pymgarch 0.1.0__tar.gz

pymgarch-0.1.0/.github/workflows/python-publish.yml

@@ -0,0 +1,86 @@
+# .github/workflows/python-publish.yml
+name: pymgarch CI/CD
+
+on:
+  push:
+    branches: [main]
+    tags: ["v*"]
+  pull_request:
+    branches: [main]
+
+permissions:
+  contents: read
+
+jobs:
+  test:
+    runs-on: ubuntu-latest
+    strategy:
+      matrix:
+        python-version: ["3.11", "3.13"]
+    steps:
+    - uses: actions/checkout@v4
+      with:
+        fetch-depth: 0  # full history for hatch-vcs
+
+    - name: Install uv
+      uses: astral-sh/setup-uv@v5
+
+    - name: Setup Python
+      uses: actions/setup-python@v5
+      with:
+        python-version: ${{ matrix.python-version }}
+
+    - name: Install dependencies
+      run: uv sync --dev --extra numba
+
+    - name: Run tests
+      run: uv run pytest tests/ -v --cov=src/pymgarch --cov-report=xml
+
+    - name: Upload coverage to Codecov
+      uses: codecov/codecov-action@v5
+      with:
+        token: ${{ secrets.CODECOV_TOKEN }}
+        files: coverage.xml
+        fail_ci_if_error: false
+
+  release:
+    if: startsWith(github.ref, 'refs/tags/v')
+    needs: test
+    runs-on: ubuntu-latest
+    permissions:
+      contents: write
+    steps:
+    - uses: actions/checkout@v4
+
+    - name: Create GitHub Release
+      # --notes-from-tag publishes the annotated tag message as the release
+      # notes (--generate-notes only summarises PRs, which direct pushes lack);
+      # explicit --title keeps the release title to the bare tag name
+      run: gh release create "${{ github.ref_name }}" --notes-from-tag --title "${{ github.ref_name }}"
+      env:
+        GH_TOKEN: ${{ secrets.GITHUB_TOKEN }}
+
+  deploy:
+    if: startsWith(github.ref, 'refs/tags/v')
+    needs: [test, release]
+    runs-on: ubuntu-latest
+    environment: pypi
+    permissions:
+      id-token: write
+    steps:
+    - uses: actions/checkout@v4
+      with:
+        fetch-depth: 0  # full history for hatch-vcs
+
+    - name: Setup Python
+      uses: actions/setup-python@v5
+      with:
+        python-version: "3.13"
+
+    - name: Build
+      run: |
+        pip install build hatch-vcs
+        python -m build
+
+    - name: Publish to PyPI
+      uses: pypa/gh-action-pypi-publish@release/v1

pymgarch-0.1.0/.gitignore

@@ -0,0 +1,30 @@
+# Python
+__pycache__/
+*.py[cod]
+*.egg-info/
+dist/
+build/
+.venv/
+.python-version
+
+# Generated version file
+src/pymgarch/_version.py
+
+# Tooling caches
+.pytest_cache/
+.ruff_cache/
+.mypy_cache/
+.hypothesis/
+.coverage
+coverage.xml
+
+# Docs build
+docs/_build/
+
+# IDE
+.idea/
+.vscode/
+
+# Local project settings
+.claude/settings.local.json
+uv.lock.orig

pymgarch-0.1.0/.readthedocs.yaml

@@ -0,0 +1,15 @@
+version: 2
+
+build:
+  os: ubuntu-24.04
+  tools:
+    python: "3.13"
+
+sphinx:
+  configuration: docs/conf.py
+
+python:
+  install:
+    - requirements: docs/requirements.txt
+    - method: pip
+      path: .

pymgarch-0.1.0/LICENSE

@@ -0,0 +1,21 @@
+MIT License
+
+Copyright (c) 2026 Marwin Steiner
+
+Permission is hereby granted, free of charge, to any person obtaining a copy
+of this software and associated documentation files (the "Software"), to deal
+in the Software without restriction, including without limitation the rights
+to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+copies of the Software, and to permit persons to whom the Software is
+furnished to do so, subject to the following conditions:
+
+The above copyright notice and this permission notice shall be included in all
+copies or substantial portions of the Software.
+
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+SOFTWARE.

