Skip to contents

Changelog

Source: NEWS.md

stochvolTMB 0.4.0

Breaking changes

  • The skew-Gaussian models estimate the skewness instead of the shape alpha. The fixed parameter alpha is replaced by logit_skewness. summary(report = "transformed") and coef() report both skewness and alpha, and sim_sv() still takes alpha.
  • residuals() uses the exact grid filter by default (method = "grid") and returns residuals for all observations. Use method = "tmb" for TMB::oneStepPredict.
  • demo() is deprecated. Use run_shiny_demo().

New features and fixes

  • New var_es() calculates the one-day Value at Risk and Expected Shortfall from the exact one-step-ahead predictive distribution, for every observation and for the next day after the data. It uses the same filter on a grid as residuals().

  • New backtest_var() backtests the Value at Risk with the unconditional coverage test of Kupiec (1995), the independence test of Christoffersen (1998) and their combination, and reports the ratio of the realized losses to the Expected Shortfall on the exceedances.

  • New forecast_var_es() simulates Value at Risk and Expected Shortfall for several days ahead, for the daily and the cumulative returns.

  • New plot_var() plots the returns with the Value at Risk and Expected Shortfall and marks the exceedances.

  • New compare_models() fits several models and ranks them by AIC. It also reports if the optimizer converged and if the Hessian is positive definite, and it keeps the fitted models.

  • predict() for the leverage model now uses the last observed return for the first volatility shock, as for the skew models with leverage. Before, the first predicted volatility ignored the correlation with the last return.

  • New model skew_gaussian_leverage that combines skew-Gaussian returns with the leverage effect. The skew-Gaussian return shock drives the volatility shock, eta_t = rho * z_t + sqrt(1 - rho^2) * e_t. This order keeps the Laplace approximation accurate. The other order (skew noise added to rho * eta_t) made the approximation overestimate the likelihood, and the fit diverged to rho = -1 and |alpha| -> Inf.

  • The skew-Gaussian models now estimate the skewness instead of the shape alpha. The fixed parameter alpha is replaced by logit_skewness, and the skewness is bounded by |skewness| < 0.99 (|alpha| <= 27). summary(report = "transformed") reports both skewness and alpha. Before, the optimizer started at alpha = 0, where the likelihood is flat (the score is exactly zero), and the fit often stayed at alpha = 0.

  • The skew-Gaussian density now uses a numerically stable log(pnorm(x)) instead of log(1e-5 + pnorm(x)).

  • New model skew_t: the skew t distribution of Jones and Faddy (2003), standardized to mean zero and unit variance, with separate degrees of freedom for the left and the right tail (df_left, df_right). Supported by estimate_parameters(), sim_sv(), predict(), residuals() and plot_returns().

  • New model skew_t_leverage: the skew t return shock drives the volatility shock, as in skew_gaussian_leverage. It estimates df_left, df_right and rho. On the S&P 500 data it has the smallest AIC of all models, and its one-step-ahead residuals have a straight Q-Q plot. residuals(method = "tmb") is not available for this model.

  • The degrees of freedom of the t and skew_t models are capped at 10,000. For larger values the lgamma differences in the density lost all precision, and the fit of the t model could run to df = 1e14 with a meaningless likelihood. At 10,000 degrees of freedom the distribution is gaussian for all practical purposes.

  • New plot_returns() plots the returns with the central interval of the fitted conditional distribution and marks the returns outside it.

  • New plot_diagnostics() plots a Q-Q plot and the autocorrelation of the one-step-ahead residuals and of the squared residuals.

  • residuals() has a new default method, method = "grid". It calculates the one-step-ahead residuals exactly with a filter on a grid for the log-volatility, for all observations in a few seconds. The old method, TMB::oneStepPredict with the Laplace approximation, takes one to three seconds per observation and is available with method = "tmb". With method = "grid", residuals() returns residuals for all observations by default.

  • residuals(method = "tmb") no longer fails with “Multiple TMB models loaded” when several TMB models are loaded, for example after devtools::load_all(). It then calculates the residuals without parallel processing.

  • residuals() is now an S3 method, residuals.stochvolTMB(). Before, the package exported a plain residuals() function that hid stats::residuals(), so for example residuals(lm(...)) failed when stochvolTMB was attached. Calls of the form residuals(fit) do not change.

  • New coef() and vcov() methods. By default they return the parameters on their natural scale (for example sigma_y, phi and rho) and their covariance matrix. Use scale = "estimation" for the parameters on the scale they were estimated.

  • The shiny demo is now started with run_shiny_demo(). demo() without arguments still runs it, with a deprecation warning. With arguments, demo() calls utils::demo(), so demo(topic) works again when stochvolTMB is attached.

  • Plots use linewidth instead of the deprecated size for lines. ggplot2 (>= 3.4.0) is required.

stochvolTMB 0.2.0

CRAN release: 2021-08-13

  • Added predict function that makes it possible to simulate future volatility and log-returns.
  • plot has a new forecast argument that makes it possible to include forecasted volatility.
  • The degrees of freedom for the t model is estimated on log scale and shifted by 2 so that it is guaranteed to be greater than 2. This is to ensure that the variance is finite.
  • Changed parameter names from phi_logit and rho_logit to logit_phi and logit_rho to be consistent with other parameter names.
  • Updated README.
  • Updated documentation.

stochvolTMB 0.1.2

CRAN release: 2020-09-04

  • CRAN release