Gini, KS, and ROC-AUC measure how well a machine learning credit scoring model separates good borrowers from bad. Gini = 2 × AUC (Area Under the Curve) − 1; KS is the maximum vertical gap between cumulative good and bad distributions; ROC-AUC can be interpreted as the probability that a randomly selected defaulter receives a higher-risk score than a randomly selected non-defaulter, assuming higher scores indicate higher risk.
Lenders in India are under pressure to approve more credit without letting losses spiral. That tension shows up in every model review: “Is this model actually good, or does it just look good on paper?”
This guide walks you through the three metrics that matter most when you validate a machine learning credit scoring model: Gini, KS (Kolmogorov-Smirnov) & ROC-AUC. You’ll see how to calculate each one, what ranges count as “good” in production, and how to explain them clearly to a risk committee.
By the end, you should be able to –
- Derive KS, Gini, and ROC-AUC from score or probability outputs
- Spot common calculation mistakes that inflate performance
- Connect metric results to real decisions on cutoffs and approval strategy
If you’re already using or evaluating automated scoring, Fintly’s AI-powered machine learning scoring platform helps teams build, validate, and monitor these models end to end.
Why these three metrics matter for Indian lenders
For model governance, the model name or algorithm matters less than whether its performance, limitations, data, validation methodology, and monitoring framework are documented and defensible.
Gini, KS & ROC-AUC all measure discriminatory power: the ability to rank higher-risk applicants above lower-risk ones. They do not tell you whether your predicted probabilities are perfectly calibrated, but they do tell you if your ranking is useful.
For Indian banks and NBFCs, these metrics feed directly into –
- Model risk documentation for internal audit and regulators
- Pricing and limit strategies by risk band
- Early warning signals when portfolio quality shifts
A model that cannot separate goods from bads will leak losses no matter how “accurate” its overall hit rate looks.
What is a machine learning credit scoring model, and where do these metrics fit?
A machine learning credit scoring model is a statistical or ML-based system that takes applicant and bureau data and outputs a risk score or probability of default.
Think of it as a more flexible version of a traditional scorecard. Instead of fixed points for each attribute, the model learns patterns from historical data to predict who is more likely to default.
These metrics: Gini, KS, and ROC-AUC, come in after the model is trained. They live in the validation stage, where you test whether the model’s rankings hold up on unseen data.
If you’re new to how these models plug into decisioning, this beginner’s guide to credit scorecard engines in banking explains the broader workflow.
How is the Kolmogorov-Smirnov (KS) statistic calculated mathematically?
KS answers one question: what is the maximum separation between the cumulative distribution of scores for good and bad borrowers?
In practice, you –
- Sort all accounts by score (or predicted probability) from low to high.
- Split them into groups (often deciles or score bands).
- For each group, compute cumulative % of goods and cumulative % of bads.
- Take the absolute difference between cumulative goods and cumulative bads at each point.
- KS is the maximum of those differences, usually expressed as a percentage (0–100).
Note: The calculation is based on the ordering of observations, so monotonic transformations of a score do not change ROC-AUC. However, reversing the score direction changes the interpretation of AUC/Gini.
Mathematically, if FG(s) and FB(s) are the cumulative distribution functions of scores for goods and bads , then –
KS = maxs∣FG(s) − FB(s)∣×100
What counts as a “good” KS?
Industry practice suggests –
- KS < 20: weak separation, often not acceptable for production
- KS 20–40: acceptable for many retail and MSME portfolios
- KS > 40: strong separation, typical of well-built bureau-backed models
Indian lenders with mature analytics teams often target KS in the mid-30s or higher for unsecured retail, depending on data quality and segment.
What is the Gini coefficient, and how do you compute it from ROC-AUC?
The Gini coefficient is another way to express discriminatory power, closely tied to the ROC curve.
In simple terms, Gini tells you how much better your model is than a random ranking. A Gini of 0 means no discrimination; 1 means perfect separation.
Most teams do not compute Gini from scratch. Instead, they derive it from ROC-AUC –
Gini = 2×AUC−1
Where AUC is the area under the ROC curve (explained in the next section).
