Interactive Tableau dashboards for healthcare, risk, and business analytics.

Business question

Which credit card customers are likely to default next month, how should the portfolio be graded and approved, and how much could losses grow in a downturn?

What it shows

• Default rate by risk grade (A to E)
• Information value of each scorecard feature
• Default rate among approved accounts at each cutoff score
• Expected loss under baseline, adverse, and severe scenarios

Result

Grade E defaults at 53.7% against about 10% for grades A and B. A 520 cutoff approves 71.8% of accounts and cuts their default rate from 22.1% to 12.6%; a severe downturn raises expected loss by 50%.

Business question

Can four hospitals train one diagnostic model without sharing patient records, and what does differential privacy cost in accuracy?

What it shows

• Accuracy as privacy noise increases
• Pooled vs federated vs single-hospital accuracy
• Accuracy when each hospital trains alone
• Case mix (malignant share) by hospital

Result

Federated training matched pooled-data accuracy (97.2%) with zero records shared and beat hospitals training alone (95.5% on average). Accuracy falls to 86.2% at the highest noise level.

Business question

How do GRE, GPA, and undergraduate institution rank relate to admission, and how should a shortlisting threshold be set?

What it shows

• Admission rate by institution rank
• Odds ratios from logistic regression
• Admission rate by GPA band
• Precision and recall at each decision threshold

Result

Rank 1 applicants are admitted at 54.1% against 17.9% for rank 4. A 0.40 threshold catches 78% of admitted students for human review; overall predictive power is modest (ROC-AUC 0.67).

Business question

How well can routine clinical measurements identify heart disease, and which measurements matter most?

What it shows

• Disease rate by chest pain type
• Disease rate by number of blocked major vessels
• Held-out test results for k-NN and logistic regression
• Permutation importance of each measurement

Result

A tuned k-NN reached 86.8% accuracy and ROC-AUC 0.95 on unseen patients. Asymptomatic patients had the highest disease rate (72.9%), and blocked vessels carried the most signal.

Business question

Which appointments are most likely to be missed, and where should reminder calls be focused?

What it shows

• No-show rate by booking lead time
• SMS reminder effect within the same lead time
• No-show rate by predicted risk decile
• No-show rate by age group

Result

No-shows rise from 4.6% for same-day bookings to 33.0% at 31+ days. Within the same lead time, SMS reminders lower no-shows; the riskiest model decile misses 41.5% of visits.

Business question

Who survived, and is Cherbourg's higher survival rate real or explained by something else?

What it shows

• Survival rate by class and sex
• Survival rate by port of embarkation
• Class mix by port (the confounder)
• Survival rate by age group

Result

First-class women survived at 96.8% against 13.5% for third-class men. Cherbourg's higher survival is explained by its large first-class share; a tuned random forest reached 83.9% accuracy.

Business question

What drives a used car's price, and how accurately can it be estimated for a dealer?

What it shows

• Median price by vehicle age
• Median price for the 10 most common brands
• Median price by fuel type and transmission
• Median price by kilometres driven

Result

Prices peak around two years of age and then fall steadily. A ridge model on log(price) explains 85.7% of price variation (R² 0.857) with a typical error of $1,740.