stefan-jansen/machine-learning-for-trading

Machine Learning for Trading — 3rd Edition

Build, test, and deploy ML-driven trading strategies — from data sourcing to live execution.

This repository hosts the code for Machine Learning for Trading, 3rd Edition by Stefan Jansen — a ground-up rebuild, organized around one end-to-end workflow: how you define a research idea and develop it iteratively into a strategy you can actually run, and keep running, in a live market.

• Nine case studies illustrate the workflow throughout the 27 chapters of the book, from raw data through features, models, backtests, costs, and risk to deployment.
• Generative AI and autonomous agents are new to this edition and cut across that workflow, bringing retrieval-augmented generation, knowledge graphs, and multi-agent systems to financial research.
• The companion website features 112 primers, 61 agent skills, and six production Python libraries that facilitate substantial parts of the workflow.

Start here: Installation walks a blank Linux, Windows or macOS machine to a running notebook, prerequisites included. The short version is under Quick Start below.

Free reader's guide: Join Navigate ML for Trading, 3rd Edition on July 30, 2026 at 11:00 AM ET for a 30-minute map of the book, case studies, code, and companion resources. See all current courses and workshops; the cohort courses are listed under Courses below.

---

What's New in the Third Edition

The whole book traces one path: from data infrastructure and strategy research, across an evidence boundary that separates tuning from evaluation, to deployment and monitoring — with a feedback loop that retrains, pauses, or retires a strategy as its edge decays.

Where earlier editions moved technique by technique, the third edition runs that one process end to end — and adds substantial new material:

• A wider model toolkit: from gradient boosting (XGBoost, LightGBM, CatBoost) to deep time-series architectures (PatchTST, iTransformer, TSMixer, TCN, Mamba) and newer tabular and latent-factor models (TabPFN, TabM, conditional and supervised autoencoders).
• Dedicated strategy-design chapters: transaction costs and risk management are now full chapters, neither of which existed before, joining portfolio construction and strategy synthesis so a raw signal is carried through to a sized, cost- and risk-aware portfolio.
• A full production track: live trading systems (Interactive Brokers, Alpaca, QuantConnect), MLOps and governance (drift detection, safe rollout, circuit breakers, feature stores, experiment tracking), and the operational reality of running strategies, not just building them.
• Generative AI: retrieval-augmented generation grounded in SEC filings, knowledge graphs and Graph RAG, and autonomous, multi-agent research systems.
• Causal machine learning: Double ML, Bayesian structural time series, and causal discovery for separating real effects from spurious correlation.
• Reinforcement learning: optimal execution, market making with inventory, and deep hedging.
• Synthetic financial data: TimeGAN, Tail-GAN, Sig-CWGAN, and diffusion-based generators for validation when history is short.

Methodological rigor is treated as a first-class topic rather than an afterthought. The book draws an explicit line between exploration and confirmation — the evidence boundary — uses walk-forward cross-validation throughout, and confronts the multiple-testing and overfitting problems that quietly invalidate most backtests, with tools like the Deflated Sharpe Ratio, the Rademacher Anti-Serum, and White's Reality Check, plus conformal prediction for honest uncertainty estimates.

The data layer moves to Polars for fast, expression-based manipulation, and every chapter ships in reproducible Docker environments so results repeat across machines; PyTorch, LightGBM, Optuna, and Plotly round out the modeling and visualization stack.

Nine Case Studies

The structural centerpiece of the third edition is nine case studies that run the length of the book. ETFs, crypto perpetuals, intraday equities, options, FX, futures, and equity factor panels are each carried through the same pipeline — from raw data and labels to features, models, backtests, costs, risk overlays, and a final deployment assessment. One disciplined process applied to nine very different markets shows where it works, where it breaks, and why.

