Getting Started with wARMASVp

In wARMASVp: Winsorized ARMA Estimation for Higher-Order Stochastic Volatility Models

knitr::opts_chunk$set(
  collapse = TRUE,
  comment = "#>",
  fig.width = 6, fig.height = 4
)

The wARMASVp package provides closed-form estimation, simulation, hypothesis testing, filtering, and forecasting for higher-order stochastic volatility SV(p) models. It supports Gaussian, Student-t, and Generalized Error Distribution (GED) innovations, with optional leverage effects.

The SV(p) Model

The stochastic volatility model of order $p$ is:

$$y_t = \sigma_y \exp(w_t / 2)\, z_t$$ $$w_t = \phi_1 w_{t-1} + \cdots + \phi_p w_{t-p} + \sigma_v v_t$$

where $z_t$ is an i.i.d. innovation (Gaussian, Student-t, or GED) and $v_t \sim N(0,1)$ drives the log-volatility.

Simulation and Estimation

Gaussian SV(1)

library(wARMASVp)
set.seed(123)

# Simulate
sim <- sim_svp(2000, phi = 0.95, sigy = 1, sigv = 0.3)
y <- sim$y

# Estimate
fit <- svp(y, p = 1, J = 10)
summary(fit)

Gaussian SV(2)

y2 <- sim_svp(2000, phi = c(0.20, 0.63), sigy = 1, sigv = 0.5)$y
fit2 <- svp(y2, p = 2, J = 10)
summary(fit2)

Student-t Innovations

yt <- sim_svp(2000, phi = 0.90, sigy = 1, sigv = 0.3,
              errorType = "Student-t", nu = 5)$y
fit_t <- svp(yt, p = 1, errorType = "Student-t")
summary(fit_t)

GED Innovations

yg <- sim_svp(2000, phi = 0.90, sigy = 1, sigv = 0.3,
              errorType = "GED", nu = 1.5)$y
fit_ged <- svp(yg, p = 1, errorType = "GED")
summary(fit_ged)

Leverage Effects

When return and volatility shocks are correlated ($\rho \neq 0$), use the leverage option:

sim_lev <- sim_svp(2000, phi = 0.95, sigy = 1, sigv = 0.3,
                   leverage = TRUE, rho = -0.5)
fit_lev <- svp(sim_lev$y, p = 1, leverage = TRUE)
summary(fit_lev)

Leverage is supported for all three distributions:

sim_lev_t <- sim_svp(2000, phi = 0.90, sigy = 1, sigv = 0.3,
                     errorType = "Student-t", nu = 5,
                     leverage = TRUE, rho = -0.5)
fit_lev_t <- svp(sim_lev_t$y, p = 1, errorType = "Student-t", leverage = TRUE)
summary(fit_lev_t)

Hypothesis Testing

The package provides Local Monte Carlo (LMC) and Maximized Monte Carlo (MMC) tests based on Dufour (2006).

AR Order Testing

Test whether SV(1) is sufficient versus SV(2):

y_test <- sim_svp(2000, phi = 0.95, sigy = 1, sigv = 0.3)$y

# H0: SV(1) vs H1: SV(2) — should not reject
test_ar <- lmc_ar(y_test, p_null = 1, p_alt = 2, N = 49)
print(test_ar)

Leverage Testing

test_lev <- lmc_lev(y_test, p = 1, N = 49, Amat = "Weighted")
print(test_lev)

Distribution Testing

Test for heavy tails against a specific null value of the tail parameter:

# Test H0: nu = 10 (mild tails) on Student-t data with true nu = 5
test_t <- lmc_t(yt, nu_null = 10, N = 49, Amat = "Weighted")
print(test_t)

# Directional test: H1: nu < 10 (heavier tails than null)
test_t_dir <- lmc_t(yt, nu_null = 10, N = 49, Amat = "Weighted", direction = "less")
print(test_t_dir)

AR Order Selection via Information Criteria

In addition to the LMC/MMC pairwise AR-order test above, the package selects the SV(p) lag order by information criteria. svp_IC() computes the criteria for a single fitted model; svp_AR_order() sweeps over p = 1, ..., pmax and reports the argmin for each criterion.

fit_ic <- svp(y_test, p = 2, J = 10)
svp_IC(fit_ic)

sel <- svp_AR_order(y_test, pmax = 4, J = 10)
sel$argmin

Four criteria are returned by default, spanning two estimation families and two penalty philosophies: BIC_Kalman / AIC_Kalman use the QML log-likelihood from the Gaussian mixture Kalman filter, while BIC_HR / AIC_HR use a two-stage Hannan-Rissanen ARMA(p, p) residual variance. Additional criteria (AICc_Kalman, BIC_Whittle, and the Yule-Walker variants) are available opt-in via the criteria argument. Both functions read errorType and leverage from the fitted model, so heavy-tailed and leverage specifications are handled automatically. See Ahsan, Dufour, and Rodriguez-Rondon (2026) for the theoretical motivation and consistency simulations.

Filtering

Three methods are available via filter_svp(), which takes a fitted model:

• Corrected Kalman Filter (CKF): Gaussian approximation, fast.
• Gaussian Mixture Kalman Filter (GMKF): KSC (1998) 7-component mixture. Recommended — dominates CKF even under Gaussianity.
• Bootstrap Particle Filter (BPF): Exact density weights. Gold-standard benchmark.

# Fit model
fit_filt <- svp(y, p = 1, J = 10)

# GMKF (recommended)
filt <- filter_svp(fit_filt, method = "mixture")
plot(filt)

Forecasting

Multi-step ahead volatility forecasts using Kalman filtering. Pass a fitted model object from svp():

fit_fc <- svp(sim_lev$y, p = 1, leverage = TRUE)
fc <- forecast_svp(fit_fc, H = 20)
plot(fc)

Output scales can be chosen: "log-variance" (default), "variance", or "volatility". All three are always computed and stored.

References

• Ahsan, M. N. and Dufour, J.-M. (2021). Simple estimators and inference for higher-order stochastic volatility models. Journal of Econometrics, 224(1), 181-197.
• Ahsan, M. N., Dufour, J.-M., and Rodriguez-Rondon, G. (2025). Estimation and inference for higher-order stochastic volatility models with leverage. Journal of Time Series Analysis, 46(6), 1064-1084.
• Ahsan, M. N., Dufour, J.-M., and Rodriguez-Rondon, G. (2026). Estimation and inference for stochastic volatility models with heavy-tailed distributions. Bank of Canada Staff Working Paper 2026-8. doi:10.34989/swp-2026-8
• Dufour, J.-M. (2006). Monte Carlo tests with nuisance parameters: A general approach to finite-sample inference and nonstandard asymptotics. Journal of Econometrics, 133(2), 443-477.

wARMASVp documentation built on May 15, 2026, 5:07 p.m.