Each card below is a standalone data story. Use the tags as mental filters.
Retail • Marketing & Privacy
How Target Predicted Pregnancy Before Anyone Was Told
Data used: purchase history • Model: propensity scoring
Small shifts in shopping — unscented lotion, vitamins, cotton products —
quietly roll into a model that scores the likelihood a customer is pregnant.
Even a modestly accurate prediction, deployed across millions of shoppers,
changes how coupons and offers are targeted.
Streaming • Recommendation Systems
Netflix and the $1M Bet on Ensemble Models
Data used: viewing history • Model: ensembles, collaborative filtering
Netflix improved its movie recommendation accuracy by blending over a hundred
models into one “ensemble”. No single algorithm won; the crowd of
models did — proving that many good predictors, combined, beat one great one.
Banking • Risk & Default
How a Bank Scores Mortgage Risk Before You Miss a Payment
Data used: credit, income, behavior • Model: risk scoring
A risk model doesn’t ask “Will this person default?” It asks,
“How likely is default relative to everyone else?” By ranking
borrowers by risk, the bank can adjust terms, set limits, and intervene early
— long before a crisis shows up on the books.
AI • Natural Language
Watson on Jeopardy: How a Machine Learned to Buzz In
Data used: text corpora • Model: ensemble NLP scoring
IBM Watson didn’t “know” trivia. It generated candidate answers,
scored each one by probability, and only buzzed in when its confidence crossed
a threshold. Under the hood: layers of models voting on how likely each answer
was to be right.
Workforce • Retention
Predicting Who Will Quit: Inside Workforce Analytics
Data used: tenure, performance, behavior • Model: churn prediction
By treating employee turnover like customer churn, HP built models that estimated
which employees were most likely to leave. That gave managers a shortlist for
early conversations, targeted incentives, and retention programs.
Healthcare • Outcomes
The 30-Day Readmission Score That Saves Lives
Data used: labs, vitals, history • Model: risk stratification
Hospitals assign a risk score for who is likely to be readmitted within 30 days.
That one number guides follow-up calls, home visits, and extra monitoring,
turning raw EHR data into targeted care that can reduce both cost and harm.
Public Safety • Ethics
Predicting Crime Hotspots: Where Models Cross the Line
Data used: incident logs, geography • Model: hotspot forecasting
Police departments have tested algorithms that forecast where crimes are likely
to occur. The upside is better resource allocation; the downside is feedback
loops and fairness concerns when past bias gets baked into future patrols.
Finance • Fraud Detection
Catching the One Bad Swipe in a Million
Data used: transaction stream • Model: anomaly & risk scoring
Every card transaction gets a fraud score in milliseconds. The model considers
location, amount, history, and patterns across millions of cards. Most scores are low.
A few spike. Those are the ones that trigger a text, a decline, or a call.
Campaigns • Persuasion
The Voter Persuasion Score Behind Modern Campaigns
Data used: demographics, history, surveys • Model: uplift / persuasion
Campaigns don’t just predict who will vote. They model who can be persuaded by an ad
or a knock on the door. That “uplift” prediction tells them where every dollar
and volunteer hour is most likely to change minds.