Centrale Méditerranée · DDEFI option · 2026–2027
Machine Learning
We spend hours deriving gradients, yet the course starts elsewhere: the decision someone has to make, the target, the metric, the guard-rails. Only then the maths, checked in code, on a real credit-default score.
- 36 hours of teaching and one defence day
- 14 and 15 October, 15 and 16 December, Thursday 11 and Friday 12 February 2027 (exact day to be confirmed)
- Taught in French
- Technopôle de Château-Gombert, Marseille
To begin: AI at a crossroads
Wednesday 14 October at 08:30, before session 1: seventy years of AI and its two winters, the 2026 figures on jobs for young developers and data scientists, the engineer and the project manager turned managers of agents, and the procurement bet. First exercise: the groups and subjects questionnaire.
Add me on Kaggle and on GitHub: kaggle.com/sitrakamatthew · github.com/Sitraka17
Next: Kaggle, GitHub, Python
From 10:05 to 10:30, before session 1: what the three tools are for, how to create the accounts, what the final project's GitHub repository will contain (scope document, code, batch scripts, slides, web interface), and where to catch up on Python if needed.
Read: the attention mechanism and the transformer, step by step
The sessions: block 1 (14 and 15 October), block 2 (15 and 16 December), block 3 (Thursday 11 and Friday 12 February 2027, exact day to be confirmed)
Half-day 1: Cadrer avant de modéliser, puis les moindres carrés
Wednesday 14 October 2026 · 08:30–12:30 · on campus
Half-day 2: Vraisemblance, généralisation, lancement du projet
Wednesday 14 October 2026 · 13:30–18:30 · on campus
Half-day 3: Optimisation convexe et régularisation
Thursday 15 October 2026 · 08:30–12:30 · on campus
Half-day 4: Décider sous coût, livrer un dépôt reproductible
Thursday 15 October 2026 · 13:30–18:30 · on campus
Half-day 5: Titanic à finir ensemble, noyaux, SVM et arbres de décision
Tuesday 15 December 2026 · 14:00–18:00 · on campus
Half-day 6: Forêts, boosting et la comparaison honnête
Wednesday 16 December 2026 · 08:30–12:30 · on campus
Half-day 7: Expliquer, auditer l'équité, découvrir sans étiquettes
Wednesday 16 December 2026 · 14:00–18:00 · on campus
Half-day 8: Réseaux de neurones, dérive, conformité, mise en service
Friday 12 February 2027 · 09:00–16:00 · on campus
Half-day 9: Journée de présentation finale : les soutenances
Thursday 11 and Friday 12 February 2027 (exact day to be confirmed) · 08:30–17:30 · on campus
The slides exist in French and in English (same content, same password) and open with the session password. On a laptop they display in reading mode with the notes under each slide, so an absent student can follow.
The homework: one project that moves from block to block
Three team assignments, one between each block, on the project your team chose, up to the final delivery. Each deck reads on its own: deliverables, calendar, grading, traps.
Jalon 1 : votre projet et sa problématique (noté sur 5)
À partir du vendredi 20 novembre 2026 ; bonifié avant le mardi 15 décembre ; dernier délai le mercredi 16 décembre, 23:59 · graded out of 5, 10 % of the grade
Jalon 2 (25 %) : une OLS ou une classification simple sur votre projet
À rendre le vendredi 29 janvier 2027 · 25 % of the grade
Présentation finale avec le code (50 %) : le projet livré
Jusqu'à la veille de la présentation finale (bloc des jeudi 11 et vendredi 12 février 2027, jour exact à confirmer) · 50 % of the grade
A filled example of milestone 1: credit risk (Home Credit, Kaggle)
The docs/cahier-des-charges.md file as we expect it, filled in for a fictional persona on a real dataset, Home Credit Default Risk (a 2018 Kaggle competition). Short version; the full one is in the milestone 1 deck, slides 5 and 6. If your team picks this dataset, the persona, the question and the figures must be your own.
| Element | Content |
|---|---|
| The question | Which of this month's credit applications should the credit officer send to manual review to reduce defaults without refusing more creditworthy clients? |
| 1. The decision | For each application, the credit officer grants, refuses or sends it to manual review. A default costs far more than a creditworthy client refused; the ratio is set with the risk director. |
| 2. The target | TARGET = 1 if payment difficulties on the first instalments (Home Credit's definition); predicted at application, known after several instalments. It captures neither the loss in euros nor late defaults. |
| 3. The metric | ROC AUC (the competition metric) and PR-AUC; business: expected cost per application at the threshold. Useful if it beats the current rule. |
| 4. The safeguards | Gender excluded, age audited (non-discrimination), no variable from after the application, no directly identifying data, final decision by a human: credit scoring of individuals is high risk (AI Act, Annex III). |
| 5. The data | Home Credit Default Risk, Kaggle, 2018: about 307,000 applications, plus the credit bureau and previous applications. Licence: the competition rules. Missing: amounts recovered, economic context. |
| 6. The plan | Naive: threshold on an external score (EXT_SOURCE); best: logistic regression (milestone 2); heavy: LightGBM (final presentation). Roles Data, Model, Review. |
| Who we are | The persona: Aurore Data Conseil, a fictional data consulting firm hired by the risk department of a consumer-credit lender; the risk director signs. Then the real people, one line per member. |
Open the example in the milestone 1 deck (slide 5, password given in class)
Grading
Four assessments make the final grade, out of 100%. Final grading scheme of 19 October 2026.
