smcs 0.1.1

JAX-based Sequential Monte Carlo library for time series prediction

smcs

smcs is a JAX-based Sequential Monte Carlo library for time series prediction. It combines the academic rigor of particles, the functional design patterns of BlackJAX, and the state-space model abstractions of Dynamax.

Features

• JAX-native: Full JIT compilation and GPU/TPU support
• Type-safe: Comprehensive jaxtyping annotations with runtime checking
• Multiple SMC algorithms:
  • Bootstrap Particle Filter
  • Auxiliary Particle Filter
  • Liu-West Filter (online parameter learning)
  • Storvik Filter (sufficient statistics)
  • SMC² (nested SMC for parameters)
  • PMMH (Particle MCMC)
  • Waste-Free SMC
• State-space models:
  • Dynamic Linear Models (Local Level, Local Linear Trend)
  • ARIMA/SARIMA
  • Stochastic Volatility
  • GARCH family
  • Dynamic Factor Models
  • Regime-switching models
• High-level forecasting agents for easy use
• Pandas DataFrame integration

Installation

pip install smcs

For development:

pip install smcs[dev]

Quick Start

Using High-Level Agents

import jax.numpy as jnp
from smcs import LocalLevelAgent, SMCConfig
from smcs.io import from_dataframe

# Load data
data, timestamps = from_dataframe(df)

# Create and configure agent
config = SMCConfig(n_particles=1000, seed=42)
agent = LocalLevelAgent(config)

# Fit model
agent.fit(data, timestamps)

# Generate forecasts
forecast = agent.predict(horizon=10)
print(f"Forecast mean: {forecast.mean}")
print(f"95% interval: [{forecast.quantiles[0.05]}, {forecast.quantiles[0.95]}]")

# Online update with new observation
agent.update(jnp.array([new_value]))

Using Low-Level API

import jax
import jax.numpy as jnp
from smcs import (
    run_bootstrap_filter,
    LocalLevelModel,
    LocalLevelParams,
)

# Define model and parameters
model = LocalLevelModel()
params = LocalLevelParams(
    sigma_obs=0.5,
    sigma_level=0.1,
    m0=0.0,
    C0=1.0,
)

# Generate synthetic data
key = jax.random.PRNGKey(42)
observations = jax.random.normal(key, shape=(100, 1))

# Run particle filter
filter_key = jax.random.PRNGKey(123)
state, info = run_bootstrap_filter(
    filter_key,
    observations,
    model,
    params,
    n_particles=1000,
)

print(f"Log-likelihood: {state.log_likelihood:.4f}")
print(f"Final ESS: {info.ess[-1]:.1f}")

Parameter Learning with Liu-West Filter

from smcs import run_liu_west_filter, LocalLevelModel

model = LocalLevelModel()

# Define parameter conversion
def param_to_model_params(param_vec):
    return LocalLevelParams(
        sigma_obs=jnp.exp(param_vec[0]),
        sigma_level=jnp.exp(param_vec[1]),
        m0=0.0,
        C0=1.0,
    )

# Run Liu-West filter
state, info = run_liu_west_filter(
    key,
    observations,
    model,
    param_to_model_params,
    initial_state_sampler,
    initial_param_sampler,
    n_particles=1000,
    delta=0.98,
)

# Get estimated parameters
estimated_params = state.weighted_param_mean()

Architecture

smcs/
├── src/smcs/
│   ├── core/           # Particles, resampling, ESS computation
│   ├── algorithms/     # SMC algorithm implementations
│   ├── models/         # State space model definitions
│   ├── agents/         # High-level forecasting agents
│   ├── config/         # Pydantic configuration
│   └── io/             # DataFrame utilities
├── tests/
└── docs/

SMC Algorithms

Algorithm	Use Case	Complexity
Bootstrap PF	Basic filtering	O(NT)
Auxiliary PF	Informative observations	O(NT)
Liu-West	Online parameter learning	O(NT)
Storvik	Models with sufficient statistics	O(NT)
SMC²	Full online parameter learning	O(Nθ×Nx×T)
PMMH	Batch parameter learning	O(N×MCMC)
Waste-Free	Efficient MCMC utilization	O(NT)

State Space Models

Model	Description
Local Level	Random walk + noise
Local Linear Trend	Level + slope dynamics
ARIMA	Autoregressive integrated moving average
Stochastic Volatility	Time-varying volatility
GARCH	Deterministic volatility dynamics
Dynamic Factor	Multivariate with latent factors
Regime-Switching	Markov-switching dynamics

Configuration

from smcs import SMCConfig

config = SMCConfig(
    n_particles=1000,          # Number of particles
    seed=42,                   # Random seed
    ess_threshold=0.5,         # ESS/N ratio for resampling
    resampling_method="systematic",  # Resampling algorithm
    liu_west_delta=0.98,       # Liu-West discount factor
    jit_compile=True,          # JIT compilation
)

Development

# Clone repository
git clone https://github.com/smcs-authors/smcs.git
cd smcs

# Install in development mode
pip install -e ".[dev]"

# Run tests
pytest

# Run linter
ruff check src tests

# Type checking
mypy src

References

1. Gordon, N. J., Salmond, D. J., & Smith, A. F. (1993). Novel approach to nonlinear/non-Gaussian Bayesian state estimation. IEE Proceedings F.
2. Pitt, M. K., & Shephard, N. (1999). Filtering via simulation: Auxiliary particle filters. JASA.
3. Liu, J., & West, M. (2001). Combined parameter and state estimation in simulation-based filtering.
4. Chopin, N., Jacob, P. E., & Papaspiliopoulos, O. (2013). SMC²: an efficient algorithm for sequential analysis of state space models. JRSS-B.
5. Andrieu, C., Doucet, A., & Holenstein, R. (2010). Particle Markov chain Monte Carlo methods. JRSS-B.
6. Dau, H. D., & Chopin, N. (2022). Waste-free sequential Monte Carlo. JRSS-B.

License

MIT License - see LICENSE for details.