112 Primer Topics

Free concept explainers for every idea the book relies on. Each part links to its full list; a few topics show the range:

• Foundations: 8 topics spanning limit order book mechanics, bitemporal data models, and the stylized facts a simulator must reproduce.
• Research Design and Feature Engineering: 21 topics, including multiple testing in factor research, fractional differencing, and path signatures for financial sequences.
• Model Development: 22 topics, among them regularization geometry, conformal prediction in finance, and the mechanism behind double machine learning.
• Strategy Implementation: 27 topics, from the deflated Sharpe ratio and hierarchical risk parity to Almgren-Chriss optimal execution.
• Advanced AI: 8 topics such as Markov decision processes, the policy-gradient theorem, and proper scoring rules for event forecasts.
• Production: 2 topics, champion-challenger evaluation and training-serving skew with feature stores.
• Cross-cutting concepts: 24 building blocks referenced across chapters, for example momentum and mean reversion, the bias-variance tradeoff, and walk-forward validation.

61 Agent Skills

Reusable, guard-railed tasks for coding agents, each with built-in defenses against lookahead bias, data leakage, and multiple-testing errors. Each category links to its full set; a few skills show the range:

• Concepts: 10 skills, including lookahead bias, data leakage, and the information coefficient.
• Data Acquisition: 7 skills spanning fetching data, building bars, and data validation.
• Feature Engineering: 10 skills, among them computing features, triple-barrier labels, and feature selection.
• Evaluation & Validation: 8 skills, from walk-forward CV and purging-and-embargo to the deflated Sharpe ratio.
• Backtesting: 5 skills such as running backtests, cost models, and tear sheets.
• Portfolio Management: 5 skills, including position sizing, risk metrics, and kill switches.
• Infrastructure: 4 skills, for example the canonical schema, the registry system, and Polars patterns.
• Workflows: 5 skills covering factor research, model validation, and production readiness.
• Production: 2 skills, live trading and monitoring & alerting.
• Advanced AI: 5 skills covering research operators, agent memory, forecasting, governance, and RAG evaluation.

Courses

Cohort-based courses on Maven that work through the material live, with direct feedback:

• Machine Learning for Trading: From Research to Production: take a research idea all the way to a deployed, monitored strategy.
• Engineering a Multi-Agent Forecasting System: design auditable multi-agent systems for financial research.

Each course runs as a scheduled cohort; the links above always point to the next one, where you can enroll or join the waitlist. Stay current between cohorts with the Insights newsletter.

---

The ML4T Libraries

The notebooks are built on six production Python packages, each documented and usable on its own — one per stage of the workflow:

---

An introduction and a closing chapter bookend six workflow-aligned parts. Each chapter title links to its guide; notebooks are added directory by directory as they are finalized.

Introduction

1. The Process Is Your Edge

Why process discipline beats model sophistication. Introduces the ML4T workflow as a research-to-production system, regime detection on factor returns and macro indicators, and the evidence boundary that separates exploration from confirmation.

Part I — Financial Data (Chapters 2–5)

The markets, instruments, and infrastructure the rest of the book builds on: a taxonomy of sources, raw exchange messages turned into feature-ready bars, point-in-time fundamentals, and synthetic histories for robust validation.

2. The Financial Data Universe

A taxonomy of market, fundamental, and alternative data. Surveys eight asset classes, quantifies survivorship bias, benchmarks storage formats (Parquet, DuckDB, kdb+, TimescaleDB), and establishes the data-quality framework used throughout the book.

3. Market Microstructure

From raw exchange messages to feature-ready bars. Parses NASDAQ ITCH, reconstructs limit order books from multiple data sources, validates Lee-Ready trade classification, and compares bar-sampling methods — dollar bars deliver the best return normality.

4. Fundamental and Alternative Data

Point-in-time pipelines for SEC EDGAR filings, entity resolution across identifier systems, macro and commodity fundamentals, and alternative-data evaluation — including on-chain crypto fundamentals and prediction markets (Kalshi, Polymarket).

5. Synthetic Financial Data

Generating alternative market histories for robust validation. Implements TimeGAN, Tail-GAN, Sig-CWGAN, Diffusion-TS, and LLM-based tabular generation, evaluated through a fidelity–utility–privacy framework.

