Quantitative Methods in International Business with Python
MATH60033A · HEC Montréal
Foreword
Ukraine supplied around seventy per cent of the world’s neon. Russia controlled some forty-four per cent of its palladium. Taiwan fabricates close to two-thirds of its semiconductors. None of those three facts was secret, and none of them appeared on the balance sheet of a company that depended on all three — until a pandemic and a war made them appear at once (Warin 2022). Firms discovered the shape of their own supply chains by watching them break. The research literature had been arguing for years that concentration and fragility in global production were badly measured and badly understood; what it lacked was an event that made the point unarguable (Baldwin and Freeman 2022).
That is the difficulty this book is written against. The transformations that matter most in international business are not hidden. They are unmeasured by the people they will affect, and the received account of how the world economy works quietly supplies the missing numbers with assumptions. Chief among them is the assumption that global value chains, having been optimised for decades, must be efficient — an assumption that survives largely because so few people test it against data (Warin 2025c).
A second difficulty is that the unit of analysis is wrong in most public discussion. Countries do not trade; firms do. This is not a rhetorical point: heterogeneous-firm trade theory reorganised the field around it, showing that aggregate trade responds mainly through which firms enter and leave export markets rather than through all firms adjusting together (Melitz 2003; Warin 2026a). Aggregate statistics on bilateral flows are the sum of those decisions, taken under constraints — freight, insurance, policy wedges, the cost of switching a supplier — that the aggregate cannot show. The gravity equation remains the workhorse for describing those flows (Anderson and Wincoop 2003), and describing them is not the same as judging whether they are the ones that should be happening (Warin 2025a). Work at the level where the decision is actually made requires data at that level, and methods equal to it.
And the ground itself is moving. The central economic shift of this decade is not digitisation but a pivot from producing goods and knowledge toward valuing the data generated in the course of producing them — in agriculture, manufacturing, health, logistics and public administration alike (Warin 2026b). An economy organised around data capture and platform intermediation concentrates advantage differently than one organised around factories, and it rewards the institutions and the analysts who can read what the data say (Warin 2023).
The methods have moved with it. Econometrics has absorbed a toolkit built for prediction rather than for inference, and the two demand different care (Varian 2014); the data themselves are increasingly administrative, transactional and unstructured rather than surveyed (Einav and Levin 2014). Quantities nobody could measure at all are now constructed — economic policy uncertainty read off newspaper archives has entered mainstream macroeconomics (Baker, Bloom, and Davis 2016), and the productive capabilities of a whole economy can be inferred from the composition of what it exports (Hidalgo and Hausmann 2009). None of this removes the need for the discipline this book teaches. It multiplies the number of places to go wrong.
None of that is an argument that quantitative methods produce certainty. It is an argument that they are the only way to hold a claim about the world economy to account — to say what would have to be true for it to hold, and to notice when it stops holding. Trade theory has been rewritten repeatedly, from the mercantilists through Ricardo to the platform economy, and each rewriting followed evidence that the previous account could not absorb (Warin 2025b).
This book teaches the methods, on real data, from the first week. Teaching statistics in a business school pulls in two directions: real data are messy, and pedagogical data are false. This course resolves it in favour of the mess (Warin 2024). Every technique here is applied to a panel carrying the missing values, the influential observations and the awkward panel structure of the real European sources — as teaching fixtures generated from a known structure, so that the methods behave as they would on the real series while no number here can be mistaken for a fact about Europe. Chapter 1 says what that means and what it costs.
Twelve chapters follow the arc from describing data to defending a causal claim. Each one asks a single question — can this regression be trusted?, did the policy do anything? — and each ends where a referee would begin.
There is a companion volume in R. Data Science for International Business with R (warin.ca/ds4ibr) covers much of the same ground — exploratory analysis, regression and panel data, logistic models, unsupervised learning, structural equation modelling — in the other language this field is written in. The two books are deliberately close in structure, so a reader fluent in one can follow the other and see what is a property of the method and what is a property of the tooling. This one is the Python volume.
How to use it
| If you want to | Go to |
|---|---|
| Prepare for next week | the session’s Before the session |
| Re-read a derivation you lost in class | the session’s lecture sections |
| Know what the group work asks for | the session’s In the practice |
| Find where something was covered | the search box, top of the sidebar |
| Know how you are graded | the syllabus, in the course repository |
| Get your machine working | Setting up your tools |
The search box is the reason this book exists. Slides are good for a room and bad for finding something three weeks later. Search covers every session at once, so a term you half-remember — leverage, soft-thresholding, parallel trends — takes you to the session that introduced it.
The slides
Each chapter links to the three decks it was built from. The slides remain the version delivered in class; the book is the same material set as continuous prose, which is the better form for reading afterwards.
The twelve chapters
| Chapter | Question | |
|---|---|---|
| 1 | Foundations: Scenarios, Tools, and the Syllabus | Before we model the future, what are we claiming to know? |
| 2 | Exploratory Data Analysis, and the First Model | Before you model anything, what does the data actually look like? |
| 3 | Regression: Adequacy, Validity, and Robustness | You have fitted a regression. Can it be trusted? |
| 4 | Logistic Regression: Binary, Ordinal, and Multinomial | The outcome is a category, not a number. Now what? |
| 5 | Regularisation: Ridge, Lasso, and the Elastic Net | When is a deliberately biased estimator the better one? |
| 6 | Regression: Advanced Considerations (asynchronous) | Does the relationship hold across countries, and across years? |
| 7 | Principal Component and Factor Analyses | How many independent things are actually being measured? |
| — | Midterm — on paper, closed book | Chapters 1–7 |
| 8 | K-Nearest Neighbours and the Bias–Variance Trade-off | Flexible, or just unstable? |
| 9 | Structural Equation Modelling | Can you measure something you cannot observe? |
| 10 | Causal Inference I: Counterfactuals, Randomisation, Matching | Did the policy do anything, or were the groups different to begin with? |
| 11 | Causal Inference II: Difference-in-Differences | What would have happened otherwise? |
| 12 | Final Group Presentations | Can you make a decision-maker act on this — and say what would change your mind? |
Source: github.com/warint/quantitative-methods · 6 November 2026