Typical Gini ranges in production:
- Gini < 0.4: weak model
- Gini 0.4–0.6: moderate
- Gini 0.6–0.7: strong
- Gini > 0.7: excellent, often seen in bureau-driven or highly enriched models
For a Gini coefficient explained simply: imagine two curves. One for goods, one for bads. The more area between them, the higher the Gini.
How does the Gini coefficient relate to ROC-AUC?
ROC-AUC (Area Under the Receiver Operating Characteristic Curve) measures the probability that a randomly chosen defaulter will have a higher risk score than a randomly chosen non-defaulter.
Gini and ROC-AUC are mathematically linked –
- AUC ranges from 0.5 (random) to 1.0 (perfect).
- Gini rescales this to 0–1 using Gini = 2×AUC−1.
- An AUC of 0.75 translates to a Gini of 0.5; an AUC of 0.85 becomes a Gini of 0.7.
Both capture the same underlying ranking quality, but Gini is often preferred in credit risk reporting because it reads more intuitively as “percentage of maximum possible separation achieved.”
Here’s how the three metrics compare in practice –
| Metric | What it measures | Range | Sensitive to cutoff? | Typical “strong” threshold |
| KS | Max gap between cumulative goods & bads | 0–100 | No | > 35–40 |
| ROC-AUC | Probability correct ranking of pairs | 0.5–1.0 | No | > 0.80 |
| Gini | Rescaled version of AUC | 0–1 (or %) | No | > 0.60 |
Risk committees often ask for all three because they reinforce the same story from slightly different angles.
What is a good KS score for a production model in India?
A “good” KS depends on product type, data richness, and vintage stability.
For unsecured personal loans and credit cards using bureau and bank statement data, many Indian lenders see –
- KS in the low 30s as a practical floor for new models
- Mid-to-high 30s as a target for mature, well-tuned models
- 40+ when bureau depth and alternative data are strong
In MSME and small business lending, KS tends to be lower due to noisier financials and higher heterogeneity. A KS in the high 20s to low 30s can still be acceptable if backed by strong policy and monitoring.
Regulatory guidance in India emphasizes robust model validation and ongoing monitoring, even if it doesn’t prescribe exact KS or Gini thresholds.
Illustrative example: Validating a personal loan scorecard for an NBFC
Consider an NBFC that built a machine learning credit scoring model for its personal loan book. On a holdout sample of 10,000 accounts (1,200 bads, 8,800 goods), the risk team groups scores into deciles.
For simplicity, assume the cumulative distributions look like this at the KS peak:
- Cumulative % of goods up to that score: 55%
- Cumulative % of bads up to that score: 85%
Then:
KS = ∣85%−55%∣ = 30
A KS of 30 suggests acceptable separation for an unsecured retail portfolio, assuming stable vintages.
Next, the team computes ROC-AUC from the same holdout and gets 0.78. Using the Gini formula:
Gini = 2×0.78−1 = 0.56
A Gini of 0.56 is moderate-to-strong, consistent with the KS of 30.
The risk committee interprets this as:
- The model is meaningfully better than random.
- It can support risk-based pricing and cutoffs, but not extreme segmentation.
- Ongoing monitoring is needed to ensure KS and Gini don’t decay as the book grows.
This type of validation is an important component of model governance for teams building and monitoring ML-based scoring systems.
Where do teams go wrong? Data leakage, sample bias, and overfitting
Even experienced teams can overstate model performance by making subtle mistakes.
Common pitfalls include –
- Target leakage: Using variables that indirectly encode the outcome (e.g., “days past due in last 3 months” in a model predicting 3+ DPD).
- Sample bias: Training only on approved accounts without correcting for rejection inference. Because outcomes are generally observed only for applicants who were approved and originated, the performance data may not represent applicants who were rejected, creating a selection problem known as reject inference.
- Mixing development and validation data: Reporting metrics on the same dataset used for tuning.
- Ignoring vintage effects: Validating on early vintages that haven’t had time to mature.
These errors can push KS and Gini artificially high, only for performance to collapse in production.
Contrary view: Automation isn’t always the answer. In thin-file segments or new products, a simple rules-based approach with manual review may outperform a complex model that looks good on backtests but fails in the wild. The key is honesty about data quality and business context.
