Pandas groupby Explained: From Basics to Real Analyst Workflows

If you learn one thing in pandas well, make it groupby. Almost every real analysis — revenue by region, average by category, retention by cohort — is a groupby underneath. This guide builds from the absolute basics to the patterns working analysts use daily.

The mental model (start here)

Every groupby is three steps, often called split-apply-combine:

  1. Split the rows into groups by some key (e.g. region).
  2. Apply a function to each group (e.g. sum the sales).
  3. Combine the results back into one table.

Once you see every problem this way, groupby stops being confusing.

Beginner: your first groupby

To get total sales per region:

sales.groupby("region")["amount"].sum()

Read it as: split by region, take the amount column, sum each group. Swap .sum() for .mean(), .count(), .max(), .min() as needed. This one line covers a huge share of everyday analysis.

To sort the result (most common follow-up):

sales.groupby("region")["amount"].sum().sort_values(ascending=False)

Intermediate: multiple keys and multiple metrics

Group by more than one column — sales per region and product:

sales.groupby(["region", "product"])["amount"].sum()

Compute several metrics at once with named aggregations (the clean, modern way):

sales.groupby("region").agg(
    total=("amount", "sum"),
    orders=("order_id", "count"),
    avg_order=("amount", "mean"),
)

This produces a tidy table with three named columns — exactly what you'd hand to a stakeholder.

The concept that levels you up: transform

groupby().sum() collapses each group to one row. But often you want to keep every row and attach a group-level value — for example, each order alongside its region's average. That's transform:

sales["region_avg"] = sales.groupby("region")["amount"].transform("mean")
sales["vs_avg"] = sales["amount"] - sales["region_avg"]

transform returns a result the same length as the original, so it aligns row-by-row. This is how you compute "above/below the group average", share of group total, and z-scores within a group.

Advanced: real analyst patterns built on groupby

Share of total — what fraction each category contributes:

by_cat = sales.groupby("category")["amount"].sum()
share = (by_cat / by_cat.sum() * 100).round(1)

Top-N per group — the best product in each region — combines groupby with sorting or ranking.

Cohort analysis — group customers by the month of their first purchase, then count how many stay active each following month. It's groupby on a derived "cohort" key.

RFM segmentation — group by customer to compute recency, frequency and monetary value, then bucket them. Again: groupby at its core.

Time-based grouping — group by month or week using the datetime accessor:

sales.groupby(sales["date"].dt.to_period("M"))["amount"].sum()

Common mistakes to avoid

  • Forgetting to reset the index when you need a flat table: add .reset_index().
  • Chained assignment warnings — assign transform results back to a column directly, as shown above.
  • Silently dropping NaN groups — groupby ignores NaN keys by default; pass dropna=False if you need them.
  • Using apply when agg/transform would do — apply is flexible but slow; prefer the built-ins for speed and clarity.

How to get fluent

groupby fluency is pure practice. Take a real dataset and answer twenty questions with it — totals, averages, per-group ranks, shares, cohorts. After a few dozen reps, the split-apply-combine pattern becomes automatic and most analysis questions feel routine.

The bottom line

groupby is the workhorse of pandas. Master the three-step model, learn the difference between agg (collapse) and transform (keep shape), and you'll be able to answer the large majority of analytical questions with a couple of lines of code.