| Assessment | Weight |
|---|---|
| Milestone 1: your project and its problem statement (graded out of 5)From 20 November to 16 December 2026, 23:59; extra credit before 15 December | 10% |
| Paper test, 16 December 202645 minutes, alone, no AI, no documents; morning or afternoon, time to be confirmed | 15% |
| Milestone 2: an OLS or a simple classification on your projectDue Friday 29 January 2027 | 25% |
| Final presentation with the code (delivered project)Thursday 11 and Friday 12 February 2027 (exact day to be confirmed); one team grade, adjusted for each member by the oral | 50% |
| Total | 100% |
No final exam: Friday 12 February 2027 stays a teaching day, which starts with a project work session. The end-of-day tests and the randomly drawn orals are practice, not graded. Outside the scheme: the 300-word reflection and the Kaggle bonuses.
The reference deck
To reopen whenever a model choice is unclear, during the project and on the job.
Le fil rouge, pas à pas : le dépôt du score de défaut, en captures d'écran
Le dépôt que chaque équipe a reçu, page par page : la pull request à fusionner, make reproduce et make test, le notebook guide et ses pièges, pourquoi un modèle et six tests de robustesse, branches et CI, et deux entraînements conseillés (non notés) avant les jalons sur votre propre projet.
Prédire le prix d'un actif : séries temporelles, XGBoost et le sentiment des news
Un projet de bout en bout, à lire avant de choisir un sujet de finance : cadrage, rendements, validation en marche avant, XGBoost sur variables retardées, sentiment des news, combinaison, pièges.
Le code de chaque modèle : importer, télécharger par API, ajuster, prédire
Pour ceux qui n'ont presque jamais codé : les six briques de Python, des données téléchargées par appel d'API, puis un modèle par section avec son listing exécuté, sa sortie et l'explication de chaque ligne.
Données ouvertes et projets de machine learning par secteur : finance, assurance, logistique, tech, sport
Où trouver des données ouvertes et à quelles conditions, un cahier des charges type, et des projets réels ou d'équipe secteur par secteur, avec la structure du dépôt, les paquets et ce que ça rapporte.
Le manuel de l'expert data et IA : cadrer, outiller, modéliser, déployer, observer
À lire en ligne et à garder sous la main : le cahier des charges, la boîte à outils avec les logos (Airflow, Spark, Docker, Kubernetes, Datadog, Elasticsearch, Grafana…), le B.A.-BA des modèles, déployer et observer une application, les équipes, vingt projets de plus en finance et IA.
Data engineering, SRE et cloud : les commandes, les principes, les métiers, du junior à l'expert
Le pipeline de données, les commandes de base (shell, Git, SQL, Docker, Kubernetes, cloud), les principes clés, les métiers, et ce qu'on attend d'un junior, d'un senior et d'un expert.
Panorama des modèles : lesquels, pour quoi faire, dans quels secteurs
Une famille par slide, avec un schéma, les bibliothèques et les usages en finance, logistique, marketing, industrie, santé.
Le deep learning, l'essentiel : les concepts clés, pourquoi les connaître, les maths au tableau, et un mini-TD
Neuf notions au-delà de la séance 8, chacune en deux pistes, maths au tableau et dix lignes de PyTorch : graphe de calcul, pas de gradient et Hessienne, batch norm et son gradient, profondeur et ResNet, convolution, RNN et LSTM, attention, VAE et GAN ; neuf exercices corrigés à faire en binôme mixte.
Course notes, open to everyone
The slides are encrypted for the students. Each deck also has a public summary: the key ideas, the formulas and a worked example, in English, French and German.
All the notes, as a reading list
All 23 notes
Introduction
AI at a crossroads: the two AI winters, the 2026 job market for young engineers, and a method to work in itTwo AI winters, the perceptron, expert systems, the 2026 job market for junior developers, and the method an engineer needs to stay useful.
Tools
Kaggle, GitHub and Python: the accounts, the Git workflow and the repository of a machine learning projectKaggle competitions, the daily Git cycle, branches and pull requests, and the layout of a reproducible machine learning project repository.
Session 1
Frame before you model, then least squares: decision, leakage, normal equations and conditioningFrame a machine learning project (decision, target, metric, guard-rails, leakage), then derive least squares, the hat matrix and the condition number.
Session 2
Maximum likelihood, logistic regression and honest validation, step by stepFrom the Bernoulli likelihood to the log-loss, the sigmoid, its gradient and convexity, odds ratios, bias-variance and leak-free validation.
