GAS models¶
Example¶
1 2 3 4 5 6 7 8 | import numpy as np
import pandas as pd
import pyflux as pf
d = pd.read_csv("https://vincentarelbundock.github/io/Rdatasets/csv/MASS/drivers.csv")
accidents = d['drivers'].values
model = pf.GAS(data=accidents,ar=1,sc=1,dist='Poisson')
|
Class Arguments¶
The GAS() model class has the following arguments:
- data : requires a pd.DataFrame object or an np.array
- dist : one of [‘Normal’, ‘Laplace’, ‘Poisson’, ‘Exponential’]
- ar : the number of autoregressive lags
- sc : the number of conditional score lags
- integ : (default : 0) order of integration (0 : no difference, 1 : first difference, ...)
- target : (default: None) specify the pandas column name or numpy index if the input is a matrix. If None, the first column will be chosen as the data.
Class Attributes¶
A GAS() object holds the following attributes:
Model Attributes:
- dist : one of [‘Normal’, ‘Laplace’, ‘Poisson’, ‘Exponential’]
- ar : the number of autoregressive lags
- sc : the number of conditional score lags
- integ : order of integration (0 : no difference, 1 : first difference, ...)
- param_no : number of model parameters
- link : the model link function
- scale : whether or not the model has a scaling parameter
- index : the timescale of the time-series
- data : the dependent variable held as a np.array
- data_name : string variable containing name of the time series
- data_type : whether original datatype is numpy or pandas
Parameter Attributes:
The attribute param.desc is a dictionary holding information about individual parameters:
- name : name of the parameter
- index : index of the parameter (begins with 0)
- prior : the prior specification for the parameter
- q : the variational distribution approximation
Inference Attributes:
- params : holds any estimated parameters
- ses : holds any estimated standard errors for parameters (MLE/MAP)
- ihessian : holds any estimated inverse Hessian (MLE/MAP)
- chains : holds trace information for MCMC runs
- supported_methods : which inference methods are supported
- default_method : default inference method
Class Methods¶
adjust_prior(index,prior)
Adjusts a prior with the given parameter index. Arguments are:
- index : taking a value in range(0,no of parameters)
- prior : one of the prior objects listed in the Bayesian Inference section
1 2 | model.list_priors()
model.adjust_prior(2,ifr.Normal(0,1))
|
fit(method)
Fits parameters for the model. Arguments are:
- method : one of [‘BBVI’,MLE’,’MAP’,’M-H’,’Laplace’]
- printed : (default: True) whether to print output
- nsims : (default: 100000) how many simulations if M-H is chosen
- cov_matrix (default: None) covariance matrix for M-H
- iterations : (default: 30000) how many iterations if BBVI is chosen
- step : (default: 0.001) step size for BBVI
1 | model.fit("M-H",nsims=20000)
|
list_priors()
Lists the current prior specification.
plot_fit()
Graphs the fit of the model.
predict(h)
Predicts h timesteps ahead. Arguments are:
- h : (default: 5) how many timesteps to predict ahead
- past_values : (default: 20) how many past observations to plot
- intervals : (default: True) whether to plot 95/90 prediction intervals
1 | model.predict(h=12,past_values=36)
|