Machine learning

Credit Card Default Prediction

hard4–5 hours2 datasetsPremium

The whole brief is open to read. This lab's datasets come with Premium.

Company: Capital One
Job positions: Data ScientistML EngineerData Analyst
Topics: ClassificationCredit riskData leakageExpected profitscikit-learn

The scenario

Capital One: Build a model that predicts which credit card applicants will default, decide who to approve, and explain each rejection.

When somebody applies for a Capital One credit card, the company decides whether to approve them, knowing some will stop repaying, which is what defaulting means. A rejected applicant is entitled by law to the main reasons why, which lenders call adverse action reasons, and age cannot be one of them. You're a data scientist on the credit policy team, which wants to approve more applicants without more loss.

Your task

Predict each applicant's chance of defaulting, meaning 90 or more days past due in the first year, set the approval cutoff that makes the most profit, and give reasons for every rejection.

Instructions

1Load and clean the applications. Check for repeated records and missing values, and explain how you handled each.
2Decide which columns a decision could actually use. Some are filled in only after an account has been opened, and the decision cannot rest on an applicant's age.
3Explore how often accounts default, and how that rate moves with the bureau score, balances, recent credit checks and past late payments.
4Build a baseline from the bureau score alone, then train and compare at least two models on area under the ROC curve, or AUC, validating on later applications than you trained on.
5Choose an approval cutoff from expected profit. An approved account that keeps paying is worth $420 in its first year, and one that defaults costs 55 percent of the credit line the applicant asked for.
6For every application you would reject, give the main reasons behind it, ranked, in words the applicant would understand.
7Score applications_test.csv and save predictions.csv with the columns application_id, default_probability, decision (approve or decline) and decline_reasons, left empty for approvals.

Datasets

All files come in one download.

applications_train.csv

About 15,000 card applications from January to September 2024, each with how the account went in its first year.

15,207 rows · 21 columns · 1.4 MB

Column	Type	Description
application_id	integer	Unique ID of the application.
applied_on	date	Date the application was submitted.
applicant_age	integer	The applicant's age in years.
annual_income_usd	integer	Income the applicant stated, in US dollars. Zero for some retired applicants, who reported no earned income.
employment_type	text	salaried, self_employed, contract or retired.
employment_years	integer	Years with the current employer. Empty where the applicant left it blank.
housing	text	mortgage, rent, own or other.
state	text	US state the applicant lives in.
zip_median_income_usd	integer	Median household income of the applicant's ZIP code, from census data.
bureau_score	integer	Credit bureau score from 300 to 850. Empty for applicants with too little credit history to score.
credit_history_months	integer	Months since the applicant's oldest credit account was opened.
open_accounts	integer	Credit accounts currently open.
total_credit_limit_usd	integer	Total limit across the applicant's existing revolving accounts.
revolving_balance_usd	integer	Balance currently carried on those accounts.
inquiries_last_6m	integer	Credit checks run on the applicant in the last six months.
delinquencies_last_24m	integer	Payments 30 or more days late in the last two years.
months_since_last_delinquency	integer	Months since the applicant was last late. Empty means they have never been late.
requested_credit_line_usd	integer	Credit line the applicant asked for.
assigned_credit_line_usd	integer	Credit line the account was given. Set when the account is opened.
first_year_spend_usd	integer	Spend on the account over its first year.
ever_90dpd_12m	integer	1 if the account went 90 or more days past due within twelve months, otherwise 0. This is what you predict.

Preview the first 5 rows

application_id	applied_on	applicant_age	annual_income_usd	employment_type	employment_years	housing	state	zip_median_income_usd	bureau_score	credit_history_months	open_accounts	total_credit_limit_usd	revolving_balance_usd	inquiries_last_6m	delinquencies_last_24m	months_since_last_delinquency	requested_credit_line_usd	assigned_credit_line_usd	first_year_spend_usd	ever_90dpd_12m
700001	2024-01-01	42	65000	self_employed	19	mortgage	FL	69500	619	179	4	24100	16912	4	0	empty	12000	5300	5750	1
700002	2024-01-01	62	20808	retired	0	rent	VA	50500	698	299	4	11900	3708	4	4	7	12000	10500	3950	0
700003	2024-01-01	41	73000	salaried	3	own	NY	97500	592	144	6	11900	6418	2	0	62	12000	11700	5500	0
700004	2024-01-01	47	117500	salaried	15	own	IL	112500	724	267	5	28500	0	0	0	empty	6000	5400	2950	0
700005	2024-01-01	42	83000	salaried	2	rent	TN	126000	682	142	3	37200	30610	3	0	80	2500	2000	900	0

applications_test.csv

About 5,000 applications from October to December 2024, without the outcome.