Part II — Research Design and Feature Engineering (Chapters 6–10)

Define the trading problem, then turn data into model-ready signals: research design, labels, features, and the evaluation that determines what any model can learn.

6. Strategy Research Framework

Defining the trading game before building models: universe rules, decision schedule, cost model, evaluation protocol, and run logging. Introduces the nine case studies and the walk-forward cross-validation discipline that anchors Chapters 7–20.

7. Defining the Learning Task

Label engineering (forward returns, triple-barrier, trend scanning), univariate feature evaluation (information coefficients, quantile analysis, feasibility screens), multiple-testing control (BH-FDR, Deflated Sharpe Ratio), and causal plausibility checks.

8. Financial Feature Engineering

Five feature families from price data (momentum, reversal, volatility, liquidity, microstructure), structural and cross-instrument features (yield curve, term structure, relative value), contextual features (macro regime, calendar, sentiment), and feature selection with robustness testing.

9. Model-Based Feature Extraction

Features from fitted models: stationarity diagnostics, Kalman filters, Fourier and wavelet spectral features, GARCH volatility, and HMM regime probabilities — with point-in-time correctness enforced throughout.

10. Text Feature Engineering

From bag-of-words through transformers: TF-IDF, Word2Vec and GloVe embeddings, LSTM sequence models, FinBERT sentiment, financial NER fine-tuning, and news-return signal construction.

Part III — Model Development (Chapters 11–15)

Five model families applied to the same nine case studies, each building on the linear baseline.

11. The ML Pipeline

Regularized linear models (Ridge, LASSO, Elastic Net) as the baseline every later model must beat. Logistic regression for direction, SHAP interpretability, conformal prediction for uncertainty, and a cross-dataset comparison across all nine case studies.

12. Gradient Boosting and Advanced Tabular Models

XGBoost, LightGBM, and CatBoost with Optuna multi-objective tuning, plus deep-learning tabular alternatives (TabPFN, TabM). TreeSHAP explainability and cross-dataset results, where gradient boosting is the strongest tabular model in most case studies.

13. Deep Learning for Time Series

LSTM, N-BEATS, Transformers (PatchTST, iTransformer, TFT), TSMixer, TCN, and Mamba, set against the LTSF-Linear debate. A practitioner selection framework and cross-dataset evidence on when deep learning helps and when simpler models suffice.

14. Latent Factor Models

PCA eigenportfolios, IPCA with time-varying loadings, conditional and supervised autoencoders, adversarial SDF estimation, and yield-curve decomposition — with cross-dataset results on when latent factors add predictive value.

15. Causal Machine Learning

Double Machine Learning for isolating factor treatment effects, Bayesian Structural Time Series for event impact, and causal discovery (PCMCI, NOTEARS, VAR-LiNGAM), applied across the nine case studies.

Part IV — Strategy Implementation (Chapters 16–20)

From predictions to deployable strategies — backtesting, portfolio construction, costs, risk, and synthesis.

16. Strategy Simulation

Backtesting as falsification: trading-protocol specification, vectorized vs event-driven engines, an ETF baseline strategy, core metric reporting, regime diagnostics, and strategy-level overfitting control (Deflated Sharpe Ratio, Rademacher Anti-Serum, White's Reality Check).

17. Portfolio Construction

From scores to portfolios: mean-variance optimization and its pitfalls, Hierarchical Risk Parity, the Kelly criterion, conformal position sizing, deep portfolio allocation, and a controlled allocator comparison across case studies.

18. Transaction Costs

Cost taxonomy, spread estimation, market-impact calibration, execution algorithms (VWAP, TWAP, Almgren-Chriss optimal execution), transaction-cost analysis, and practical guardrails — with breakeven costs that vary widely by asset class.

