-box prediction exercise, the project focuses on answering three questions:
Exploratory analysis identifies distributions, relationships, anomalies, and borrower subgroups.
Logistic regression, statistical coefficients, WoE, scorecard concepts, and model metrics provide interpretable explanations.
The decision dashboard translates model probabilities into lending-policy scenarios and demonstrates the relationship between approval volume and portfolio risk.
Predict risk. Understand risk. Explain risk. Make better decisions.
The project demonstrates how interpretable statistical modelling can be combined with interactive analytics to create a credit-risk system that is not only predictive, but also understandable, auditable, and useful for business decision-making.
DataQuest 2026
Domain: Credit Risk Analytics Primary Model: Logistic Regression Application: Interactive Credit Risk Workstation Focus: Interpretability, Exploratory Analysis, Risk Modelling & Decision Support
Artificial intelligence was used during the development of this project as a research, reasoning, explanation, and validation assistant. AI-generated suggestions were critically evaluated by the project team and were not automatically treated as correct. Final methodological, modelling, architectural, and implementation decisions were made based on the project requirements, analytical reasoning, and evaluation of the available data.