Session 3
Convex optimisation and regularisation: gradient descent, step size, ridge and lassoConvexity, gradient descent and its step size, SGD, then ridge and lasso as shrinkage, with worked numbers and lambda chosen by cross-validation.
Session 4
Metrics, calibration and a decision threshold under asymmetric costs, then a repository someone else can rerunConfusion matrix, ROC and AUC, precision-recall, Brier score, a decision threshold derived from costs, and a reproducible ML repository.
Homework
Framing a machine learning project before any code: the question, the decision, the target and a six-line scope documentFrame a data or ML project before coding: a one-sentence question, the decision and its error costs, a dated target, a metric, safeguards, data and plan.
Session 5
The kernel trick, support vector machines, kernel ridge and decision trees, with worked numbersMaximum margin, hinge loss and the SVM dual, the kernel trick and kernel ridge, then decision trees with Gini, entropy and pruning, on worked examples.
Session 6
Random forests, gradient boosting and XGBoost, then the honest comparison of two modelsBagging and random forests, gradient boosting and XGBoost derived step by step, then nested cross-validation and the Nadeau and Bengio corrected t-test.
Session 7
Explaining a Credit Score with SHAP, Auditing its Fairness, and Finding Structure with PCA, k-means and EMShapley values computed by hand, fairness measures and their impossibility, the AI Act for credit, then PCA, k-means and EM on tiny worked examples.
Homework
A first honest model on a new project: OLS or logistic regression, a frozen split, a naive baseline and a metric you can defendHow to set up a first honest OLS or logistic model on new data: leak-free split, naive baseline, justified metric, validation results, limits.
Session 8
From the neural network to the model in service: backpropagation, drift, conformal prediction and an APIForward pass and backpropagation by hand, networks on tabular data, PSI drift monitoring, split conformal prediction and a FastAPI model service.
Homework
Delivering a machine learning project: repository, tested service, traceable report and the documents a committee readsWhat a delivered ML project contains: a rerunnable repository, a tested prediction API, a traceable report, model card, monitoring plan and client memo.
Session 9
Defending a machine learning project: the pitch, the live demonstration, the questions, and what comes afterWhat a machine learning project defence must show, the questions a jury asks, the classic traps, and how to reuse the project in interviews.
Reference
A credit default scoring project, step by step: the reproducible repository behind the UCI Default of Credit Card Clients dataA reproducible credit default scoring project on the UCI credit card data: repository, make reproduce, tests, cost threshold, robustness, CI.
Reference
Predicting an asset price: time series, XGBoost and news sentiment, measured honestlyA worked ML project: log returns, walk-forward validation, lagged features, FinBERT and GDELT sentiment, Diebold-Mariano tests and Sharpe net of costs.
Reference
The code of every model: import, download by API, fit, predictPython basics, how data become numbers, downloading data by API, then the same fit and predict code for every scikit-learn model, with real outputs.
Reference
Deep learning, the essentials: backpropagation, batch norm, ResNet, convolution, LSTM, attention, VAE and GAN, derived and checkedBackpropagation, batch norm and its gradient, ResNet, convolution, LSTM, attention, VAE and GAN, each derived on the board and checked in PyTorch.
Reference
Open data and machine learning projects by sector: finance, insurance, logistics, tech, sportWhere to find open data and under which licence, a six-line project brief, and ML project ideas in finance, insurance, logistics, tech and sport.
Reference
The data and AI expert handbook: scope, tool up, model, deploy, observeScope a data project, pick tools (Airflow, Spark, Docker, Kubernetes), choose a model, deploy an API and app, then observe it in production.
Reference
Data engineering, SRE and cloud: the commands, the principles and the jobs, from junior to expertPipelines, Parquet, SQL on DuckDB, idempotent loads, SLOs and error budgets, Terraform, Docker, Kubernetes, FinOps, and what each career level expects.
Reference
A Panorama of Machine Learning Models: Which Ones, What For, in Which SectorsMachine learning model families: what each is for, which Python libraries, the main pitfall, and where it serves in finance, logistics and industry.
Article
How it works
1
Frame first
Decision, question, target, business metric and model metric, guard-rails, working frame, data pipeline. Before any formula.
2
Derive, then check
Every session starts from a derivation and ends with code that reproduces it and compares it with scikit-learn.
3
One project all the way
A credit-default score on open data, then your own subject with the same method, in teams of two to four: calibrated, thresholded by cost, tested, served behind an API.
4
AI allowed, and declared
For the plumbing, yes. The grade rests on what you can derive and explain without it: the 16 December paper test (15%), without AI, and the oral of the final presentation with the code (50%).
The full programme
All 26 sessions, the milestones, the grading grid and the AI policy are on the course companion page, with four field pages: ML in finance, generative AI, the jobs, and the basics in animations.
The lecturer
Sitraka Forler works with data and systems every day in industry (Enterprise System Architect, POST Luxembourg) and teaches at Centrale Méditerranée, Aix-Marseille School of Economics, IAE Metz, the Université Catholique de Lille and Durham University Business School.