How to present Gini, KS, and ROC-AUC to a risk committee
Risk committees don’t need derivations. They need a clear story: how well the model separates risk, how stable it is, and what that means for losses and approvals.
A clean one-page summary might include –
- KS, Gini, and ROC-AUC on development, validation, and latest vintage
- Stability charts showing metric trends over time
- A brief note on any known limitations (e.g., new product, limited bads)
- Recommended actions: proceed, proceed with conditions, or rework
Tie metrics to business impact. Illustrative example: “At a KS of 32, shifting the cutoff by 20 points could change the approval rate by ~5% and expected loss rate by ~0.8%, depending on the portfolio and score distribution.”
For teams implementing scorecard-based decisioning, this guide on optimizing credit approval rates using scorecard cutoff strategies shows how to translate model performance into policy.
Key takeaways
- Gini, KS, and ROC-AUC all measure discriminatory power, but each tells the story in a slightly different way.
- In Indian retail and MSME lending, KS in the 30s and Gini above 0.6 are common targets for strong models, but context matters more than hard thresholds.
- Clean data, proper validation samples, and honest governance matter more than chasing the highest possible metric.
- These metrics should drive concrete decisions on cutoffs, pricing, and monitoring, not just sit in a model pack.
Conclusion: What good looks like for Indian lenders in 2026
Good model validation isn’t about impressive numbers. It’s about confidence that your machine learning credit scoring model will keep separating good from bad as the portfolio scales.
If your KS, Gini, and ROC-AUC are stable across vintages and segments, and your team can explain them in plain language, you’re in a strong position.
If you’re building or refreshing your scoring stack and want a second pair of eyes on your metrics and workflows, you can contact Fintly to discuss how we support banks, NBFCs, and fintechs with end-to-end ML scoring and validation.
QUICK ANSWERSFrequently Asked Questions
Your most common questions, answered with precision and insight
There is no universal good KS score. In some credit-scoring environments, values in the 30s may be viewed as useful discriminatory performance, but acceptability depends on portfolio, product, sample, default rate, benchmark models and stability over time.
The Gini coefficient is a rescaled version of ROC-AUC, calculated as Gini = 2 × AUC − 1. While AUC ranges from 0.5 to 1.0, Gini ranges from 0 to 1, making it easier to interpret as “percentage of maximum possible discrimination achieved” by the model.
The Kolmogorov-Smirnov (KS) statistic is the maximum absolute difference between the cumulative distribution of scores for good and bad borrowers. You compute cumulative % goods and bads across score bands, then take the largest gap between the two curves as the KS value.
There is no universal Gini target for every credit model. Values above 0.6 are sometimes used as a practitioner benchmark for strong discrimination, but the appropriate target depends on the portfolio, data available, model purpose and comparable historical models.
KS and Gini usually move together because both reflect discriminatory power, but they can appear to conflict if one metric is computed on a different sample or time window. Always ensure KS, Gini, and ROC-AUC are measured on the same validation dataset before drawing conclusions.
Avoid data leakage by excluding any variables that directly or indirectly encode the outcome, using only pre-application or truly exogenous features. Validate metrics on a holdout sample that was never used for training or hyperparameter tuning & check for vintage maturity before reporting results.
Discrimination measures how well a model ranks higher-risk borrowers above lower-risk ones (captured by Gini, KS, ROC-AUC). Calibration measures how close predicted probabilities are to actual default rates. A model can discriminate well but still be poorly calibrated, requiring separate checks and adjustments.
There is no universal quarterly rule for every lender. Monitoring frequency should be defined by the model-risk framework, portfolio characteristics, model materiality and regulatory requirements. Active retail portfolios may be monitored monthly or quarterly, with more frequent review during periods of rapid growth, material drift or economic stress.
Author
Subject Matter Expert (Lending) Fintly.co
Vijay Mali is a results-driven professional with deep expertise in HFC/NBFC startups, compliance, and underwriting. He specializes in delivering end-to-end solutions for financial institutions, focusing on Business Rule Engines (BRE), workflow automation, and AI-driven credit decision-making. He is passionate about leveraging Machine Learning (ML) scorecards and AI-powered risk assessment to optimize lending processes and drive digital transformation in the financial sector.