pymgarch-0.1.0/PKG-INFO

@@ -0,0 +1,107 @@
+Metadata-Version: 2.5
+Name: pymgarch
+Version: 0.1.0
+Summary: Multivariate GARCH for Python: DCC, ADCC, CCC on top of arch marginals, validated against rmgarch/tsmarch.
+Project-URL: Documentation, https://pymgarch.readthedocs.io
+Project-URL: Repository, https://github.com/marwinsteiner/pymgarch
+Project-URL: Issues, https://github.com/marwinsteiner/pymgarch/issues
+Author: Marwin Steiner
+License: MIT
+License-File: LICENSE
+Requires-Python: >=3.11
+Requires-Dist: arch>=7.0
+Requires-Dist: numpy>=1.26
+Requires-Dist: pandas>=2.0
+Requires-Dist: scipy>=1.11
+Provides-Extra: numba
+Requires-Dist: numba>=0.59; extra == 'numba'
+Description-Content-Type: text/markdown
+
+# pymgarch
+
+[![CI/CD](https://github.com/marwinsteiner/pymgarch/actions/workflows/python-publish.yml/badge.svg)](https://github.com/marwinsteiner/pymgarch/actions/workflows/python-publish.yml)
+[![PyPI](https://img.shields.io/pypi/v/pymgarch)](https://pypi.org/project/pymgarch/)
+[![Python](https://img.shields.io/pypi/pyversions/pymgarch)](https://pypi.org/project/pymgarch/)
+[![Downloads](https://static.pepy.tech/badge/pymgarch)](https://pepy.tech/project/pymgarch)
+[![codecov](https://codecov.io/gh/marwinsteiner/pymgarch/branch/main/graph/badge.svg)](https://codecov.io/gh/marwinsteiner/pymgarch)
+[![Docs](https://readthedocs.org/projects/pymgarch/badge/?version=latest)](https://pymgarch.readthedocs.io)
+[![License: MIT](https://img.shields.io/badge/license-MIT-blue.svg)](LICENSE)
+
+Multivariate GARCH for Python: DCC, ADCC, and CCC correlation dynamics built
+on top of [arch](https://github.com/bashtage/arch) univariate marginals,
+validated against R's rmgarch/tsmarch.
+
+## Why
+
+Python has no maintained general-purpose multivariate GARCH framework. The
+existing packages cover Gaussian DCC(1,1) at most, while R users have had
+DCC, ADCC, GO-GARCH and copula-GARCH in rmgarch (now tsmarch) for a decade.
+pymgarch closes that gap incrementally, starting with the correlation layer:
+
+- stage 1 (univariate volatility) is delegated to `arch`, the ecosystem's
+  dominant, battle-tested GARCH package;
+- stage 2 (correlation dynamics) is what this library implements, with
+  correct two-stage Engle-Sheppard standard errors and replication tests
+  against rmgarch's fitted parameters and likelihoods.
+
+## Install
+
+```
+pip install pymgarch          # or: pip install pymgarch[numba]
+```
+
+The optional `numba` extra JIT-compiles the correlation recursions; without
+it everything runs in pure NumPy.
+
+## Quickstart
+
+```python
+import pymgarch as mg
+
+# returns: (T, N) DataFrame, percent scale recommended
+res = mg.DCC(dist="t").fit(returns)
+print(res.summary())
+
+res.conditional_correlations   # (T, N, N)
+res.conditional_covariances    # (T, N, N)
+
+fc = res.forecast(horizon=10)                    # analytic
+fc = res.forecast(horizon=10, method="simulation", n_paths=2000)
+flt = res.filter(new_returns)                    # fixed params, new data
+```
+
+Marginals default to constant-mean GARCH(1,1). Customize per-column via a
+spec, or bring your own fitted arch results:
+
+```python
+spec = mg.UnivariateSpec(vol="GARCH", p=1, o=1, q=1, dist="t")  # GJR-t
+res = mg.ADCC().fit(returns, marginals=spec)
+
+from arch import arch_model
+fitted = [arch_model(returns[c], rescale=False).fit(disp="off") for c in returns]
+res = mg.DCC().fit(returns, marginals=fitted)
+```
+
+## Models (v0.1)
+
+| Model | Distribution | Estimation |
+| ----- | ------------ | ---------- |
+| CCC (Bollerslev 1990) | Gaussian | closed form given marginals |
+| DCC(1,1) (Engle 2002) | Gaussian, Student-t | two-stage QML, correlation targeting |
+| ADCC (Cappiello-Engle-Sheppard 2006) | Gaussian, Student-t | two-stage QML, PSD-constrained targeting |
+
+Standard errors are Engle-Sheppard (2001) two-stage sandwich estimates: the
+marginal and correlation scores are stacked so stage-2 uncertainty reflects
+stage-1 estimation error. Correlation targets are held fixed (same
+approximation rmgarch makes). If the stacked system is singular the library
+falls back to a stage-2-only sandwich and says so in `summary()`.
+
+## Roadmap
+
+- v0.2: GO-GARCH via ICA
+- v0.3: copula-GARCH (Gaussian and Student-t copulas)
+- v0.4: composite likelihood for large cross-sections, scalar/diagonal BEKK
+
+## License
+
+MIT

