A Packt book-launch workshop with Stefan Jansen, author of Machine Learning for Trading.
Run the notebooks in order. Each Colab notebook installs this repository in a fresh runtime and uses the bundled ETF data; no local setup is required. This is the compact ETF quick start for the full Machine Learning for Trading companion repository.
| Notebook | Open in Colab |
|---|---|
| 1. Data and eligibility | Open 01_data.ipynb |
| 2. Features and labels | Open 02_features_labels.ipynb |
| 3. Model training | Open 03_model_training.ipynb |
| 4. Cost-aware backtest | Open 04_backtest.ipynb |
Notebooks 3 and 4 regenerate any missing intermediate files when opened directly in a fresh Colab runtime. Running 1-4 in order avoids that repeated work and follows the workshop narrative.
Prefer local files? Download the complete repository as a ZIP,
extract it, and follow SETUP.md.
Build, test, and validate an end-to-end ML trading strategy on real market data - and use coding and research agents to build and validate it with current ML4T libraries and explicit checks at each stage.
This repo is everything you need to run the hands-on side of the session: the free 100-ETF dataset, four working notebooks that take you from raw prices to a cost-aware backtest, and the briefs behind the two live agent segments. Clone it before the workshop and you're ready; nothing here depends on a paid data subscription, a brokerage account, or a GPU. The slide decks are provided through the workshop platform, not this repo.
Keep the slide deck (provided separately) and this repo's current notebook open side by side. When a green Guided checkpoint slide appears, run the named notebook section with Stefan, stop at the expected output, and compare your result before moving on. The notebooks are prepared: you are checking the research decisions and their outputs, not recreating the whole pipeline from an empty file.
Run the notebooks in this order:
01_data.ipynb- inspect the data and point-in-time eligibility.02_features_labels.ipynb- build features, verify the label, and save the model panel.03_model_training.ipynb- train eight walk-forward validation folds, keep the 2024-2025 holdout sealed, and compare IC statistics.04_backtest.ipynb- run the same strategy with and without trading costs.
The coding-agent build is a presenter-led demonstration; the research-agent
segment traces a real, already-run pipeline result in 05_research_agent_trace.ipynb
rather than firing a live LLM call. You receive both briefs, but you do not
need an LLM API key or coding-agent subscription to complete the
five-notebook build.
By the end of the session you'll have run, end to end, on 18 years of real ETF data:
- A point-in-time-correct feature set (momentum, volatility, RSI, liquidity) on a 100-ETF universe - and know exactly which line of code would leak the future if you got it wrong.
- A LightGBM model validated with walk-forward, leakage-aware cross-validation, scored with the Information Coefficient - naive t-stat and the HAC-corrected one, and why they can disagree by a factor of three.
- A cost-aware backtest with real commission and spread modeling, so you've seen the actual dollar gap between "the backtest" and "the backtest with costs switched on" - not a rule-of-thumb haircut.
- A live look at a coding agent building part of this pipeline from a spec, and a walkthrough of a real, already-run trace of the book's own multi-agent forecasting pipeline (Chapter 24) producing a calibrated, auditable probability estimate.
docs/prerequisites.md- what you need installed (spoiler: Python 3.12+, and a Docker/Colab fallback if that's inconvenient).SETUP.md- three ways to get running: localuv, Docker, or Colab. Pick one and verify it before the workshop.notebooks/- run in order:01_data→02_features_labels→03_model_training→04_backtest→05_research_agent_trace. Each one runs standalone against the data in this repo; each ends by pointing at the next.
notebooks/ 01-05, run in order - the hands-on spine of the workshop
data/ the free 100-ETF dataset, plus one bundled real agent trace (see data/README.md)
docs/
prerequisites.md what you need before Saturday
coding_agent_brief.md the spec used in the live coding-agent build
research_agent_walkthrough.md why the research-agent segment traces a real run instead of a live LLM call
SETUP.md uv / Docker / Colab, pick one
210 minutes, one break: the ML4T workflow - data → alpha factors → labels & validation → model → cost-aware backtest → live handoff - with two live agent segments woven through it (a coding agent builds part of the pipeline; the book's own research-agent pipeline runs a live forecast).
Stefan Jansen is the author of Machine Learning for Trading (Packt), now in
its 3rd edition, and founder of Applied AI. The book's companion code -
450+ notebooks across 27 chapters and 9 case studies, including the full,
non-simplified version of the ETF pipeline this workshop draws on - is at
github.com/stefan-jansen/machine-learning-for-trading.
This workshop is a compact entry point into a larger workflow. Continue with the third-edition book and companion code, free Lightning Lessons, Foundations, Loop Engineering: Reliable Work From Coding Agents, Engineering a Multi-Agent Forecasting System, or Machine Learning for Trading: From Research to Production. Choose by the work you want to do next, not by a fixed sequence.
- Book, companion code, and Foundations: ml4trading.io
- Lightning Lessons, workshops, and live course: maven.com/stefan-jansen
Code and notebooks: MIT, see LICENSE. Bundled market data: see
data/README.md for its own terms.