[project]
name = "ml4t"
version = "3.0.0"
description = "Machine Learning for Algorithmic Trading — 3rd Edition"
readme = "README.md"
requires-python = ">=3.14"
license = "MIT"
authors = [{ name = "Stefan Jansen" }]

dependencies = [
    # ===================
    # Core Data Science
    # ===================
    "numpy>=1.26",
    # scipy<1.18: 1.18 removed the private scipy.cluster.hierarchy._LINKAGE_METHODS
    # that PyPortfolioOpt 1.6.0 (latest) still references for HRP linkage validation
    # (Ch17 08_library_comparison). Lift once pypfopt ships a scipy>=1.18 fix.
    "scipy>=1.12,<1.18",
    "pandas>=2.2",
    "polars>=1.0",
    "pyarrow>=15.0",
    # ===================
    # Visualization
    # ===================
    "plotly>=5.20",
    "matplotlib>=3.8",
    "seaborn>=0.13",
    "kaleido==0.2.1",
    # ===================
    # Machine Learning
    # ===================
    "scikit-learn>=1.4",
    "xgboost>=3.1",
    "lightgbm>=4.6",
    "catboost>=1.2.8",
    "optuna>=3.6",
    "optuna-integration[lightgbm]>=4.8",
    "shap>=0.45",
    "tabpfn>=2.0",
    "gymnasium>=1.2",
    "stable-baselines3>=2.7",
    # ===================
    # ML4T Libraries (PyPI, pre-release)
    # ===================
    "ml4t-data>=0.1.0b15",
    "ml4t-diagnostic>=0.1.0b21",
    "ml4t-engineer>=0.1.0b8",
    "ml4t-backtest>=0.1.0b20",
    "ml4t-live>=0.1.0b3",
    # ===================
    # Deep Learning
    # ===================
    "torch>=2.2",
    "pytorch-lightning>=2.2",
    "torchdiffeq>=0.2",
    # torchode (Lienen & Günnemann 2022) would offer per-sample adaptive
    # integration in Ch5 NB04 but its torchtyping dependency is incompatible
    # with torch>=2.10, so the notebook uses torchdiffeq instead.
    "transformers>=4.40,<5",
    "sentencepiece>=0.2",
    "sentence-transformers>=2.6",
    "einops>=0.7",
    "granite-tsfm>=0.3",
    "chronos-forecasting>=1.4",
    # ===================
    # Time Series
    # ===================
    "statsmodels>=0.14",
    "arch>=8.0",
    "pmdarima>=2.0",
    "sktime>=0.30",
    "darts>=0.28",
    "statsforecast>=1.7",
    "hmmlearn>=0.3",
    "filterpy>=1.4",
    "pykalman>=0.11",
    "pywavelets>=1.6",
    "ruptures>=1.1",
    # ===================
    # Causal Inference
    # ===================
    "dowhy>=0.11",
    "econml>=0.15",
    "causalml>=0.15",
    "tigramite>=5.2",
    # tfcausalimpact (BSTS, Ch15 NB06): py312 image only — requires Python <3.13.
    # See envs/py312/pyproject.toml; pycausalimpact (ARIMA, 2020-dead) removed.
    "causal-learn>=0.1",
    "linearmodels>=6.0",
    # ===================
    # NLP / Text
    # ===================
    # gensim: word2vec/Doc2Vec only (Ch10 NB01-02). No Python 3.14 support.
    # Use Docker: docker compose --profile word2vec run --rm word2vec python ...
    "datasets>=2.19",
    "evaluate>=0.4",
    # ===================
    # Portfolio / Finance
    # ===================
    "PyPortfolioOpt>=1.5",
    "riskfolio-lib>=6.0",
    "skfolio>=0.3",
    "vectorbt>=0.26",
    # ===================
    # Data Sources
    # ===================
    "yfinance>=0.2",
    "edgartools>=5.9",
    # sec-edgar-api: removed — no notebook imports it (they use edgartools).
    # sec-edgar-api 1.1.0 is broken with pyrate-limiter 4.x anyway.
    "secedgar>=0.6",
    "iex-parser>=1.7",
    "beautifulsoup4>=4.12",
    "cot-reports>=0.1",
    "databento>=0.47.0",
    "gdown>=5.1",
    "oandapyV20>=0.7",
    # ===================
    # Technical Analysis
    # ===================
    "ta-lib>=0.6.8",
    "ta>=0.11",
    # ===================
    # Synthetic Data
    # ===================
    "be-great>=0.0.5",
    "opacus>=1.4",
    # signatory/esig: separate env (envs/py312/). Only Ch5 NB03, Ch9 NB06/12.
    # ===================
    # Strategy
    # ===================
    "exchange-calendars>=4.5",
    # ===================
    # Utilities
    # ===================
    "python-dotenv>=1.0",
    "pyyaml>=6.0",
    "tqdm>=4.66",
    "rich>=13.7",
    "requests>=2.31",
    "nest-asyncio>=1.6",
    "neo4j>=5.28",
    "thefuzz>=0.22",
    "rapidfuzz>=3.6",
    "psutil>=5.9",
    "sympy>=1.12",
    "openpyxl>=3.1",
    "numba>=0.59",
    "papermill>=2.6",
    # ===================
    # Jupyter
    # ===================
    "jupyter>=1.0",
    "notebook>=7.0",
    "jupytext>=1.16",
    "jupyterlab>=4.1",
    "ipython>=8.22",
    "anywidget>=0.9",
    "nbconvert>=7.16",
    "nbformat>=5.10",
    # ===================
    # File Formats
    # ===================
    "tables>=3.10",
    "duckdb>=1.0",
    "mlflow>=3.10.1",
    "arviz>=0.23.4,<1.0",
    "feast>=0.61",
    "pymc>=5.25.1",
    "pyvis>=0.3.2",
    "llama-index-core>=0.14.20",
    "llama-index-embeddings-huggingface>=0.7.0",
    "chromadb>=1.5.5",
    "pfhedge>=0.22.0",
    "ml4t-models>=0.1.0a4",
    "llama-index-vector-stores-chroma>=0.5.5",
    "llama-index-llms-openai>=0.7.7",
    # Ch24 Autonomous Agents (required at notebook runtime; Docker installs
    # main deps only, so these must not live in an optional-dependencies group)
    "crewai>=0.86",
    # litellm enters transitively via crewai. Pin the security floor directly:
    # crewai's newest 0.x (0.118) caps litellm <1.83, which carries proxy-server
    # CVEs (GHSA OIDC bypass / proxy-config priv-esc). The book only uses litellm
    # as a client lib, but we keep the patched line anyway; this floor holds
    # crewai at 0.95 (last 0.x allowing litellm>=1.83). crewai 1.x needs
    # chromadb<1.2, which conflicts with the Ch22 RAG stack.
    "litellm>=1.83.0",
    "langgraph>=0.2",
    "google-genai>=0.3",
    # Anthropic is the first provider in the Ch24 agent auto-detect cascade
    # (agent_providers.create_llm_client). It is a runtime dependency, not a
    # dev tool — readers with ANTHROPIC_API_KEY set need it in the Docker image.
    "anthropic>=0.89.0",
]

