What a Great Notebook Walkthrough Actually Sounds Like

Interview Insights · Data Analytics

What a Great Notebook Walkthrough Actually Sounds Like

SECTION 01Open with the question, then skip ahead on purpose

The first thirty seconds decide whether the interviewer follows you or drifts. A great walkthrough opens on a markdown cell that states the question in one sentence, who it is for and what decision it supports. For example: 'Do customers who join in festive months stay as long as customers who join in other months? The marketing team wanted to know whether festive campaigns bring loyal customers or one-time buyers.'

After that, a strong candidate says something like: 'I will skip the imports and loading, and spend most of our time on the retention analysis in section three.' That one sentence shows control. You are not hiding the early cells, you are offering them. If the interviewer wants to see cleaning, they will ask, and you will go there by choice instead of reading through it by habit.

# Festive-season cohorts: do they stay?
**Question:** Do customers acquired in festive months
retain as well as customers acquired in other months?
**For:** Marketing team planning next year's campaigns.
**Decision supported:** How much budget to give festive
acquisition versus retention offers.

## 1. Data and cleaning
## 2. Building monthly cohorts
## 3. Retention by cohort  <- main analysis
## 4. Findings, limits and next steps

"This notebook answers one question for the marketing
team: do customers who join during festive months stay
as long as others?

I will cover the data quickly, then spend most of the
time in section 3, the retention table and heatmap,
because that is where the answer is.

Happy to open the cleaning steps if you want to see them."

Numbered markdown headers also help during a screen share. With the table of contents panel open in JupyterLab, you can jump to section 3 in one click instead of scrolling for half a minute while the interviewer waits.

SECTION 02Narrate the key cell as decisions, not syntax

Once you reach the core analysis, resist the urge to read the code. The interviewer can see the code. What they cannot see is why it is shaped that way. Narrate the cell as a short list of decisions: what each row represents, how the groups were formed and what the output means.

Here is a cohort retention cell for the festive question. A cohort is the group of customers who made their first purchase in the same month. For each cohort, we count how many were still active one, two, three months later. A great narration covers three points: how the first purchase month was found, how the month offset was calculated, and why the result is shown as a share of the cohort size rather than raw counts.

import pandas as pd

orders = pd.read_csv('data/orders.csv', parse_dates=['order_date'])
orders['order_month'] = orders['order_date'].dt.to_period('M')

# Cohort = month of each customer's first order
orders['cohort'] = (orders
                    .groupby('customer_id')['order_month']
                    .transform('min'))

# Months since joining (0 = joining month)
orders['offset'] = (orders['order_month'] - orders['cohort']).apply(lambda d: d.n)

# Unique active customers per cohort and offset
counts = (orders
          .groupby(['cohort', 'offset'])['customer_id']
          .nunique()
          .unstack(fill_value=0))

# Share of each cohort still active
retention = counts.div(counts[0], axis=0).round(3)
retention.head()

"Each customer is placed in the month of their first
order. That is their cohort.

For every later order I calculate how many months after
joining it happened, so the joining month is zero.

Then I count unique customers, not orders, at each offset,
because one loyal customer with many orders should not
look like many customers.

Finally I divide by the cohort size, so a large October
cohort and a small March cohort can be compared directly."

That narration takes under a minute and contains no syntax at all, yet it proves you understand every line. If the interviewer wants syntax, they will ask about transform or the period arithmetic, and you can answer from a position of control.

SECTION 03Read the chart aloud, then close on what it cannot prove

The weakest sentence in any notebook walkthrough is 'and this is the chart'. A chart does not speak for itself in an interview, especially over a shared screen where the text may be small. Tell the interviewer how to read it: what the axes are, where to look, and what the pattern means for the question.

Then close properly. A great walkthrough ends by stating the limits of the analysis. Cohort retention shows a pattern, but it does not prove the festive campaign caused it. Festive cohorts might include more discount-driven buyers, or a different product mix. Saying that clearly, and naming the next test you would run, ends the tour on judgement rather than on a picture.

import matplotlib.pyplot as plt

fig, ax = plt.subplots(figsize=(9, 6))
im = ax.imshow(retention.iloc[:, :7], cmap='Blues', vmin=0, vmax=1)

ax.set_xticks(range(7))
ax.set_xticklabels([f'M{m}' for m in range(7)])
ax.set_yticks(range(len(retention)))
ax.set_yticklabels(retention.index.astype(str))
ax.set_xlabel('Months since first order')
ax.set_ylabel('Cohort (first order month)')
ax.set_title('Share of cohort still ordering')
fig.colorbar(im, ax=ax, label='Retention share')
plt.tight_layout()
plt.show()

## 4. Findings, limits and next steps

**Finding:** In this data, festive-month cohorts
retained less after the second month than other cohorts.

**Limits:**
- This shows association, not cause.
- Festive cohorts may differ in discount use and
  product mix.
- Recent cohorts have fewer months of history.

**Next step:** Compare festive customers with and
without discount codes, and propose a small retention
offer test for the next festive cohort.

Notice that the closing cell does not claim more than the data allows. Saying 'in this data' and 'association, not cause' shows statistical maturity. Then the next step turns the limit into a plan, which is the note you want the interviewer to remember.

SECTION 04Turn your own notebook into a five-minute tour

Prepare once, and every future walkthrough becomes easier. These tabs help you restructure the notebook, plan the tour and rehearse it.

BEFORE THE INTERVIEW, ADD TO YOUR NOTEBOOK:

[ ] Title cell with question, audience and decision
[ ] Numbered section headers (1, 2, 3, 4)
[ ] One markdown line under each header: what this does
[ ] Mark the main analysis section clearly
[ ] Move dead-end exploration to an appendix section
[ ] Final cell: findings, limits, next steps
[ ] Restart kernel and run all, so outputs are fresh

MY 5-MINUTE TOUR

Stop 1 (0:00) Question: ______________________
Stop 2 (0:30) Data and grain: ________________
              One cleaning choice: ____________
Stop 3 (1:15) Key cell: section __ cell __
              Narration: 'Each ____ ...'
              Chart reading: axes / look at / means
Stop 4 (4:00) Limit 1: ____  Limit 2: ____
              Next step: ______________________
Hand-off line: 'Would you like to see ____?'

REHEARSE THREE TIMES

Run 1: alone, with a timer. Note where you go over.
Run 2: screen share to a friend. Ask them:
       - What was the question?
       - What was the answer?
       - What can the analysis NOT tell us?
Run 3: same friend interrupts twice with
       'Why did you do that?' at random cells.

Ready when your friend answers all three
questions correctly after Run 2.

SECTION 05Habits that make a notebook tour feel effortless

These small habits make a big difference when you share your screen.

• Zoom in before sharing — Increase the browser zoom so code and chart labels are readable on the interviewer's screen. Small text forces them to guess.
• Collapse long outputs — Hide huge DataFrame prints and warnings before the interview, so scrolling to the key section is quick and clean.
• Point with words — Say 'the second row from the top' or 'the column labelled M3' instead of 'here' or 'this one', because your cursor may not be visible.
• Keep one number per finding — Give a single number that supports each finding, rather than a list. It is easier to remember and easier to defend.
• Use an appendix — Move exploratory charts to an appendix at the end. You can still show them if asked, but they will not clutter the main tour.

SECTION 06Interview Q&A — Notebook walkthrough

SECTION 07Test yourself — Notebook walkthrough quiz

SECTION 08Frequently asked questions

SECTION 09Related reads