pymgarch-0.1.0/README.md

@@ -0,0 +1,88 @@
+# pymgarch
+
+[![CI/CD](https://github.com/marwinsteiner/pymgarch/actions/workflows/python-publish.yml/badge.svg)](https://github.com/marwinsteiner/pymgarch/actions/workflows/python-publish.yml)
+[![PyPI](https://img.shields.io/pypi/v/pymgarch)](https://pypi.org/project/pymgarch/)
+[![Python](https://img.shields.io/pypi/pyversions/pymgarch)](https://pypi.org/project/pymgarch/)
+[![Downloads](https://static.pepy.tech/badge/pymgarch)](https://pepy.tech/project/pymgarch)
+[![codecov](https://codecov.io/gh/marwinsteiner/pymgarch/branch/main/graph/badge.svg)](https://codecov.io/gh/marwinsteiner/pymgarch)
+[![Docs](https://readthedocs.org/projects/pymgarch/badge/?version=latest)](https://pymgarch.readthedocs.io)
+[![License: MIT](https://img.shields.io/badge/license-MIT-blue.svg)](LICENSE)
+
+Multivariate GARCH for Python: DCC, ADCC, and CCC correlation dynamics built
+on top of [arch](https://github.com/bashtage/arch) univariate marginals,
+validated against R's rmgarch/tsmarch.
+
+## Why
+
+Python has no maintained general-purpose multivariate GARCH framework. The
+existing packages cover Gaussian DCC(1,1) at most, while R users have had
+DCC, ADCC, GO-GARCH and copula-GARCH in rmgarch (now tsmarch) for a decade.
+pymgarch closes that gap incrementally, starting with the correlation layer:
+
+- stage 1 (univariate volatility) is delegated to `arch`, the ecosystem's
+  dominant, battle-tested GARCH package;
+- stage 2 (correlation dynamics) is what this library implements, with
+  correct two-stage Engle-Sheppard standard errors and replication tests
+  against rmgarch's fitted parameters and likelihoods.
+
+## Install
+
+```
+pip install pymgarch          # or: pip install pymgarch[numba]
+```
+
+The optional `numba` extra JIT-compiles the correlation recursions; without
+it everything runs in pure NumPy.
+
+## Quickstart
+
+```python
+import pymgarch as mg
+
+# returns: (T, N) DataFrame, percent scale recommended
+res = mg.DCC(dist="t").fit(returns)
+print(res.summary())
+
+res.conditional_correlations   # (T, N, N)
+res.conditional_covariances    # (T, N, N)
+
+fc = res.forecast(horizon=10)                    # analytic
+fc = res.forecast(horizon=10, method="simulation", n_paths=2000)
+flt = res.filter(new_returns)                    # fixed params, new data
+```
+
+Marginals default to constant-mean GARCH(1,1). Customize per-column via a
+spec, or bring your own fitted arch results:
+
+```python
+spec = mg.UnivariateSpec(vol="GARCH", p=1, o=1, q=1, dist="t")  # GJR-t
+res = mg.ADCC().fit(returns, marginals=spec)
+
+from arch import arch_model
+fitted = [arch_model(returns[c], rescale=False).fit(disp="off") for c in returns]
+res = mg.DCC().fit(returns, marginals=fitted)
+```
+
+## Models (v0.1)
+
+| Model | Distribution | Estimation |
+| ----- | ------------ | ---------- |
+| CCC (Bollerslev 1990) | Gaussian | closed form given marginals |
+| DCC(1,1) (Engle 2002) | Gaussian, Student-t | two-stage QML, correlation targeting |
+| ADCC (Cappiello-Engle-Sheppard 2006) | Gaussian, Student-t | two-stage QML, PSD-constrained targeting |
+
+Standard errors are Engle-Sheppard (2001) two-stage sandwich estimates: the
+marginal and correlation scores are stacked so stage-2 uncertainty reflects
+stage-1 estimation error. Correlation targets are held fixed (same
+approximation rmgarch makes). If the stacked system is singular the library
+falls back to a stage-2-only sandwich and says so in `summary()`.
+
+## Roadmap
+
+- v0.2: GO-GARCH via ICA
+- v0.3: copula-GARCH (Gaussian and Student-t copulas)
+- v0.4: composite likelihood for large cross-sections, scalar/diagonal BEKK
+
+## License
+
+MIT

