Look before you leap

Exploratory Data Analysis (EDA) is the detective phase — before any modeling, you get to know your data: its shape, its quirks, what's missing, and how variables relate. Skipping it is the classic rookie error, because every later step inherits whatever you didn't notice. An afternoon of EDA routinely saves days of debugging a model that was doomed by a problem visible in the raw data.

First passes that always pay off

A few commands give you a fast read on almost any dataset:

df.shape                   # how big
df.head()                  # what the rows look like
df.describe()              # summary stats for numeric columns
df.isna().sum()            # missing values per column
df.corr(numeric_only=True) # linear relationships between columns

These answer the essential opening questions: how much data, what a row means, what's typical, where the holes are, and what moves together.

What to look for

• Distributions — is a column bunched, spread, or heavily skewed? Skew changes which summaries and models make sense.
• Outliers — extreme values that are either errors to fix or the most interesting cases to study.
• Missing data — how much, and why? Missing-at-random differs from "the sensor fails when it's cold."
• Relationships — which variables correlate, and do any surprise you?

Pair numbers with pictures

Summary statistics can hide what a quick plot reveals — two datasets can share a mean and look completely different. So EDA is always numbers and charts: describe, then histogram and scatter.

EDA isn't a formality before the "real" work — it is a lot of the real work. You can't honestly model data you haven't looked at.

Try this: On any dataset, run describe() and isna().sum(), then pick the most surprising thing you see — a wild maximum, a half-empty column — and investigate why. That instinct to chase the surprise is the core habit of a good analyst.