19. Risk Management

VaR/CVaR tail measurement, drawdown and path-risk controls, factor and sector decomposition, stress testing, adaptive risk overlays, deep hedging, and kill switches. Overlay effectiveness turns out to be strategy-specific.

20. Strategy Synthesis

What nine experiments reveal about translating ML predictions into strategies: IC–Sharpe decorrelation, Fundamental Law diagnostics, the model-family cascade, cost-survival analysis, holdout failure modes, and a practitioner's decision framework.

Part V — Advanced AI (Chapters 21–24)

Reinforcement learning, large language models, knowledge graphs, and autonomous agents for finance.

21. Reinforcement Learning for Execution and Hedging

MDP formulation for finance, DQN/PPO/SAC algorithms, optimal execution, market making with inventory management, deep hedging with PFHedge, inverse RL for strategy recovery, and the sim-to-real gap.

22. RAG for Financial Research

Retrieval-augmented generation grounded in SEC filings: ingestion, domain-specific embeddings, hybrid retrieval with re-ranking, constraint-based prompting, RAG evaluation and failure diagnostics, and the transition to agentic workflows.

23. Knowledge Graphs

When graphs earn their infrastructure cost: KG construction from SEC filings, Graph RAG for multi-hop reasoning, graph features for ML (GNN embeddings, centrality, community detection), financial networks, and temporal-leakage prevention.

24. Autonomous Agents

Agent architectures (ReAct, Tree of Thoughts, Reflexion), memory systems, tool contracts, the engineering stack (LangGraph, Claude SDK), a stateful equity-research agent, multi-agent forecasting with adversarial debate, and production reliability.

Part VI — Production (Chapters 25–26)

Taking strategies live — trading systems and the operational infrastructure that keeps them running.

25. Live Trading Systems

A unified framework bridging research and production: Interactive Brokers and Alpaca integration, managed platforms (QuantConnect), order-lifecycle management, pipeline verification, and operational readiness.

26. MLOps and Governance

An ML failure taxonomy (pipeline divergence vs performance decay), drift detection, safe model rollout, circuit breakers, feature stores, experiment tracking, and the MLOps infrastructure financial ML systems need.

Conclusion

27. The Systematic Edge

The systematic philosophy, quant career paths, learning resources, research frontiers, and how to build your own edge. The closing bookend to Chapter 1: the process is the edge.

---

Quick Start

New here? Read these three, in order:

1. What this repository is, and what it is not - what you can reproduce with one command, what a configuration change buys you, what needs real compute or licensed data, and what is not promised. Five minutes, and it sets expectations before you install anything.
2. Installation - Linux, Windows WSL2, macOS, Docker, and GPU.
3. Running notebooks - the case-study pipeline, the run log, and how to experiment without disturbing the downloaded results.

These commands are typed into a terminal on your own computer, not into GitHub. New to the command line? Start with Before You Begin.

Run everything from the repository root. Clone and set up with Docker or a local uv environment:

git clone https://github.com/stefan-jansen/machine-learning-for-trading.git
cd machine-learning-for-trading
cp .env.example .env

docker compose pull ml4t # Option A — Docker (recommended)

Option B is a local uv environment, on macOS, Linux, or inside WSL2. Install uv with its own installer rather than with pip, which is missing or refuses to install on most current systems:

curl -LsSf https://astral.sh/uv/install.sh | sh
source $HOME/.local/bin/env   # the installer's own line; puts uv on PATH here and now
uv sync

Option B compiles several dependencies from source, scikit-learn among them, so it needs a C/C++ compiler and the Python headers: on Ubuntu, Debian and WSL2 sudo apt install build-essential python3-dev, on macOS xcode-select --install. Docker carries its own and needs none of this.

macOS readers: on Apple Silicon take Option B. It is the path walked on real hardware before each release, and it needs only the Xcode command-line tools for the packages that build from source. Docker there is worth its disk only for the twelve ml4t-py312 notebooks, which have no arm64 build and ship pre-executed, and for Chapter 2's containerized database benchmarks. On an Intel Mac take Option A: PyTorch publishes no macOS x86_64 wheel, so Option B cannot work there.