pymgarch-0.1.0/docs/api.md

@@ -0,0 +1,40 @@
+# API reference
+
+## Models
+
+```{eval-rst}
+.. autoclass:: pymgarch.DCC
+   :members:
+
+.. autoclass:: pymgarch.ADCC
+   :members:
+
+.. autoclass:: pymgarch.CCC
+   :members:
+```
+
+## Marginals
+
+```{eval-rst}
+.. autoclass:: pymgarch.UnivariateSpec
+   :members:
+
+.. autoclass:: pymgarch.MarginalSet
+   :members:
+```
+
+## Results
+
+```{eval-rst}
+.. autoclass:: pymgarch.MGARCHResult
+   :members:
+
+.. autoclass:: pymgarch.MGARCHForecast
+   :members:
+```
+
+## Simulation
+
+```{eval-rst}
+.. autofunction:: pymgarch.simulate_dcc
+```

pymgarch-0.1.0/docs/conf.py

@@ -0,0 +1,26 @@
+project = "pymgarch"
+author = "Marwin Steiner"
+copyright = "2026, Marwin Steiner"
+
+extensions = [
+    "myst_parser",
+    "sphinx.ext.autodoc",
+    "sphinx.ext.napoleon",
+    "sphinx.ext.intersphinx",
+    "sphinx.ext.viewcode",
+]
+
+myst_enable_extensions = ["dollarmath", "amsmath"]
+
+intersphinx_mapping = {
+    "python": ("https://docs.python.org/3", None),
+    "numpy": ("https://numpy.org/doc/stable/", None),
+    "pandas": ("https://pandas.pydata.org/docs/", None),
+    "scipy": ("https://docs.scipy.org/doc/scipy/", None),
+}
+
+html_theme = "furo"
+html_title = "pymgarch"
+
+source_suffix = {".rst": "restructuredtext", ".md": "markdown"}
+exclude_patterns = ["_build"]

pymgarch-0.1.0/docs/index.md

@@ -0,0 +1,27 @@
+# pymgarch
+
+Multivariate GARCH for Python: DCC, ADCC, and CCC correlation dynamics on top
+of `arch` univariate marginals, validated against R's rmgarch.
+
+Python has had no maintained general-purpose multivariate GARCH framework:
+the existing packages stop at Gaussian DCC(1,1), while R users have had a
+full model stack in rmgarch (now tsmarch) for a decade. pymgarch closes that
+gap incrementally. Stage 1 (univariate volatility) is delegated to `arch`;
+stage 2 (correlation dynamics) is what this library implements, with
+Engle-Sheppard two-stage standard errors and replication tests against
+rmgarch's fitted parameters and likelihoods.
+
+```{toctree}
+:maxdepth: 2
+
+quickstart
+models
+inference
+api
+```
+
+## Install
+
+```
+pip install pymgarch          # or: pip install pymgarch[numba]
+```

