hystar_fit {hystar}R Documentation

Estimate the HysTAR model using conditional least squares estimation

Description

This function allows you to estimate the parameters of the hysteretic threshold autoregressive (HysTAR) model.

Usage

hystar_fit(
  data,
  r = c(0.1, 0.9),
  d = 0L,
  p0 = 1L,
  p1 = 1L,
  p_select = "bic",
  thin = FALSE,
  tar = FALSE
)

Arguments

data

a vector, matrix or data.frame containing the outcome variable y in the first column and the threshold variable z in the second. Other columns are ignored. A vector, is taken to be both the outcome and control variable, so, in that case, a self-exciting HysTAR is fitted.

r

A vector or a matrix with search values for \hat{r}_0, \hat{r}_1. Defaults to c(.1, .9).

  • A vector r must contain two values a and b in [0, 1]. The search space for the thresholds will be observed values of z between its 100a\% and 100b\% percentiles.

  • A matrix r allows for a custom search. It must have two columns, such that each row represents a pair r_0 \le r_1 to test. You can use a matrix with one row if you don't want to estimate the thresholds. Note that the values in these matrix should be on the scale of z.

d

A numeric vector with one or more values for the search space of the delay parameter. Defaults to 1. Typically, d is not very large, so a reasonable search space might be 0, 1, 2, ..., 5.

p0

A numeric vector with one or more values for the search space of the autoregressive order of Regime 0. Defaults to 1.

p1

Same as p0, but for regime 1. Note that it does not need to be equal to p0.

p_select

The information criterion that should be minimized to select the orders p_0 and p_1. Choices:

  • "aic" (Akaike Information Criterion)

  • "aicc" (Corrected Akaike Information Criterion)

  • "bic" (default, Bayesian Information Criterion)

thin

TRUE (default) or FALSE. Only relevant when r is a vector.

  • If TRUE (default), the search space for the thresholds are the 100a\%, 100(a+0.01)\%, \dots, 100b\% percentiles of z. This drastically reduces computation costs while keeping a reasonably large search space for the thresholds. Note that this is a purely practical choice with no theoretical justification.

  • If FALSE, all observed unique values of z between the 100a\% and 100b\% percentiles of z will be considered.

tar

TRUE or FALSE (default). Choose TRUE if you want to fit a traditional 2-regime threshold autoregressive (TAR) model. In this model, there is only one threshold (or equivalently, a HysTAR model with r_0 = r_1).

Details

In regime 0, y_{t} is predicted by values up to y_{t - p_0}. This implies that the first p_0 time points can not be predicted. E.g., if p_0 = 2, y_1 would miss a value from y_{-1}. Similarly, the value of the delay parameter implies that the regime is unknown for the first d time points. To ensure that the same data are used on all options for d, p0 and p1, the first max(d, p0, p1) observations are discarded for estimation of the parameters.

Value

An object of S3 class hystar_fit, which is a list containing the following items:

Implemented generics for the hystar_fit class:

The HysTAR model

The HysTAR model is defined as:

y_t = \begin{cases} \phi_{00} + \phi_{01} y_{t-1} + \cdots + \phi_{0 p_0} y_{t-p_0} + \sigma_{0} \epsilon_{t} \quad \mathrm{if}~R_{t} = 0 \\ \phi_{10} + \phi_{11} y_{t-1} + \cdots + \phi_{1 p_1} y_{t-p_1} + \sigma_{1} \epsilon_{t} \quad \mathrm{if}~R_{t} = 1, \\ \end{cases}

with R_t = \begin{cases} 0 \quad \quad \mathrm{if} \, z_{t-d} \in (-\infty, r_{0}] \\ R_{t-1} \quad \mathrm{if} \, z_{t-d} \in (r_0, r_1] \\ 1 \quad \quad \mathrm{if} \, z_{t-d} \in (r_1, \infty), \\ \end{cases}

where p_j denotes the order of regime j \in \{0,1\} with coefficients \phi_{j0}, \dots, \phi_{j p_j \in (-1, 1)}, \sigma_{j} is the standard deviation of the residuals, and d \in \{0, 1, 2, \dots\} is a delay parameter. The parameters of primary interest are the thresholds r_0 \le r_1. We let t = 0, 1, 2, ..., T, where T is the number of observations.

Author(s)

Daan de Jong.

References

Li, Guodong, Bo Guan, Wai Keung Li, en Philip L. H. Yu. ‘Hysteretic Autoregressive Time Series Models’. Biometrika 102, nr. 3 (september 2015): 717–23.

Zhu, Ke, Philip L H Yu, en Wai Keung Li. ‘Testing for the Buffered Autoregressive Process’. Munich Personal RePEc Archive, (november 2013).

Examples

z <- z_sim(n_t = 200, n_switches = 5, start_regime = 1)
sim <- hystar_sim(z = z, r = c(-.5, .5), d = 2, phi_R0 = c(0, .6), phi_R1 = 1)
plot(sim)
fit <- hystar_fit(sim$data)
summary(fit)

[Package hystar version 1.0.0 Index]