Financial Machine Learning

Yonsei University · for 3rd- and 4th-year undergraduates, beginners welcome · 16 units

Browse and download → https://keerhee.github.io/financial-machine-learning/

Sixteen units, four optional special sessions beyond them, and a supplementary packages on linear and logistic regression with KOSPI data, outlier detection, meta-labeling and a map of the algorithm families. The sessions: S1 and S2 on recommenders, S3 and S4 on the AI-native firm.

From what machine learning is, through the algorithms, to the parts that only matter in finance: bars and the CUSUM filter instead of the calendar, the triple barrier instead of "will it go up tomorrow", purged cross-validation instead of a naive backtest, and a Deflated Sharpe Ratio to tell you when the result was luck.

Built on Hull's Machine Learning in Business, López de Prado's Advances in Financial Machine Learning, and Causal Factor Investing — and carried through three industrial applications: product recommendation, fraud detection, and an AI chatbot.

66 PDFs · 1,729 slides · 45 practice data files. For each of the sixteen units a primer, a lecture deck and a problem set; plus the syllabus, a midterm practice set with its answer key and cheat sheet, four special sessions, five supplementary packages in English and Korean, and the companion Colab notebook the labs run from. Only PDFs, CSVs, Markdown and the notebook are tracked — the editable PPTX and DOCX sources are excluded by .gitignore.

---

Layout

Path	Contents
U01_ML_and_the_Finance_Problem/ … U16_Integration_and_Capstone/	Per unit: the primer, the lecture deck, the problem set, and a data/ directory with the CSVs its three problems use. Each has a README naming the special sessions it leads to
course/	Syllabus, the curriculum map, and the data setup notebook the labs use
special_sessions/	Four optional sessions outside the sixteen units, each a deck and a companion note
supplementary/	Supplementary packages beside the units — a deck (with its data and notebook for the KOSPI labs), in English and Korean
lecture_sd/	In-class decks — the short decks used in the face-to-face class (Week 7: classifiers, classification metrics, outlier detection), in English and Korean
midterm/	Midterm practice set (units 1–7, 48 questions), its answer key, and a cheat sheet
site/	The GitHub Pages listing page (index.html, one file)

Page counts match the source slide counts one for one. GitHub renders both PDFs and CSVs in the browser, so nothing needs cloning to read.

File names

Form	Example
Primer	U07_Primer_Trees_and_Ensembles.pdf
Lecture deck	U07_Lecture_Trees_and_Ensembles.pdf
Problem set	U07_ProblemSet_Trees_and_Ensembles.pdf
Practice data	U07_Trees_and_Ensembles/data/fml_w7_p1_loans.csv
Special session	special_sessions/S3_AI_Native_Asset_Manager.pdf
Supplementary package	supplementary/SUP01_Linear_Regression_KOSPI_EN.pdf (_KR for Korean)
In-class deck	lecture_sd/W07_Classification_Metrics_EN.pdf (_KR for Korean)

U07 is unit 7 — the seventh block of material, not the seventh week of term. The data files keep their original w7 naming from when the course was written; unit 7 and its data are the same thing.

---

The four parts

Part	Units	Topic
I · Foundations	1–3	ML paradigms · bars and CUSUM · triple barrier and meta-labeling
II · Core Algorithms	4–9	Linear · logistic, KNN, NB · SVM · trees and ensembles · deep learning · unsupervised
III · Financial ML in Depth	10–13	Purged CV, DSR and PBO · HRP and NCO · NLP, LLMs and RAG · reinforcement learning
IV · Frontier and Causality	14–16	Causal inference · SHAP and fairness · TFT, GNN and agents · integration

Anomaly detection runs as its own thread across the algorithms — LOF in unit 5, OCSVM in unit 6, Isolation Forest in unit 7, the autoencoder in unit 8 — because in production no single detector is enough.

---

Sixteen units, fourteen teaching weeks

Week 8 is the midterm and week 16 is the final, so the sixteen units are taught in fourteen weeks. Two weeks carry two units: week 5 takes the three simple classifiers together with the support vector machine, and week 14 takes causality and fairness together with the frontier.

