Image	Docker Hub	Python	Platforms	Size
ml4t	`ml4t/ml4t:latest`	3.14	amd64 + arm64	~12 GB / ~3 GB
py312	`ml4t/ml4t-py312:latest`	3.12	amd64 only	~9.6 GB
benchmark	`ml4t/ml4t-benchmark:latest`	3.14	amd64 + arm64	~1.7 GB
rapids	(build locally)	3.12	amd64 + NVIDIA GPU	~15 GB

ml4t (Main)

Covers all 27 chapters and 9 case studies. Includes PyTorch with CUDA 12.8 support, LightGBM, scikit-learn, Polars, Plotly, and all ML4T libraries.

docker compose pull ml4t
docker compose up ml4t                    # Jupyter Lab at http://localhost:8888
docker compose run --rm ml4t python nb.py # Run a notebook directly

GPU passthrough (same image, NVIDIA runtime required):

docker compose --profile gpu run --rm ml4t-gpu python notebook.py

py312 (Python 3.12 Dependencies)

For notebooks requiring libraries without Python 3.14 wheels:

Notebook	Library
Ch05 `03_sigcwgan_signatures`	signatory
Ch09 `06_path_signatures`, `12_wasserstein_regimes`	signatory, esig
Ch10 `01_word2vec`, `02_asset_embeddings`, `03_sentiment_evolution`	gensim
Ch15 `06_fed_announcement_bsts`	tfcausalimpact (TFP BSTS)
Ch21 `05_deep_hedging_pfhedge`	pfhedge

docker compose --profile py312 pull py312
docker compose --profile py312 run --rm py312 python 05_synthetic_data/03_sigcwgan_signatures.py
docker compose --profile py312 run --rm py312 \
  /opt/bsts/bin/python 15_causal_estimation/06_fed_announcement_bsts.py

The Chapter 15 command uses the isolated /opt/bsts interpreter. The default /opt/ml4t interpreter and its NumPy 2 signature, gensim, and pfhedge dependencies remain unchanged for every other py312 notebook.

No native arm64 image. On Apple Silicon, read the committed .ipynb outputs, or run the amd64 image under Rosetta with DOCKER_DEFAULT_PLATFORM=linux/amd64.

benchmark (Storage Benchmarks)

For Chapter 2 storage benchmarks comparing file formats and databases (TimescaleDB, ClickHouse, QuestDB, InfluxDB).

docker compose pull benchmark

# Start database services
docker compose --profile benchmark up -d timescaledb clickhouse questdb influxdb

# Run benchmark
docker compose --profile benchmark run --rm benchmark \
  python 02_financial_data_universe/21_storage_benchmark_database.py

# Stop databases
docker compose --profile benchmark down

rapids (GPU Benchmarks)

For Chapter 12 GBM GPU benchmark with RAPIDS cuML and LightGBM CUDA. Requires NVIDIA GPU. Must be built locally:

docker compose --profile rapids build rapids
docker compose --profile rapids run --rm rapids python 12_gradient_boosting/02_gbm_comparison.py

Directory Structure

envs/
├── README.md                  # This file
├── ml4t/Dockerfile            # Main image (Python 3.14)
├── py312/
│   ├── Dockerfile             # Python 3.12 for signatory/esig/gensim/pfhedge/tfcausalimpact
│   └── pyproject.toml         # py312-specific dependencies
├── benchmark/
│   ├── Dockerfile             # Benchmark image with DB clients
│   └── pyproject.toml         # Benchmark-specific dependencies
├── rapids/
│   └── Dockerfile             # RAPIDS cuML + LightGBM CUDA
└── test_all_imports.py        # Import verification script (63 packages)

Import Verification

Each Docker image has a self-test that verifies all packages needed by its notebooks are importable. Run it after pulling or building to confirm the environment is healthy:

# ml4t image (baked-in command)
docker compose run --rm ml4t ml4t-test-imports

# Or explicitly
docker compose run --rm ml4t python envs/test_all_imports.py

# py312 image
docker compose --profile py312 run --rm py312 python envs/test_all_imports.py --image py312

# Benchmark image
docker compose --profile benchmark run --rm benchmark python envs/test_all_imports.py --image benchmark

# Test a specific chapter
docker compose run --rm ml4t python envs/test_all_imports.py --chapter 15 --verbose

What's tested per image:

Image	Packages	ML4T Libs	Utils Modules	Chapters
ml4t	50 third-party	6 (data, engineer, models, diagnostic, backtest, live)	22 (4 repo + 18 case study)	Ch01-Ch26
py312	5 (signatory, esig, gensim, pfhedge, tfcausalimpact)	1 (diagnostic)	—	Ch05, Ch09, Ch10, Ch12, Ch14, Ch15, Ch21
benchmark	5 (duckdb, tables, DB clients)	0	—	Ch02

The test groups packages by chapter, so failures map directly to which notebooks are affected. Exit code is 0 (all pass) or 1 (failures). py312-only packages are shown as informational when running the ml4t test.

Building Locally

If you prefer to build from source instead of pulling from Docker Hub:

docker compose build ml4t                       # ~45 min on x86, ~15 min on ARM64
docker compose --profile py312 build py312      # ~30 min
docker compose --profile benchmark build benchmark  # ~10 min