What Happens When a Candidate Cannot Explain Their Notebook

Interview Insights · Data Analytics

What Happens When a Candidate Cannot Explain Their Notebook

SECTION 01The moment the notebook stops counting as your work

Here is the honest sequence from my side of the table. I open the notebook you linked, scroll past the imports and stop at the first cleaning cell. I ask a simple question: 'Why did you fill missing values with zero here?' If you give a reason, even an imperfect one, we have a conversation about your choices. If you say 'that removed the error' or 'I am not sure, I followed a tutorial', I make a quiet note and move to a second cell.

One unexplained cell is forgivable. Everyone copies a line now and then. Two in a row form a pattern, and at that point I stop using the notebook as evidence. The rest of the round turns into basic questions, such as the difference between loc and iloc or how merge works, because I can no longer assume you know them. That shift is what candidates feel as the interview 'going cold'.

df = pd.read_csv('orders.csv')
df = df.drop(['col7', 'Unnamed: 0', 'discount'], axis=1)
df = df.fillna(0)
df = df[df['amount'] < 50000]
df['month'] = pd.to_datetime(df['date']).dt.month

# Interview questions this invites:
# Why drop 'discount'? Why fill every column with 0?
# Where does 50000 come from? Month without year?

df = pd.read_csv('orders.csv', index_col=0)  # first column is the old index

# 'discount' is blank for most rows and only filled from
# a later system version, so I keep it out of the trend analysis.
df = df.drop(columns=['discount'])

# Missing 'returns' means no return was logged -> 0 is correct.
# Missing 'city' is unknown, so I label it instead of zero.
df['returns'] = df['returns'].fillna(0)
df['city'] = df['city'].fillna('Unknown')

# Orders above the 99.9th percentile were bulk B2B orders,
# confirmed by their customer type, so I analyse them separately.
cap = df['amount'].quantile(0.999)
bulk = df[df['amount'] > cap]
df = df[df['amount'] <= cap]

df['order_month'] = pd.to_datetime(df['date']).dt.to_period('M')

The second version is longer, but every line now answers its own interview question. Notice also the fix at the end: taking only .dt.month merges January of two different years, while to_period('M') keeps year and month together. Bugs like that are what I find when I probe a cell, and finding them yourself first is much better.

SECTION 02The cells interviewers probe first, and why

Interviewers do not pick cells at random for long. With experience, you learn where weak notebooks hide their problems. The imports and the final chart are rarely interesting. The interesting parts are the decisions in the middle: how missing values were handled, which rows were removed, how tables were joined, and how a metric was defined.

These are the places where a copied notebook shows its seams, because the person who wrote the tutorial made those decisions for their own data, not yours. If your notebook drops the same columns or uses the same thresholds as a popular tutorial on a different dataset, the cell cannot be explained, because there is no reason behind it in your context. Prepare each of the cell types below before you share any notebook.

orders = pd.read_csv('orders.csv')
customers = pd.read_csv('customers.csv')

print('orders before:', len(orders))

# Each order has one customer; each customer appears once.
# validate raises an error if that assumption is false.
merged = orders.merge(
    customers,
    on='customer_id',
    how='left',
    validate='many_to_one',
    indicator=True,
)

print('after merge:', len(merged))
print(merged['_merge'].value_counts())
# 'left_only' rows are orders with no matching customer.

Interviewer: Your row count is the same before and
after the merge. How do you know that is right?

Candidate: I expected it to stay the same, because it is
a left join from orders to a customer table where each
customer should appear once. I used validate='many_to_one',
so pandas would have raised an error if a customer ID
was duplicated.

The indicator column showed a few orders with no
matching customer. I kept them and labelled them
as unknown customers rather than dropping sales.

The validate and indicator arguments of merge are small, but they let you answer the join question with evidence rather than hope. Very few freshers use them, so they also make your notebook stand out when someone reads it.

SECTION 03Hidden state: when the notebook itself contradicts you

Notebooks have a trap that scripts do not. Cells can run in any order, and variables stay in memory after you delete the cell that created them. A notebook that worked on your machine may fail when someone else runs it top to bottom, or the outputs saved in it may come from code that no longer exists.

Interviewers notice this quickly. The execution counters next to each cell, like [12], [3], [27], show the order things ran in. If they jump around, or a cell uses a variable that is never defined above it, the notebook looks assembled rather than built. Worse, if I ask you to rerun a section during a screen share and it breaks, you now have to explain an error under pressure. Clearing hidden state before an interview is basic hygiene.

# From a terminal, in the project folder:
# executes every cell top to bottom in a fresh kernel
# and saves a new copy, so the original stays untouched.

jupyter nbconvert --to notebook --execute analysis.ipynb \
    --output analysis_clean_run.ipynb

# If any cell fails, the command stops and shows the error.
# Fix it in the original notebook, then run again until it passes.

# Fragile: works only on one laptop
df = pd.read_csv('C:/Users/me/Desktop/final/orders.csv')

# Robust: data lives in a folder next to the notebook
from pathlib import Path

DATA = Path('data')
df = pd.read_csv(DATA / 'orders.csv')

# Add a short README line: 'Put orders.csv in ./data'.

Running the notebook end to end in a fresh kernel takes a few minutes and removes a whole category of awkward moments. If it reveals a cell you cannot fix or explain, that is a strong sign the cell should be removed or rewritten before anyone else sees it.

SECTION 04A notebook self-audit before you share the link

Run these three checks on every notebook listed on your resume or portfolio. Each one targets a type of question that causes candidates to freeze.

For EVERY cell that changes data, answer:

[ ] What does this cell change? (rows, columns, values)
[ ] Why is that change correct for THIS dataset?
[ ] What would happen if I skipped it?
[ ] Is the reason written in a comment or markdown cell?

Mark cells you cannot answer with  # TODO-EXPLAIN
Fix or delete every TODO-EXPLAIN before the interview.

Find every hard-coded number in the notebook.

Number | Cell | Where it came from | Written down?
-------+------+--------------------+--------------
50000  | [7]  | __________________ | Y / N
0.3    | [15] | __________________ | Y / N
30     | [21] | __________________ | Y / N

Any number with no source is a question
you will be asked and cannot answer.

Ask a friend to open your notebook and pick
any three cells without telling you which.

For each cell you must, within 30 seconds:
1. Say what it does in one sentence.
2. Say why it is needed.
3. Name one alternative you considered.

Score: 3/3 cells clear = ready.
Anything less = rewrite those cells and repeat.

SECTION 05Habits that make any notebook easy to explain

These habits make explanation part of the notebook, so the interview becomes a reading of reasons you already wrote down.

• Write the why above the code — Use a short markdown cell or comment before each decision, describing what the data meant and why the step follows from it.
• Print shapes after changes — After filters and merges, print the row count. It gives you evidence to point at when asked whether a step lost data.
• Name variables for meaning — Use names such as orders_clean or bulk_orders instead of df2 and temp, so the interviewer can follow without your help.
• Delete what you cannot defend — If a cell came from a tutorial and you cannot explain it for your data, remove it or redo it. A shorter notebook is safer.
• Keep a decisions log — End the notebook with a short list of the main decisions and alternatives. It doubles as your interview revision sheet.

SECTION 06Interview Q&A — Notebook explanation

SECTION 07Test yourself — Notebook explanation quiz

SECTION 08Frequently asked questions

SECTION 09Related reads