Note: This version has never been officially released. Several models have been refactored, improved or bugfixed in 0.8.
The following major new features appear in this version.
Principal Component Analysis¶
Author: Kevin Sheppard
A new class-based Principal Component Analysis has been added. This class replaces the function-based PCA that previously existed in the sandbox. This change bring a number of new features, including:
Options to control the standardization (demeaning/studentizing)
Information criteria for selecting the number of factors
R-squared plots to assess component fit
NIPALS implementation when only a small number of components are required and the dataset is large
Missing-value filling using the EM algorithm
import statsmodels.api as sm from statsmodels.multivariate.pca import PCA data = sm.datasets.fertility.load_pandas().data columns = map(str, range(1960, 2012)) data.set_index('Country Name', inplace=True) dta = data[columns] dta = dta.dropna() pca_model = PCA(dta.T, standardize=False, demean=True) pca_model.plot_scree()
Note : A function version is also available which is compatible with the call in the sandbox. The function version is just a thin wrapper around the class-based PCA implementation.
Regression graphics for GLM/GEE¶
Author: Kerby Shedden
Added variable plots, partial residual plots, and CERES residual plots are available for GLM and GEE models by calling the methods plot_added_variable, plot_partial_residuals, and plot_ceres_residuals that are attached to the results classes.
State Space Models¶
Author: Chad Fulton
State space methods provide a flexible structure for the estimation and analysis of a wide class of time series models. The statsmodels implementation allows specification of state models, fast Kalman filtering, and built-in methods to facilitate maximum likelihood estimation of arbitrary models. One of the primary goals of this module is to allow end users to create and estimate their own models. Below is a short example demonstrating the ease with which a local level model can be specified and estimated:
import numpy as np import statsmodels.api as sm import pandas as pd data = sm.datasets.nile.load_pandas().data data.index = pd.DatetimeIndex(data.year.astype(int).astype(str), freq='AS') # Setup the state space representation class LocalLevel(sm.tsa.statespace.MLEModel): def __init__(self, endog): # Initialize the state space model super(LocalLevel, self).__init__( endog, k_states=1, initialization='approximate_diffuse') # Setup known components of state space representation matrices self.ssm['design', :] = 1. self.ssm['transition', :] = 1. self.ssm['selection', :] = 1. # Describe how parameters enter the model def update(self, params, transformed=True): params = super(LocalLevel, self).update(params, transformed) self.ssm['obs_cov', 0, 0] = params self.ssm['state_cov', 0, 0] = params def transform_params(self, params): return params**2 # force variance parameters to be positive # Specify start parameters and parameter names @property def start_params(self): return [np.std(self.endog)]*2 @property def param_names(self): return ['sigma2.measurement', 'sigma2.level'] # Fit the model with maximum likelihood estimation mod = LocalLevel(data['volume']) res = mod.fit() print res.summary()
The documentation and example notebooks provide further examples of how to form state space models. Included in this release is a full-fledged model making use of the state space infrastructure to estimate SARIMAX models. See below for more details.
Time Series Models (ARIMA) with Seasonal Effects¶
Author: Chad Fulton
A model for estimating seasonal autoregressive integrated moving average models with exogenous regressors (SARIMAX) has been added by taking advantage of the new state space functionality. It can be used very similarly to the existing ARIMA model, but works on a wider range of specifications, including:
Additive and multiplicative seasonal effects
Flexible trend specification
Regression with SARIMA errors
Regression with time-varying coefficients
Measurement error in the endogenous variables
Below is a short example fitting a model with a number of these components, including exogenous data, a linear trend, and annual multiplicative seasonal effects.
import statsmodels.api as sm import pandas as pd data = sm.datasets.macrodata.load_pandas().data data.index = pd.DatetimeIndex(start='1959-01-01', end='2009-09-01', freq='QS') endog = data['realcons'] exog = data['m1'] mod = sm.tsa.SARIMAX(endog, exog=exog, order=(1,1,1), trend='t', seasonal_order=(0,0,1,4)) res = mod.fit() print res.summary()
Generalized Estimating Equations GEE¶
Author: Kerby Shedden
Enhancements and performance improvements for GEE:
EquivalenceClass covariance structure allows covariances to be specified by arbitrary collections of equality constraints #2188
add weights #2090
refactored margins #2158
Author: Kerby Shedden with Saket Choudhary
Enhancements to MixedLM (#2363): added variance components support for MixedLM allowing a wider range of random effects structures to be specified; also performance improvements from use of sparse matrices internally for random effects design matrices.
Other important new features¶
GLM: add scipy-based gradient optimization to fit #1961 (Kerby Shedden)
wald_test_terms: new method of LikelihoodModels to compute wald tests (F or chi-square) for terms or sets of coefficients #2132 (Josef Perktold)
add cov_type with fixed scale in WLS to allow chi2-fitting #2137 #2143 (Josef Perktold, Christoph Deil)
VAR: allow generalized IRF and FEVD computation #2067 (Josef Perktold)
get_prediction new method for full prediction results (new API convention)
Major Bugs fixed¶
see github issues for a full list
bug in ARMA/ARIMA predict with exog #2470
bugs in VAR
x13: python 3 compatibility
Backwards incompatible changes and deprecations¶
List backwards incompatible changes
Development summary and credits¶
Thanks to all of the contributors for the 0.7 release:
Hans-Martin von Gaudecker
Louis-Philippe Lemieux Perreault
These lists of names are automatically generated based on git log, and may not be complete.