July 10, 2026 · 5 min read

EDA checklist: the questions I ask every new dataset

A repeatable first-hour routine for meeting a new dataset — before any modeling happens.

Data Analysis · Learning in Public

This is a SAMPLE article demonstrating the writing layout. Replace it with a real post in src/data/articles.ts.

Every dataset lies to you at first. Not maliciously — but missing values hide behind sentinel numbers, categories are misspelled, and dates arrive in three formats. Before I trust any chart or model, I run through the same short checklist.

1. What does one row mean?

The single most important question. Is a row a customer, a transaction, a customer-month? Getting this wrong quietly breaks every aggregation that follows.

2. Audit the basics

df.info()
df.describe(include="all")
df.isna().mean().sort_values(ascending=False)

Row and column counts vs. what the documentation claims
Missing values — and whether they're truly missing or encoded as 0 / -1 / 'unknown'
Duplicates on the supposed primary key
Ranges that are physically impossible (negative ages, future dates)

3. Look at distributions before correlations

Skew, outliers, and class imbalance change which methods are appropriate. A correlation matrix on wildly skewed features mostly measures the outliers.

None of this is glamorous, and that's the point: the first hour with a dataset is about earning the right to make claims later.