[project.optional-dependencies]
# Database benchmark extras (requires Docker for servers)
db-benchmark = [
    "arcticdb>=4.5",
    "clickhouse-connect>=0.7",
    "psycopg2-binary>=2.9",
    "influxdb-client>=1.45",
    "questdb>=2.0",
]
# Bayesian modeling
bayesian = [
    "pymc>=5.14",
    "arviz>=0.18,<1.0",
]
# MLOps (Ch26)
mlops = [
    "feast>=0.61",
    "mlflow>=2.16",
]
# Live Trading (Ch25)
live = [
    "ib_insync>=0.9",
    "python-okx>=0.3",
    "alpaca-py>=0.33",
]
# Development tools
dev = [
    "pytest>=8.0",
    "pytest-rerunfailures>=15.0",
    "jupytext>=1.16",
    "pre-commit>=3.6",
    "ruff>=0.8",
]

[tool.uv]
# ml4t-engineer only has pre-release versions on PyPI
prerelease = "if-necessary-or-explicit"

# Force protobuf >= 5.0 for Python 3.14 compatibility.
# protobuf 4.x has a C extension metaclass bug on Python 3.14.
# opentelemetry-proto 1.27.0 pins protobuf<5; override until it upgrades.
override-dependencies = [
    "protobuf>=5.0",
]

[build-system]
requires = ["setuptools>=77", "wheel"]
build-backend = "setuptools.build_meta"

[dependency-groups]
dev = [
    "httpx>=0.28.1",
]

[tool.setuptools.packages.find]
include = ["utils", "data", "case_studies"]
namespaces = false

# Auto-loaded at interpreter startup to put numbered chapter dirs on sys.path
# (see sitecustomize.py). Shipped as a top-level module so both the editable
# install (local) and a wheel install resolve it.
[tool.setuptools]
py-modules = ["sitecustomize"]

[tool.ruff]
line-length = 100
# Vendored upstream code (MIT-licensed PatchTST reference) — don't lint/modify.
extend-exclude = ["**/_reference/**"]
# Keep at py312 even though requires-python is >=3.14. With
# target-version = "py314", ruff/pyupgrade rewrites `except (A, B):`
# to PEP 758's unparenthesized `except A | B:` form. Until our minimum
# runtime support and reader tooling catches up to py314, we pin py312
# so ruff preserves the parenthesized tuple-catch form.
target-version = "py312"

[tool.ruff.lint]
select = ["E", "F", "I", "UP", "B", "SIM"]
ignore = [
    "E501",  # Line length handled by formatter
    "E402",  # Module level import not at top - OK for notebooks
    "F401",  # Imported but unused - OK for conditional imports in notebooks
    "F811",  # Redefinition of unused variable - OK in notebook cells
    "B006",  # Mutable argument defaults - common pattern in notebooks
    "B007",  # Loop variable not used - intentional in some cases
    "B018",  # Useless expression - intentional in notebooks to display output
    "B905",  # zip without strict= - bare zip reads cleaner in teaching code
    "F822",  # Undefined in __all__ - forward declarations
    "E741",  # Ambiguous variable names (l, I, O) - common in math/ML code
    "F841",  # Local variable assigned but never used - common in exploratory notebooks
    "SIM102", # Use single if statement - readability preference
    "SIM105", # contextlib.suppress over try/except/pass - explicit form clearer
    "SIM108", # Use ternary operator - readability preference
    "SIM113", # Use enumerate - readability preference
    "SIM118", # Use `key in dict` - both forms acceptable
]

[tool.jupytext]
formats = "ipynb,py:percent"

[tool.pytest.ini_options]
testpaths = ["tests"]
python_files = ["test_*.py"]
python_functions = ["test_*"]
addopts = "-v --tb=short -p no:torchtyping"