Generalized Autoregressive Score (GAS) models are a recent class of observation-driven time-series model for non-normal data. For a conditional observation density p\left(y_{t}\mid{x_{t}}\right) with an observation y_{t} and a latent time-varying parameter x_{t}, we assume the parameter x_{t} follows the recursion:

x_{t} = \mu + \sum^{p}_{i=1}\phi_{i}x_{t-i} + \sum^{q}_{j=1}\alpha_{j}S\left(x_{j-1}\right)\frac{\partial\log p\left(y_{t-j}\mid{x_{t-j}}\right) }{\partial{x_{t-j}}}

For example, for a Poisson distribution density, where the default scaling is \exp\left(x_{j}\right), the time-varying parameter follows:

x_{t} = \mu + \sum^{p}_{i=1}\phi_{i}x_{t-i} + \sum^{q}_{j=1}\alpha_{j}\left(\frac{y_{t-j}}{\exp\left(x_{t-j}\right)} - 1\right)

These types of model can be viewed as approximations to parameter-driven state space models, and are often competitive in predictive performance. See GAS State Space models for a more general class of models that extend beyond the simple autoregressive form. The simple GAS models considered here in this notebook can be viewed as an approximation to non-linear ARIMA processes.


Types of GAS Model

PyFlux supports many types of distribution for GAS modelling, including

PyFlux Class
Poisson GAS GASPoisson()
t GAS GASt()
Skew t GAS GASSkewt()
Normal GAS GASNormal()
Laplace GAS GASLaplace()
Exponential GAS GASExponential()

Below we demonstrate usage with an example for count data.

Poisson GAS for Banking Crisis data

The data below records if a country somewhere in the world experiences a banking crisis in a given year.

import numpy as np
import pyflux as pf
import pandas as pd
import matplotlib.pyplot as plt
%matplotlib inline 

data = pd.read_csv("")
numpy_data = np.sum(data.iloc[:,2:73].values,axis=1)
numpy_data[np.isnan(numpy_data)] = 0
financial_crises = pd.DataFrame(numpy_data)
financial_crises.index = data.year
financial_crises.columns = ["Number of banking crises"]

plt.title("Number of banking crises across the world")
png We will fit an arbitrary GAS(2,2) model to the data and specify the family as GASPoisson():

model = pf.GAS(ar=2,sc=2,data=financial_crises,family=pf.GASPoisson())
x =
Poisson GAS(2,0,2)                                                                                        
======================================================= =================================================
Dependent Variable: Number of banking crises            Method: MLE                                       
Start Date: 1802                                        Log Likelihood: -473.5316                         
End Date: 2010                                          AIC: 957.0632                                     
Number of observations: 209                             BIC: 973.7748                                     
Latent Variable                          Estimate   Std Error  z        P>|z|    95% C.I.                 
======================================== ========== ========== ======== ======== =========================
Constant                                 0.0        0.0144     0.0      1.0      (-0.0282 | 0.0282)       
AR(1)                                    0.4144     1.0631     0.3898   0.6967   (-1.6693 | 2.498)        
AR(2)                                    0.5383     0.9959     0.5405   0.5889   (-1.4136 | 2.4902)       
SC(1)                                    0.2465     0.023      10.7356  0.0      (0.2015 | 0.2916)        
SC(2)                                    0.0725     0.2553     0.2841   0.7763   (-0.4278 | 0.5728)       

We can plot the latent variables using plot_z:

We can plot the model fit using plot_fit:


For in-sample prediction we can use plot_predict_is. The fit_once argument specifies whether to fit the model once, then predict, or fit the model after each time step (rolling):

model.plot_predict_is(h=20, fit_once=True, figsize=(15,5))
If we want to see model forecasts, we can use plot_predict:

model.plot_predict(h=10, past_values=30, figsize=(15,5))
To output the data in DataFrame format, we use predict:

Number of banking crises
2011 9.253661
2012 8.173014
2013 7.910106
2014 7.299097
2015 6.936803
2016 6.504375
2017 6.162003
2018 5.820292
2019 5.521255
2020 5.238532