Windows readers: both options run inside WSL2, not in PowerShell. Run wsl --install -d Ubuntu from an Administrator PowerShell, restart, run it a second time (the first run usually installs the WSL runtime without a distribution), and then follow the Linux instructions in the Ubuntu terminal. Installing into Windows Python is not supported and does not work — scikit-learn has no Windows wheel for this Python version and its source build fails. The installation guide has the full WSL2 walkthrough.

Budget about 16 GB for Option B (11 GB environment, 4 GB free datasets, 0.9 GB of git history) and about 12 minutes for the data.

See the installation guide for platform-specific setup and GPU instructions. Intel Macs are Docker-only: PyTorch no longer publishes macOS x86_64 wheels, so the local uv path cannot resolve there.

Download data. Most notebooks need datasets; start with the free ones (no API keys):

uv run python data/download_all.py --free-only

Docker readers run this in the Jupyter Lab terminal (File → New → Terminal) as python data/download_all.py --free-only — there is no host Python on the Docker path.

That command fetches seven datasets and takes about 4 GB and twelve minutes, almost all of it the firm-characteristics panel, first needed in Ch04. To start in about 75 MB, leave it out and fetch it when a chapter asks for it:

uv run python data/download_all.py --free-only --skip-firm-characteristics

The data guide documents every dataset, API-key setup, the loaders, and storage tiers.

(Optional) pre-computed results. To explore the nine released Ch11-20 case studies without retraining, download their verified registries, predictions, model files, and backtest artifacts:

uv run python scripts/download_artifacts.py

Run notebooks. Notebooks are paired Jupytext files (.py source + generated .ipynb). Run a quick smoke test, or open Jupyter Lab:

uv run python 01_process_is_edge/factor_regimes.py                # smoke test
ML4T_DATA_PATH="${ML4T_DATA_PATH:-$PWD/data}" uv run jupyter lab  # local: open the URL it prints
docker compose up -d ml4t                                         # Docker: same address

uv sync already installed Jupyter Lab. Start it from the repo root: the ML4T_DATA_PATH prefix gives the loaders an absolute path, because Jupyter runs each notebook with its chapter folder as the working directory, and they would otherwise search inside that folder and report the datasets as missing. It keeps a value you have already exported and defaults to this repository's data/.

See the guide to running notebooks for case-study pipelines, Papermill parameters, and the experiment workflow.

Docker images

Most notebooks run on the default ml4t image; a few need a specialized one, and each such notebook says so in its preamble. Full details in the Docker environments guide.

Looking for the second edition? It is complete and stable on the second-edition branch — git checkout second-edition, and everything is exactly where the book describes it.

---

Repository Layout

machine-learning-for-trading/
├── 01_process_is_edge/ … 27_systematic_edge/   27 chapters — Jupytext .py + .ipynb, each with a README
├── case_studies/     nine datasets carried through the full pipeline (Ch6 → Ch20)
├── data/             download scripts and loaders for every dataset      → data/README.md
├── utils/            shared config, paths, styling, modeling, and CV code → utils/README.md
├── scripts/          reader utilities (install check, notebook sync, artifacts) → scripts/README.md
├── tests/            Papermill notebook execution + unit guards, run in CI → tests/README.md
├── envs/             Dockerfiles for every image                          → envs/README.md
├── docs/             what-this-is, installation, and notebook-execution guides
├── docker-compose.yml    all Docker services
├── pyproject.toml · uv.lock    pinned dependencies (uv)
└── matplotlibrc      figure styling, auto-applied from the repo root

---

Contributing and Feedback

Found an error, a broken link, or have a suggestion? Early feedback is especially valuable before the book launches.

• Issues: open a GitHub issue
• Website and contact: ml4trading.io

---

License

Code: MIT License · Book content: © 2026 Stefan Jansen. All rights reserved.

Get the book • ml4trading.io • GitHub