pymgarch-0.1.0/docs/inference.md

@@ -0,0 +1,53 @@
+# Inference
+
+## Two-stage standard errors
+
+Two-stage QML is consistent but the stage-2 information matrix alone
+understates uncertainty: the correlation parameters are estimated on
+residuals that were themselves standardized by estimated marginals. pymgarch
+implements the Engle and Sheppard (2001) correction by stacking the
+per-observation scores of every marginal likelihood and the correlation
+likelihood into one M-estimator moment vector
+$g_t(\theta)$, $\theta = (\phi_1, \dots, \phi_N, \psi)$, and computing
+
+$$
+\widehat{\mathrm{Avar}}(\hat\theta) = \hat A^{-1} \hat B \hat A^{-\top}/T,
+\qquad
+\hat A = \frac{1}{T}\sum_t \frac{\partial g_t}{\partial \theta'},
+\qquad
+\hat B = \frac{1}{T}\sum_t g_t g_t'.
+$$
+
+$\hat A$ is block lower-triangular because marginal scores do not depend on
+$\psi$. All derivatives are central finite differences; marginal likelihoods
+are re-evaluated through arch's own variance recursions with the backcast and
+variance bounds arch used during fitting held fixed, so the perturbed
+evaluations are consistent with the objective arch actually maximized.
+
+`summary()` labels the estimates `two-stage-robust`. If the stacked bread
+matrix is numerically singular, the library falls back to a stage-2-only
+sandwich (labelled `stage2-robust`) and warns; those standard errors ignore
+marginal estimation error.
+
+## Known approximations
+
+- Correlation targets $\bar S$ and $\bar N$ are held fixed at their point
+  estimates, so targeting uncertainty is ignored. rmgarch makes the same
+  approximation.
+- Set `compute_se=False` on `fit()` to skip the sandwich entirely (useful in
+  simulation studies where only point estimates matter).
+
+## Replication against rmgarch
+
+`scripts/make_fixtures.R` fits rmgarch on a pinned dataset (percent returns
+of five dji30ret constituents) and exports parameters, likelihoods, sigma and
+residual paths, correlation targets, and one-step forecasts to JSON. The
+test suite then:
+
+1. evaluates pymgarch's likelihood at R's fitted parameters on R's own
+   volatility paths and correlation targets, isolating the correlation math
+   from optimizer and univariate-initialization differences; and
+2. runs the full pipeline on the same CSV and requires parameter agreement
+   and a log-likelihood no worse than R's optimum.
+
+The tests skip automatically when the fixture file is absent.

