Session 9 · Final Presentations and Closing

Friday 27 November 2026 13:00 – 16:00 Project · Final delivery 60% of the final grade

Deadline

The notebook, the report (max. 8 pages) and the assistant log are due on Thursday 26 November, 23:59, by email with the subject Project – Final – Group name. Slides can be sent up to 12:00 on the day of the presentation. Details are on the Project page.

❓ Today's question

You have built a model on a real problem. Can you convince a decision-maker to use it, and defend every choice behind it?

🗓️ Agenda (3 hours)

Time	Block
13:00 – 13:10	Order of presentations (drawn at random)
13:10 – 15:40	Presentations: 12 minutes + 5 minutes of questions per group, with a break halfway
15:40 – 16:00	Course closing

How to structure 12 minutes

A decision-maker remembers three things: the problem, what your model changes, and whether they can trust it. A structure that works:

Minutes	Slide(s)	What it must answer
1.5	The decision	Who decides what with your prediction? What does each kind of error cost?
1.5	The data	Unit of analysis, rows, target, and what you would not have at prediction time
3	The honest comparison	Baseline vs. your models, same CV folds, mean ± spread. One table or one chart, not five
2	The choice	Which model you would deploy, and why: not only the highest score
2.5	The recommendation	What changes because of the model, quantified, with explicit assumptions
1.5	Limits	What would break in production, and what data you would need next

One number per slide

If a slide needs a pointer to explain, it has too much on it. The chart that compares your models is the most important one in the talk; give it the most time.

The questions you will be asked

Questions are about your code and your choices. Every member of the group must be able to answer them. Expect questions like these:

Show me where you split the data. What happens to the test set after that line?
Which steps are inside your pipeline? What would change if this one were outside?
Why this metric, and why this threshold? What would a 1:1 cost ratio change?
Pick one feature: would you really know its value when the decision is made?
Your best model beats the baseline by X. Is that more than the spread across folds?
What did the assistant get wrong, and how did you notice? (from your assistant log)

Code you cannot explain is code you did not write

This is the core of the assessment. If a group member cannot explain a part of the notebook, that part counts as not done, whoever (or whatever) wrote it.

✅ Final checklist

The notebook runs from top to bottom after Restart & Run all.
The baseline appears next to every model score.
Every model was compared on the same cross-validation folds; the test set was used once.
The threshold (if classification) is justified with the costs of each error.
The report states assumptions and limitations explicitly.
The assistant log has real entries, including what you rejected or corrected.
You responded to the Checkpoint 1 feedback, and the report says how.

Course closing: what you can now do

Look back at the case. Starting from a survey file, you have:

Session	You learned to…
1–2	Read a dataset by its types and turn it into one clean row per household
3	Fit a model, and see why fitting the training data says nothing
4	Report an error you can defend: pipelines, cross-validation, no leakage
5	Turn probabilities into decisions, and price each kind of mistake
6	Use flexible models (trees and ensembles), and weigh accuracy against explainability
7	Ask which variables are worth collecting, and shrink the rest
8	Find structure without labels, and use it to design policy

Throughout, you wrote less code by hand than any cohort before you, and you had to review more of it. That habit of verifying is the part that will not go out of date.

Where to go next

Text and naïve Bayes, and neural networks: ISLR chapters 4.4 and 10, at the same level as this course.
Causal machine learning: when the question is "what is the effect of the programme?" rather than "who is poor?". Look for double/debiased machine learning (Chernozhukov et al., 2018).
Targeting in practice: compare your proxy means test with the community-based and self-targeting approaches discussed in the development literature.

Thank you for a great term. 🎓