edynamics

Empirical dynamic modelling toolkit: delay-embedding state-space reconstruction with locally-weighted projection methods, organised as a small set of composable building blocks (embeddings, observers, kernels, norms, projectors, estimators).

Install

From PyPI:

pip install edynamics

From a checkout, into a Python 3.10+ environment:

pip install -e .

Dependencies (numpy, pandas, scipy, tqdm, ray, torch) are declared in pyproject.toml and installed automatically.

If a build environment can't reach PyPI (e.g. air-gapped or Compute Canada CVMFS with pre-installed wheels), preinstall the runtime deps and add --no-build-isolation:

pip install numpy pandas scipy tqdm ray torch wheel setuptools
pip install -e . --no-build-isolation

Note on torch

torch is a heavy dependency (~1 GB). It is required because two modules (projectors/weighted_least_squares.py, estimators/local_gl_selector.py) use it for vectorised batched operations. If you only need Embedding, Lag, Minkowski, or the non-WLS kernels, the install will still pull torch because declaring it as runtime-optional would mask the import failure on modules that depend on it.

Upgrading from 0.3.x → 0.4.0

0.4.0 introduces breaking changes to the WeightedLeastSquares projector. See CHANGELOG.md for the full migration. The two largest:

• WeightedLeastSquares now requires you to first run LocalGLSelector.fit() on a library to populate per-anchor (θ*, σ*) before calling .project(). The old global_theta=... constructor argument is gone.
• The result type returned by .project() was renamed WLSResult → RoseResult. The DataFrame access pattern (result.predictions) is unchanged; the new .evaluate(embedding) method computes per-lead error and persistence-skill metrics.

Minor PEP 8 fix: kernels.constant (the class) renamed to kernels.Constant.

What's in it

• edynamics.modelling_tools.embeddings.Embedding — the delay embedding container; binds raw data to a set of Observers (e.g. Lag) and exposes a block of state-vector rows indexed by time.
• edynamics.modelling_tools.observers — Observer (ABC), Lag, LagMovingAverage, ColumnObserver.
• edynamics.modelling_tools.kernels — Kernel (ABC), Constant, Gaussian, Exponential, Epanechnikov, Tricubic.
• edynamics.modelling_tools.norms — Norm (ABC), Minkowski.
• edynamics.modelling_tools.projectors.WeightedLeastSquares — locally-weighted least-squares projection with separate drift (theta) and residual (sigma) bandwidths, selected per anchor by LocalGLSelector.
• edynamics.modelling_tools.projectors.KNearestNeighbours — simplex projection.
• edynamics.modelling_tools.estimators.LocalGLSelector — per-anchor joint (theta, sigma) grid search via the Goldenshluger-Lepski criterion.
• edynamics.modelling_tools.estimators.dimensionality — embedding- dimension prediction-skill sweep.
• edynamics.data_sets.lorenz_data — Lorenz attractor trajectory (for testing and examples).

Minimal example

import pandas as pd
from edynamics.modelling_tools import Embedding, Lag
from edynamics.modelling_tools.projectors import WeightedLeastSquares
from edynamics.modelling_tools.kernels import Gaussian
from edynamics.modelling_tools.estimators import LocalGLSelector

# 1. Build an embedding from a 1-D time series
data = pd.DataFrame(...)                # DatetimeIndex x one column
lags = [Lag(variable_name="x", tau=-i) for i in range(3)]
emb  = Embedding(data=data, observers=lags, library_times=data.index[3:-1])
emb.compile()

# 2. Set up the projector with drift + residual kernels
wls = WeightedLeastSquares(
    kernel=Gaussian(theta=1.0, dim=emb.dimension),
    residual_kernel=Gaussian(theta=1.0, dim=emb.dimension),
)

# 3. Select per-anchor (theta*, sigma*)
import numpy as np
sel = LocalGLSelector(
    theta_grid=np.logspace(-1, 1.5, 30),
    sigma_grid=np.logspace(-1, 1.5, 30),
    lwls=wls, C=2.0,
)
sel.fit(emb, library_times=emb.library_times[::10])

# 4. Project / forecast
qry = emb.get_points(emb.library_times[-5:])
result = wls.project(embedding=emb, points=qry, steps=1, step_size=1)
print(result.predictions)
print(result.evaluate(emb))             # MSE / RMSE / MAE / Skill per lead

See examples/wls_lorenz_demo.py for the full end-to-end Lorenz demo with plotting.

Development

Tests:

pip install pytest
pytest tests/

The suite covers 87 % of source lines; new contributions should include matching tests.

License

MIT — see LICENSE.