Loan Dataset Cleaning

A pandas/NumPy pipeline that takes a genuinely messy loan-applicant dataset — duplicates, typos, and all — through 7 concrete cleaning steps to something a model could actually trust.

Trail statsTelemetry · at a glance

7
cleaning steps
13
columns in the dataset
615 → 535
rows before and after cleaning
p10/p90
outliers capped at these percentiles instead of dropping rows

01 · The Map · Mission Brief

Brief

A loan-applicant dataset (13 columns — marital status, dependents, income, loan amount, term, credit score, approval status, and more) arrives the way real data always does: duplicated rows, inconsistent formatting, missing values, and a few outliers that make you question the data-entry process. This notebook walks it, step by step, into something actually safe to analyze or model: 615 rows in, 535 out, and every departure accounted for.

Don’t take my word for it — the receipts are on GitHub

02 · The Pack · Crew Duties

My Contribution

Dropped duplicate rows, both across the full row and specifically on the UID column.
Standardized string fields — uppercased marital status, and normalized the Sex column’s ‘M’/’F’ shorthand into ‘Male’/’Female’.
Corrected invalid records — replaced impossible Age values (anything below 1, negatives included) with a minimum valid age of 20.
Handled missing values: mean-imputed Loan_amount, Term_months, and Age; filled missing Is_graduate with ‘Graduate’; dropped rows with any remaining gaps.
Converted Marital_status, Sex, and Is_graduate to categorical dtypes and Income to float.
Capped Income outliers at the 10th and 90th percentiles rather than dropping them.
Dropped the irrelevant Hobby column before finalizing the dataset.

03 · Sunrise · Planet-rise

Takeaways

Cleaning a real dataset column-by-column like this is a good reminder of how much judgment is hidden in "just clean the data" — whether to cap or drop outliers, whether missing Is_graduate should default to Graduate or Not Graduate, changes the dataset in ways a model will absolutely pick up on.

Previous waypointPrevious burn · 9 of 10

Twitter Bot Detection EDA

The statistical tells that separate bots from people

Next waypointNext burn · 1 of 10

GlyphScholar

Multimodal RAG that cites the exact page

All Projects