5,022 rows · 20 columns · 469 KB

Column	Type	Description
application_id	integer	Unique ID of the application.
applied_on	date	Date the application was submitted.
applicant_age	integer	The applicant's age in years.
annual_income_usd	integer	Income the applicant stated, in US dollars.
employment_type	text	salaried, self_employed, contract or retired.
employment_years	integer	Years with the current employer. Empty where the applicant left it blank.
housing	text	mortgage, rent, own or other.
state	text	US state the applicant lives in.
zip_median_income_usd	integer	Median household income of the applicant's ZIP code, from census data.
bureau_score	integer	Credit bureau score from 300 to 850. Empty for applicants with too little credit history to score.
credit_history_months	integer	Months since the applicant's oldest credit account was opened.
open_accounts	integer	Credit accounts currently open.
total_credit_limit_usd	integer	Total limit across the applicant's existing revolving accounts.
revolving_balance_usd	integer	Balance currently carried on those accounts.
inquiries_last_6m	integer	Credit checks run on the applicant in the last six months.
delinquencies_last_24m	integer	Payments 30 or more days late in the last two years.
months_since_last_delinquency	integer	Months since the applicant was last late. Empty means they have never been late.
requested_credit_line_usd	integer	Credit line the applicant asked for.
assigned_credit_line_usd	integer	Credit line the account was given. Set when the account is opened.
first_year_spend_usd	integer	Spend on the account over its first year.

Preview the first 5 rows

application_id	applied_on	applicant_age	annual_income_usd	employment_type	employment_years	housing	state	zip_median_income_usd	bureau_score	credit_history_months	open_accounts	total_credit_limit_usd	revolving_balance_usd	inquiries_last_6m	delinquencies_last_24m	months_since_last_delinquency	requested_credit_line_usd	assigned_credit_line_usd	first_year_spend_usd
715194	2024-10-01	34	96500	contract	11	rent	SC	133000	594	111	1	20800	6920	2	0	47	12000	12200	6950
715195	2024-10-01	53	43000	salaried	13	mortgage	TN	76000	680	260	2	18000	15396	0	0	empty	2500	2800	1000
715196	2024-10-01	48	146000	salaried	21	rent	AL	178500	771	173	6	48700	12812	1	0	empty	6000	5000	2750
715197	2024-10-01	35	55500	salaried	15	rent	TX	70500	576	177	3	12300	7692	2	1	6	8000	2800	2250
715198	2024-10-01	49	102000	salaried	empty	mortgage	NY	126000	724	168	5	22700	0	4	0	empty	4000	4100	1400

Hints

A column that only exists once an account has been opened cannot help you decide a new application, however well it predicts the outcome. Check each one against the moment the form is submitted.

The reason for a decline comes from that applicant's own contribution to their score, not from which features matter most across everybody.

contribution = coefficients * applicant_scaled
contribution.sort_values(ascending=False).head(3)

Deliverable

A public GitHub repo with your analysis (a notebook or scripts), predictions.csv for the test applications, and a README with the columns you excluded, your model comparison and the cutoff you chose.

When you're done, post your repo in the Solutions tab to share it with other learners.

What grading checks

Use this checklist to review your own work before you post and share it.

Submitted GitHub repo is public and reachable.
Repo contains at least one notebook or script file.
predictions.csv has a probability, a decision and reasons for every test application.
default_probability reaches an area under the ROC curve of at least 0.81 on the test applications.
Under the cost assumptions in the brief, the approved applications are worth more than approving every applicant.
Columns that only exist after an account is opened are excluded with the reason explained, and the applicant's age is excluded from the decision.
Models are compared against a bureau score baseline, and decline reasons differ between applicants rather than repeating one list.