dynaris 0.1.2

dynaris is a JAX-powered Python library for state-space models -- from composable DLMs to nonlinear filters, switching systems, Bayesian estimation, and dynamic factor models, all with automatic differentiation and GPU acceleration.

Installation

pip install dynaris

# With Bayesian estimation support
pip install dynaris[bayesian]

Documentation

Full documentation at dynaris.readthedocs.io.

Quickstart

DLM: Trend + Seasonality

from dynaris import LocalLinearTrend, Seasonal, DLM
from dynaris.datasets import load_airline

model = LocalLinearTrend() + Seasonal(period=12)
dlm = DLM(model)
dlm.fit(load_airline()).smooth()
dlm.forecast(steps=24)
dlm.plot(kind="panel")

Nonlinear Filtering

from dynaris import SSM, LorenzAttractor

model = LorenzAttractor(dt=0.01, obs_noise=2.0)
ssm = SSM(model, filter="ukf")  # auto-selects UKF for nonlinear models
ssm.fit(observations)

Regime Switching

from dynaris import LocalLevel, MarkovSwitchingSSM
from dynaris.filters import hamilton_filter
from dynaris.smoothers import kim_smooth
import jax.numpy as jnp

switching = MarkovSwitchingSSM(
    models=(LocalLevel(1, 5), LocalLevel(5, 20)),
    transition_matrix=jnp.array([[0.95, 0.05], [0.10, 0.90]]),
    initial_probs=jnp.array([0.5, 0.5]),
)
result = hamilton_filter(switching, observations)
smoothed = kim_smooth(switching, result)

Bayesian Estimation

from dynaris import LocalLevel, fit_bayesian
from dynaris.estimation.priors import inverse_gamma_log_prior

def model_fn(params):
    return LocalLevel(sigma_level=jnp.exp(params[0]), sigma_obs=jnp.exp(params[1]))

result = fit_bayesian(model_fn, observations, jnp.zeros(2),
                      log_prior_fn=inverse_gamma_log_prior(shape=2.0, scale=1.0))
# result.samples -> (n_samples, n_params) posterior draws

Dynamic Factor Models

from dynaris.models import DFMModel

dfm = DFMModel(n_factors=2)
dfm.fit(panel_data)  # (T, m) multivariate panel
print(dfm.loadings_df())
print(dfm.factor_states_df())
dfm.forecast(steps=12)

Components

Build DLMs by combining components with +:

Filters & Smoothers

Parameter Estimation

Built-in Nonlinear Models

Datasets

Performance

All filters run inside jax.lax.scan with @jax.jit -- GPU/TPU acceleration is automatic. Additional features:

• Batch processing via jax.vmap for parallel multi-series inference
• Memory-efficient long series via jax.checkpoint (trade compute for memory)
• Parallel MCMC chains via NumPyro's chain_method="parallel"
• Pure NumPy backend for lightweight / no-GPU environments

License

MIT License. See LICENSE for details.