Skip to content

Project 2: One page, one finding

Storytelling No models Groups of 3 or 4 Presented 19 October

Your group gets a dataset and a question. You answer it with pandas and charts, and then you fit the whole answer onto a single page.

What you are doing and why

You can already make a chart. What you cannot yet do is decide which three charts matter and throw the rest away.

That is the skill this project is built around. A one-page limit is the only reliable way to learn it, because six charts is always easier than three. The same constraint runs through professional life: nobody reads your notebook, they read the one slide you put in front of them, and what you chose to leave off it is most of the work.

There are no models in this project. No training, no predictions, no accuracy. If you find yourself importing scikit-learn you have misread the brief.

The two deliverables and what each one rewards The notebook rewards thoroughness and shows everything you tried. The poster rewards ruthlessness and shows only what survived. the notebook show everything every hypothesis, including the ones that did not survive every cleaning decision and why as many charts as you need rewards being thorough the poster show almost nothing one question, one answer three charts, and not a fourth one page, no scrolling, no excuses rewards being ruthless Most people are naturally good at one of these. The project exists to force you into the other one. Deciding what to leave out is the harder skill.

Your group and your question

Each group has a different dataset and a different question. The questions are deliberately specific: none of them is "explore this data".

  • Group 1: Titanic

    Could Jack have been saved?

    What would have had to be different about him? Answer with the data, not with the film.

    Dataset · 891 passengers, 12 columns

  • Group 2: Spotify

    What makes a song popular?

    You have every audio feature Spotify measures. Find out how much of popularity they actually explain.

    Dataset · 114,000 tracks, 114 genres

  • Group 3: Netflix

    What kind of company is Netflix now, and when did it change?

    The catalogue has a history in it. Find the turning point and prove it.

    Dataset · 8,807 titles, 12 columns

  • Group 4: Student performance

    What predicts a final grade, and which of those would you refuse to use?

    Two questions. The second one is the harder half and carries equal weight.

    Dataset · 395 and 649 students, 33 columns

Every one of these datasets has a trap in it

Not a mistake in the file. A way of looking at it that produces a confident, wrong answer, and that plenty of published analyses of that exact dataset have fallen into.

You are not being told which. Finding it is part of the work, and the section on things that are wrong with your data is where you report it.

Deliverable one: the poster

One page, landscape, exported as PDF. It holds five things and nothing else.

What goes on the poster A single landscape page holding the question as a headline, the answer in one large sentence, three charts each with a finding as its title and one line of explanation, and a note on what the data could not answer. the question, as a headline the answer, in one sentence, large if you cannot fit it in one sentence you have not finished thinking a title that states the finding chart 1 one line of explanation a title that states the finding chart 2 one line of explanation a title that states the finding chart 3 one line of explanation what this data could not tell you Five elements. A fourth chart is not an improvement, it is a failure to choose.

  1. The question, as a headline.
  2. The answer, in one sentence, large enough to read from two metres away.
  3. Three charts. Not four. Each with a title that states a finding rather than naming the variables. "Third class men had almost no chance" beats "Survival by class and sex".
  4. One line under each chart saying what it shows.
  5. One line on what this data could not tell you.

Rules:

  • Every chart must come from your own code. Build the layout wherever you like, in slides, in Canva, or with plt.subplots, but the charts are yours.
  • Readable when projected. If the axis labels vanish on the screen, the chart has failed at its only job.
  • Your group's names on it.

The hardest rule is the third one. You will want six charts. Cutting to three is where the learning is, and it is also where the arguments in your group will happen.

Deliverable two: the notebook

One notebook, in a group GitHub repository, with sections in this order. It is the first-hour workflow from the handbook, done properly.

  1. Load and profile. Shape, dtypes, missing values, and what the columns mean. Two commands and a short paragraph.

  2. Three hypotheses, written before you explore. This section must come first in the notebook. If you write them afterwards you will write ones you already know are true, and you will have proved nothing.

  3. Cleaning, with a reason per decision. One line each: what you did and why. Types, missing values, duplicates, anything you dropped.

  4. Exploration. Distributions first, then relationships. This is where the charts that did not make the poster live.

  5. Each hypothesis tested, marked survived or rejected. A rejected hypothesis is a result. Do not quietly delete it.

  6. Three things wrong with this data, with evidence for each. See below.

  7. Conclusion and limitations. The answer to your question, and an honest paragraph on what would change it.

Three things wrong with this data

Every one of these four datasets has documented problems. Your job is to find three of them in yours and prove each with a line of code or a chart.

The kinds of thing that count:

  • A column that says one thing and means another.
  • Missing values that are not missing at random.
  • A group that is far too small to conclude anything about.
  • A number that cannot be true.
  • A column that would not have existed at the moment you would need to use it.
  • Whole populations absent from the data entirely.

The kind of thing that does not count: "there are some missing values". Say which, how many, and why it matters for your question.

This section is the difference between reading a dataset and interrogating one. It is also the habit that makes you useful in a job, because the first thing a real dataset does is lie to you.

How it is marked

Weight What is assessed
50 The notebook: hypotheses written first, cleaning justified, exploration that goes somewhere, honest results
30 The poster's finding: is the answer supported by what you showed?
20 The poster's craft: is it readable, edited, and does it respect the five-element limit?

The ordering matters. A beautiful poster with a weak finding scores lower than a plain one with a strong finding.

What earns marks fastest: rates where rates belong rather than raw counts, denominators stated, axes that start at zero when they should, and a limitations section that admits something real.

What loses them fastest: a chart with no title, a claim with no chart, a notebook where the hypotheses were obviously written last, and six charts on the poster.

Timeline

The three weeks to presentation day Week one, write the hypotheses and clean the data. Week two, explore and test them. Week three, cut to three charts and build the poster. Presentation on the nineteenth of October. 1 week 1 hypotheses first, then clean 2 week 2 explore and test them 3 week 3 cut to three, build the poster 19 October gallery walk Groups that leave the poster to the last evening produce six charts and no argument. Cutting takes longer than plotting.

A suggestion on the order of work, from groups who have done this before: spend week one on the data and resist the urge to plot anything pretty. The good charts come from understanding, and understanding comes from cleaning.

Before you present, check

  1. The poster has exactly three charts.
  2. Somebody outside your group can read the answer in under thirty seconds.
  3. Every chart title states a finding, not a pair of variable names.
  4. Your notebook's hypotheses section is the first thing after the profiling, and you did not edit it after exploring.
  5. Every cleaning decision has a reason next to it.
  6. You have three genuine problems with the data, each with evidence.
  7. Your limitations paragraph says something that costs you.
  8. The repository link works from a private browser window.

Questions to think about

These are worth arguing about in your group, and some of them will come back at you on the nineteenth.

  1. You cut three charts to get to three. What did the cut ones say, and are you confident none of them contradicted your answer?
  2. Your question has a yes or no shape. Did the data actually give you one, or did you round an unclear answer into a clean one?
  3. If another group had your dataset and your question, would they reach your conclusion? Which of your decisions would they most likely have made differently?
  4. Somebody reads only your poster and acts on it. What is the worst decision they could make that your poster did not warn them about?
  5. You found three things wrong with the data. Would you still publish your finding? What would you say alongside it?

A note on tools

pandas and a plotting library are required. NumPy where it helps, and there is no prize for using it if it does not.

The poster layout can be built anywhere. If you want it entirely in Python, plt.subplots with fig.savefig("poster.pdf", bbox_inches="tight") produces a single-page PDF and gives you complete control. If you would rather assemble it in slides or Canva, export your charts as PNGs at dpi=200 and place them. Neither route scores higher.