RETAIL BANK · CANADA
2024
Fraud Detection System
Logistic regression model on labelled transaction history with engineered behavioural features — improved early detection and reduced false positives.
AnalyticsRisk
Selected
work
CREDIT UNION
2023
Credit Approval & Client Segmentation
AnalyticsBanking
Live demo · Try it
Fraud Detection Scorer
Enter a transaction and get an instant fraud probability with a recommended action. Powered by a logistic regression with engineered features (amount × night, amount × distance, velocity, online category) — test AUC ≈ 0.999on a held-out sample of 50,000 transactions.
< 0.5% Low0.5–9.1% Medium> 9.1% High
How it works
- 1. A LightGBM classifier was trained on simulated card transactions (Sparkov-style schema, ~2.4% fraud base rate).
- 2. Top drivers — amount, distance from home, hour of day, velocity and merchant category — were distilled into a logistic surrogate so the model runs entirely in your browser.
- 3. The output is the estimated probability of fraud, mapped to an action: approve, step-up, or decline.
Demo only · Predictions are derived from a trained model on a simulated dataset and are not a substitute for production fraud controls.
Model performance · held-out test set
Sparkov-style simulated card transactions · 200,000 rows · 2.5% fraud base rate
AUC
0.999
Precision @ High Risk
31.1%
Recall @ High Risk
99.6%
F1
0.474
Confusion matrix · High Risk threshold
| Pred negative | Pred positive | |
|---|---|---|
| Actual neg | 45,995 | 2,755 |
| Actual pos | 5 | 1,245 |
Sample
- Test rows
- 50,000
- Positive base rate
- 2.5%
- Avg 95% CI half-width
- ±2.3pp
Batch mode · CSV
Score a whole transaction file
Upload a CSV of transactions and get a fraud probability + recommended action for every row. All processing happens locally in your browser — nothing is uploaded to a server.
Required columns (6)
amt· Amount ($) (number)category· Merchant category (string)hour· Hour (0–23) (number)distanceKm· Distance from home (km) (number)trans24h· Txns in last 24h (number)age· Cardholder age (number)
Live Demo · Try it
Credit Approval Risk Scorer
A working logistic regression model trained on 30,000 real credit card client records. Enter applicant details to estimate the probability of default and see the recommended approval decision.
- Low Risk · < 15%
- Auto-approve at standard terms.
- Medium Risk · 15–40%
- Manual review, lower limit or guarantor.
- High Risk · > 40%
- Decline or require collateral.
Model performance · held-out test set
UCI Default of Credit Card Clients (Taiwan, 2005) · 30,000 rows · 22% default base rate
AUC
0.739
Precision @ High Risk
56.8%
Recall @ High Risk
38.5%
F1
0.459
Confusion matrix · High Risk threshold
| Pred negative | Pred positive | |
|---|---|---|
| Actual neg | 5,355 | 486 |
| Actual pos | 1,020 | 639 |
Sample
- Test rows
- 7,500
- Positive base rate
- 22.1%
- Avg 95% CI half-width
- ±2.6pp
Batch mode · CSV
Score a whole applicant file
Upload a CSV of credit card applicants and get a default probability + recommended decision for every row. All processing happens locally in your browser — nothing is uploaded.
Required columns (10)
limitBal· Credit limit (NT$) (number)age· Age (number)sex· Sex (1=M, 2=F) (string)marriage· Marriage (1/2/3) (string)education· Education (1–4) (string)billAmt1· Current bill (NT$) (number)payAmt1· Last payment (NT$) (number)pay0· Repayment status — last month (number)pay2· Repayment status — 2 months ago (number)pay3· Repayment status — 3 months ago (number)