pymgarch-0.1.0/docs/models.md

@@ -0,0 +1,62 @@
+# Models
+
+All stage-2 models operate on standardized residuals $\varepsilon_t = D_t^{-1}
+(r_t - \mu_t)$, where $D_t = \mathrm{diag}(\sigma_{1t}, \dots, \sigma_{Nt})$
+comes from the fitted arch marginals. The conditional covariance is
+$H_t = D_t R_t D_t$.
+
+## CCC (Bollerslev 1990)
+
+Constant correlation $R = \mathrm{cov2cor}(E'E/T)$. No stage-2 parameters;
+closed form given the marginals. Gaussian likelihood only in v0.1.
+
+## DCC(1,1) (Engle 2002)
+
+With correlation targeting $\bar S = \mathrm{cov2cor}(E'E/T)$:
+
+$$
+Q_t = (1 - a - b)\bar S + a\,\varepsilon_{t-1}\varepsilon_{t-1}' + b\,Q_{t-1},
+\qquad
+R_t = \mathrm{diag}(Q_t)^{-1/2} Q_t \,\mathrm{diag}(Q_t)^{-1/2},
+$$
+
+with $a, b \ge 0$, $a + b < 1$, $Q_1 = \bar S$. Estimated by two-stage QML
+(SLSQP, multiple starts). Distributions: Gaussian and covariance-standardized
+multivariate Student-t (degrees of freedom estimated jointly with $a, b$).
+
+## ADCC (Cappiello, Engle and Sheppard 2006)
+
+Scalar asymmetric DCC with $n_t = \min(\varepsilon_t, 0)$ elementwise and
+$\bar N = T^{-1}\sum_t n_t n_t'$:
+
+$$
+Q_t = \left[(1-a-b)\bar S - g\bar N\right]
+ + a\,\varepsilon_{t-1}\varepsilon_{t-1}'
+ + g\,n_{t-1}n_{t-1}'
+ + b\,Q_{t-1}.
+$$
+
+The intercept is positive semi-definite iff $a + b + \delta g \le 1$ with
+$\delta = \lambda_{\max}(\bar S^{-1/2} \bar N \bar S^{-1/2})$, which is a
+linear constraint in $(a, b, g)$ and is imposed exactly during estimation.
+
+## Forecasting
+
+Marginal variance forecasts are delegated to arch. One-step correlations are
+deterministic. For $h \ge 2$, DCC uses the Engle-Sheppard approximation
+
+$$
+E[R_{T+h}] \approx \left(1 - (a+b)^{h-1}\right)\bar S + (a+b)^{h-1} R_{T+1},
+$$
+
+and every model supports Monte Carlo forecasts (`method="simulation"`) that
+propagate the exact recursions; ADCC requires them. The analytic covariance
+assembly $\hat H = \hat D \hat R \hat D$ ignores a Jensen gap that the
+simulation method does not.
+
+## Reported likelihood
+
+`loglikelihood` is the joint likelihood of the returns under the stage-2
+distribution (the same quantity rmgarch's `likelihood()` reports): the
+multivariate density of $\varepsilon_t$ at $R_t$ plus the Jacobian
+$-\sum_i \log \sigma_{it}$, summed over $t$.

pymgarch-0.1.0/docs/quickstart.md

@@ -0,0 +1,61 @@
+# Quickstart
+
+Fit a Gaussian DCC(1,1) with default constant-mean GARCH(1,1) marginals:
+
+```python
+import pymgarch as mg
+
+res = mg.DCC().fit(returns)          # (T, N) DataFrame, percent scale
+print(res.summary())
+
+res.params                            # {"alpha": ..., "beta": ...}
+res.std_errors                        # two-stage robust standard errors
+res.conditional_correlations          # (T, N, N)
+res.conditional_covariances           # (T, N, N)
+```
+
+## Distributions and asymmetry
+
+```python
+mg.DCC(dist="t").fit(returns)         # multivariate Student-t stage 2
+mg.ADCC().fit(returns)                # asymmetric (Cappiello-Engle-Sheppard)
+mg.CCC().fit(returns)                 # constant-correlation baseline
+```
+
+## Custom marginals
+
+Apply one spec to every column, or bring fitted arch results:
+
+```python
+spec = mg.UnivariateSpec(vol="GARCH", p=1, o=1, q=1, dist="t")   # GJR-t
+res = mg.DCC().fit(returns, marginals=spec)
+
+from arch import arch_model
+fitted = [
+    arch_model(returns[c], rescale=False).fit(disp="off")
+    for c in returns.columns
+]
+res = mg.DCC().fit(returns, marginals=fitted)
+```
+
+## Forecasting and filtering
+
+```python
+fc = res.forecast(horizon=10)                                   # analytic
+fc = res.forecast(horizon=10, method="simulation", n_paths=2000)
+fc.variances        # (h, N)
+fc.correlations     # (h, N, N)
+fc.covariances      # (h, N, N)
+
+flt = res.filter(new_returns)   # apply fitted params to new data
+```
+
+ADCC has no standard analytic multi-step correlation forecast; use
+`method="simulation"` there.
+
+## Scaling note
+
+Like `arch`, the optimizers behave best when returns are on a percent scale
+(roughly unit variance). pymgarch fits marginals with `rescale=False` so
+parameters stay comparable with R fits on the same series; multiply decimal
+returns by 100 before fitting.

pymgarch-0.1.0/docs/requirements.txt

@@ -0,0 +1,7 @@
+sphinx>=8.0
+furo>=2024.8.6
+myst-parser>=4.0
+arch>=7.0
+numpy>=1.26
+pandas>=2.0
+scipy>=1.11