The exam scopes fix where the line falls — the midterm covers units 1–8 and the final covers units 9–16 — so the split is even, eight units either side. The full table and the rejected pairings are in course/curriculum_map.md.

---

Midterm practice

A practice set for the midterm, its answer key, and a cheat sheet to review from — all in midterm/.

File	What it is
Midterm_Practice_U01-07.pdf	48 basic multiple-choice questions on units 1–7, about 50 minutes
Midterm_Practice_U01-07_Answer_Key.pdf	Answer grid, formula sheet, and a worked explanation for every question — why the right option is right and why each other option is wrong
Midterm_Cheat_Sheet.pdf	The rule behind every question, one topic per page: data and splits, the triple barrier, regularisation, precision–recall–F1, logistic regression, SVM, anomaly detection, trees, bagging, boosting, Isolation Forest

Try the questions before opening the key. The key is published because this is a practice set; .gitignore still refuses every other answer key by name.

---

Primers — watch before the lecture

Every unit opens with a seventeen-slide primer that assumes nothing. It is where the arithmetic goes, so the lecture does not have to stop for it: what a dot product is before the support vector machine, what a centroid is before k-means, what discounting is before reinforcement learning.

Unit	Primer	Unit	Primer
1	What a Model Is	9	No Answer Key This Time
2	How Data Becomes a Bar	10	How to Fool Yourself
3	What Are We Predicting?	11	Why Two Is Safer Than One
4	The Straight Line	12	Turning Words Into Numbers
5	From a Number to a Verdict	13	Learning by Doing
6	What Makes a Line Safe	14	Correlation Is Not Enough
7	Asking the Right Question	15	Attention and Neighbours
8	One Neuron at a Time	16	The Whole Pipeline

Twenty minutes each. If none of it is new, treat it as revision and skip to the lecture.

---

Special sessions

Four optional sessions that sit outside the sixteen units. Each is a deck with a companion note in Markdown, so it reads on GitHub without downloading anything.

#	Session	Slides	What it covers
S1	Recommender Systems	23	The third running app of the course. Content-based and collaborative filtering, matrix factorization — built from cosine similarity (unit 12) and hidden factors (unit 9).
S2	Recommender Systems · Part 2	30	Deep-learning recommenders, from factorization machines as the bridge up to graph- and language-based models.
S3	The AI-Native Asset Manager	26	The capstone view: the investment process as a team of cooperating agents — analysts, traders, risk — with humans overseeing. Pulls units 10–15 together.
S4	AI-Native Asset Manager · Part 2	28	The implementation half — building that agent firm in LangGraph as an explicit state machine, with checkpointing and human-in-the-loop.

Each unit's README points forward to the sessions that pick it up: unit 9 and unit 12 lead to S1, unit 8 and unit 15 to S2, units 10, 11, 14 and 15 to S3, and units 12, 13 and 15 to S4.

S1→S2 and S3→S4 are pairs; the second of each assumes the first. Notes: S1 · S2 · S3 · S4

---

In-class decks

Short decks used in the face-to-face class (lecture_sd/), each in English and Korean. Here W07 is the teaching week, not a unit number.

Deck	Week	PDF	Slides
Classification Models — The Basics	Week 7	EN · KR	6
Classification Metrics	Week 7	EN · KR	5
Anomaly Detection — Fraud and Unusual Transactions	Week 7	EN · KR	9

Supplementary material

Beside the units and the special sessions, self-contained packages: a deck, and for the KOSPI labs its data and notebook. Unlike the rest of the repository they ship in Korean as well as English.

