About PyBSTS

PyBSTS is an adaptation of R's implementation of Steven L. Scott's BSTS library. It has similar interface, but re-written for Python memory model. It is a Cython+Numpy based implementation and thus dependendecies for these packages.

Package Contents

PyBSTS package installs pybsts and causal_impact libraries.

Quick start

1. Install the package

pip install pybsts

2. Build BSTS model

import pybsts
import numpy as np

y = np.array([1.0, 2.0, 3.0, 4.0, 4.5])

specification = {"ar_order": 1, "local_trend": {"local_level": True},
                 "sigma_prior": np.std(y, ddof=1), "initial_value": y[0]}
b = pybsts.PyBsts("gaussian", specification, {"ping": 10, "niter":100, "seed": 1, "burn": 10})
b.fit(y, seed=1)
res = b.predict(seed=1)
print(res)

y = np.array([1.0, 2.0, 3.0, 4.0, 4.5])
X = np.array([[1.0, 2.0, 0.0, 0.0, 0.0], [0.0, 0.0, 3.0, 4.0, 0.0]])


specification = {"local_trend": {"static_intercept": True},
                 "predictors_prior": {"predictors_squared_normalized": np.dot(X, X.T)/X.shape[1]},
                 "sigma_prior": np.std(y, ddof=1), "initial_value": y[0], "mean_value": np.mean(y), 
                 "predictor_names": ["first", "second"]}
b = pybsts.PyBsts("gaussian", specification, 
                  {"ping": 10, "niter":100, "burn": 10, "forecast_horizon": 2, "seed": 1})
b.fit(X.T, y, seed=1)
res = b.predict(np.array([[1.0, 0.0], [2.0, 0.0]]), [6, 7], seed=1)
print(res)

3. Build CausalImpact model

import causal_impact
import numpy as np

y = np.array([1.0, 2.0, 3.0, 4.0, 4.5, 3.5, 2.5, 2.6])
X = np.array([[1.0, 2.0, 0.0, 0.0, 0.0, 3.5, 0.0, 0.0], [0.0, 0.0, 3.0, 4.0, 4.4, 0.0, 2.5, 2.5]])
                 
b = causal_impact.CausalImpact(X, y, range(0, 5), range(6, 8), niter=1000, burn=100, seed=1, 
                               seasons=[{"number_of_seasons": 3, "duration": 1}])
res = b.analyze()
print(res[0], res[1])
print(b.summary())

Current status

Here is a list of implemented models (see BSTS Library Documentation ):

• stationary (non-dynamic) gaussian regression with local.level + seasons
• stationary (non-dynamic) gaussian regression with local.linear.trend + seasons
• stationary (non-dynamic) gaussian regression with semilocal.linear.trend + seasons
• stationary (non-dynamic) gaussian regression with static.intercept + seasons
• any of the above + ar
• any of the above + auto.ar