What this course asks of you, what it gives you, and how you are judged
HEC Montréal
qmib.load(...), and know where the file actually isEverything you need to type is on one sheet — how to start Python in VS Codium on your platform, how to make a file and run it, and the four DataFrame moves you will use every week from now on.
Python Cheatsheet · Session 01 (PDF) · in the repository at
01-…/PYTHON-CHEATSHEET.pdf
Print it, or keep it open beside the editor. Nothing on it is examinable; all of it is assumed.
“All models are wrong, but some are useful.”
— George Box
The whole course lives in the second half of that sentence. Knowing a model is wrong is easy. Establishing that it is useful, and saying for what, is the work.
| First half | The syllabus — this deck |
| Second half | A conversation: Europe 2031, and what AI actually is |
No mathematics today. It begins in Session 02.
When can a number about the past be used to make a claim about the future?
Every method in the eleven sessions that follow is an answer to some version of that question — and every one of them comes with conditions.
You are expected to work with a language model throughout. It is a required tool, not a tolerated one.
And the standing rule: every deliverable must document at least one instance where you identified an LLM output as wrong, unverifiable, or misleading — and explain how you established that.
You are never penalised for using the model. You are penalised for using it uncritically.
Language models made information abundant. They did not make abundant the ability to ask the right question or to recognise a wrong answer. Both are functions of expertise.
A concrete case, from Session 05. Ask a model to select variables and it will run a lasso and hand you a list. The list looks authoritative. Only if you know the lasso is unstable under correlated predictors will you ask the follow-up that matters — “how often is each of these selected across bootstrap replicates?” — and turn a misleading answer into a useful one.
The model will not volunteer that question. You have to know enough to ask.
| Session | Question | |
|---|---|---|
| 02 | Exploratory data analysis, and the first model | What does the data look like? |
| 03 | Regression: adequacy, validity, robustness | Can the model be trusted? |
| 04 | Logistic regression | The outcome is a category. Now what? |
| 05 | Ridge, lasso, elastic net | Which few indicators carry the signal? |
| 06 | Regression: advanced considerations | Does it hold across countries and years? |
Midterm: Wednesday 28 October — on paper, closed book, covering Sessions 1–6.
| Session | Question | |
|---|---|---|
| 07 | Replication workshop | Can you reproduce what the papers claim, with only what you have learned? |
| 08 | KNN and the bias–variance trade-off | Flexible, or just unstable? |
| 09 | Structural equation modelling | Can you measure the unobservable? |
| 10 | Causal inference I — matching | Did the policy do anything? |
| 11 | Causal inference II — difference-in-differences | What would have happened otherwise? |
| 12 | Final group presentations | Can you make someone act on this? |
Sessions 2–6 build the model and test it; session 7 rebuilds five published results with it. Sessions 8–9 find structure nobody labelled. Sessions 10–11 confront the question the first nine cannot answer: whether anything caused anything.
| Phase | What | |
|---|---|---|
| Before | ~60–90 min | Read the paper. Load its data. Attempt the self-check |
| First half | 90 min | Lecture: the method, derived and then run in Python |
| Second half | 90 min | Your group applies it to your own project |
| Last 20 min | Two-minute reports — presenter drawn at random |
The presenter is drawn when your group is called. All three of you must understand the analysis, the number, and what would undermine it.
| Component | Weight | Form |
|---|---|---|
| Participation | 10% | individual, continuous |
| Midterm exam | 30% | individual, on paper — pen and calculator only |
| Team work | 20% | groups of three — the SDAfrique methods paper, presented Session 12 |
| Final exam | 40% | individual research paper (25) + individual oral, 15 min (15) |
Two are individual and carry 70%. The group cannot carry you, and you cannot carry the group.
The four components give a mark out of 100. That mark is then converted to a letter grade, and the conversion is curved — with the class median at B+.
Within each component, nothing is curved. Participation is scored against the published absolute bar; the midterm and the paper are marked against their rubrics. The curve applies once, at the end, so an assessment that turned out harder than intended does not cost you a letter.
It is not a quota, and it is not a race against the person next to you.
Attendance is assumed, so participation cannot mean turning up. It means active engagement — and it is earned in three ways, any of which counts.
| In the room | a question that changed how another group saw their result, a connection between two angles, a correction that landed |
| In the practice | working the ninety minutes: your own commits, the two-minute report when drawn, engaging with what other groups found |
| In the optional lab | a completed QMIB Lab knowledge check, done at home after a session |
What is marked is sustained engagement across the term — by any mix of the three. Not a count of interventions, and never a ranking against each other.
After each session the QMIB Lab offers an at-home knowledge check. It is never a prerequisite for the next lecture, and it is never marked for correctness.
warin.ca/qmib-labs — one lab per session, sessions 02 to 11.
Nothing is expected of you for today’s session: the first lab follows session 02.
A completed report counts toward participation because attempting it is the engagement being measured. It exists so you can find out privately whether you actually followed the session — and so that full marks do not depend on speaking up in a room of thirty.
Failing to deliver the two-minute report when drawn is still the clearest participation failure there is.
No machine, and no assistant. Every other assessment assumes a model beside you, because that is how the work is now done. This one deliberately does not.
It is the only instrument that isolates what you can derive unaided — which is the thing that lets you tell when the assistant is wrong, for the rest of your career.
| The SDAfrique methods paper and its repository — due Sunday 13 December | 16 |
| Presentation and defence — Session 12 | 4 |
Eleven weeks with a real client: a student in the SDAfrique programme who identified the problem and will use what you produce. She is your client, not your co-author. The project and its intellectual property stay hers: you provide intelligence — analysis, method, evaluation — and you do not code her software.
An individual multiplier from confidential peer ratings, gated on the git history, applies to this 20% only — never to the midterm, never to the final.
Default ×1.00. A mark is reduced only when two independent records agree.
A question of your own, international, using one method from the semester — the paper you would show a thesis supervisor, and could grow into your MSc supervised project or thesis.
6,000–8,000 words, including references and appendices · a research question on an international topic · 10 academic articles in the literature review · the data · the model · the results.
The paper is a methods paper written for a real reader: an SDAfrique student in Benin who works in R, on a modest machine, with limited bandwidth, on a problem she chose herself.
A paper is written slowly, with help, over weeks. Fifteen minutes without documents establishes what you can reconstruct and defend in real time — which is what a referee or an employer actually tests.
This course applies the standard Elsevier and Emerald apply to submitted work.
It is a research assistant, not a co-author. It does not do the job; it helps you do the job.
Ask it to argue against you. Its most useful output is an objection you had not considered.
Every deliverable carries an LLM statement: what you used, for what, and the one output you caught being wrong.
HEC Montréal’s Règlement sur les fraudes et plagiat applies in full. Two clarifications specific to this course:
All data in this course are public. Respect each licence, record provenance, and check before redistributing anything.
And one more, which is the hardest: a well-argued refusal earns full marks. “This cannot be claimed, and here is what would be needed” is the answer I am looking for.
A result that does not reproduce is not a result. That is not pedantry — it is the standard every serious journal now applies.
A conversation, not a lecture. Europe 2031 — the scenario you read — and what AI actually is. Bring the page you annotated and the one question you could not resolve.
Syllabus in full: SYLLABUS.md · Session 01: README.md · The conversation: 02-practice/README.md
MATH60033A · Session 01 · Thierry Warin, PhD · HEC Montréal