#	Package	Slides	What it covers
SUP01	Linear Regression on the KOSPI · 한국어	51 · 49	Least squares, gradient descent, metrics, Ridge / Lasso / ElasticNet and polynomial features — then a KOSPI lab (29 indicators, 160 months) that walks the Colab notebook cell by cell and reproduces its 0.95 R² on a random 80/20 split. Data and notebook in supplementary/SUP01_Linear_Regression_KOSPI/. Pairs with unit 4.
SUP02	Logistic Regression on the KOSPI · 한국어	58 · 58	Sigmoid, odds and maximum likelihood, the confusion matrix and its metrics, ROC / AUC and precision–recall — then a lab on the same KOSPI data that classifies next month's direction, compares the model with an always-up baseline and repeats the random split 100 times (mean accuracy 0.545 against 0.570). Data and notebook in supplementary/SUP02_Logistic_Regression_KOSPI/. Pairs with unit 5.
SUP03	Anomaly Detection · 한국어	30 · 24	Unsupervised detectors that learn what normal looks like: the One-Class SVM and SVDD (a boundary or hypersphere around normal data), Isolation Forest (outliers isolate in few random splits), the Local Outlier Factor (density relative to neighbours) and autoencoders (outliers reconstruct badly), each with its formula and a scikit-learn example. Pairs with unit 5.
SUP04	Meta-Labeling · 한국어	10 · 13	Meta-labeling in three stages: stage 1 picks the side and is tuned for recall, stage 2 — a classifier trained on triple-barrier labels — decides whether to trade or pass and raises precision, stage 3 sizes the bet with the Kelly criterion. A trading example: cumulative return 19.7% → 39.6%, maximum drawdown −61.9% → −36.8%. Pairs with unit 3.
SUP05	ML Algorithm Family Tree · 한국어	21 · 21	A one-deck map of the main algorithm families — supervised (regression and classification), unsupervised (clustering and dimensionality reduction), ensembles (bagging and boosting) and reinforcement learning — each shown with the objective it optimises, from least squares to the Bellman equation. An overview across the units.

Details in supplementary/README.md.

---

The lab notebook

Twelve of the sixteen lecture decks send you to the same companion notebook, so it lives once in course/ rather than inside any one unit:

course/FML_Data_Setup_Guide.ipynb — open it in Colab

Run section 0 once, then jump to the section for the unit you are on. Section 10 is a map from lab to section:

Unit	Lab topic	Run section
1	Returns and normality; the overfitting demo	2
2	Dollar bars; CUSUM events	3 (a)
3	Triple-barrier labeling	3 (b)
4	Linear models; index tracking	3 (b)
5–8	Fraud and default classification	4, or 5 offline
9	PCA, denoising, clustering	3 (b)
10	Leakage and multiple testing	2
11	HRP weights; comparing portfolios	3 (b)
12	Similar sentences; a tiny RAG	6
13	Q-learning; all-at-once vs sliced execution	7
14	SHAP-style explanation; fairness across groups	5
15	Attention forecast; a graph update	8
16	The capstone pipeline	9
any	No internet available	5 — offline fallback

Prices come from yfinance (free, no signup); the fraud and default sets come from Kaggle; and section 5 simulates every one of them if both are blocked. Units 12, 13, 15 and 16 need no download at all — their data is a few sentences or a few arrays, built in the notebook. Nothing needs a paid account.

Sections 0 to 5 keep the numbers the lecture decks cite, so a deck that says "run §3 (b)" still means the multi-stock panel.

---

Practice data

Every problem ships the data it needs. The files are small, synthetic or derived, and sit beside the deck that uses them:

U05_Classification_and_Anomalies/data/
    README_W5.txt
    fml_w5_p2_applicants.csv
    fml_w5_p3_likelihoods.csv

No download step, no API key, no vendor account — the numbers are chosen so the hand-calculation in the problem set and the code you write afterwards agree exactly.

---

Tooling

Python — pandas, numpy, matplotlib, scikit-learn; XGBoost for the ensembles; PyTorch for the neural network and autoencoder units.

Prerequisites are deliberately light: pandas and numpy basics, conditional probability and Bayes, matrix multiplication and eigendecomposition, and returns and risk. None of it is graded, and the syllabus lists a self-study resource for each.

---

Licence

Course materials and practice data under CC BY-NC-SA 4.0. Textbook content referenced in the slides belongs to its authors and is cited, not reproduced.