dfm: Estimate a Dynamic Factor Model

In srlanalytics/bdfm: Bayesian and Maximum Likelihood Estimation of Dynamic Factor Models

Description

Estimates a Bayesian or non-Bayesian dynamic factor Model. With the default options, dfm calls automatic procedures that works well in many circumstances.

Usage

dfm(data, factors = 1, lags = "auto", forecasts = 0,
  method = c("bayesian", "ml", "pc"), scale = TRUE, logs = "auto",
  diffs = "auto", outlier_threshold = 4, frequency_mix = "auto",
  pre_differenced = NULL, trans_prior = NULL, trans_shrink = 0,
  trans_df = 0, obs_prior = NULL, obs_shrink = 0, obs_df = NULL,
  identification = "pc_long", keep_posterior = NULL,
  interpolate = FALSE, orthogonal_shocks = FALSE, reps = 1000,
  burn = 500, verbose = interactive() &&
  !isTRUE(getOption("knitr.in.progress")), tol = 0.01)

Arguments

data	one or multiple time series. The data to be used for estimation. This can be entered as a "ts" object or as a matrix. If tsbox is installed, any ts-boxable time series can be supplied (ts, xts, zoo, data.frame, data.table, tbl, tbl_ts, tbl_time, or timeSeries)
factors	integer. The number of unobserved factors to be estimated. A larger number of factors leads to a more complex model. Denoted as 'm' in the documentation.
lags	integer. The number of lags in the transition equation. If "auto" (default), the number is equal to highest frequency in data. Denoted as 'p' in the documentation.
forecasts	integer. Number of periods ahead to forecasts.
method	character. Method to be used; one of "bayesian", "ml" or "pc". See details.
scale	logical. Should data be scaled before estimation? TRUE (default) resolves some numerical problems during estimation. FALSE ensures that the coefficient estimates are interpretable.
logs	names or index values (see details). Series of which the logarithm is taken (can be combined with diffs). If "auto" (default) this is done for all series that are differentiated and have no values < 0.
diffs	names or index values (see details). Series to be differentiated. If "auto" (default), a modified Durbin-Watson test is performed.
outlier_threshold	integer. Observations more than outlier_threshold standard deviations from the series mean are removed. This is useful to increase the stability of the estimation.
frequency_mix	integer or "auto". Number of high frequency periods in a low frequency period. If "auto" (default), this is inferred from the time series.
pre_differenced	names or index values (see details). series entered in differences (If series are specified in diffs, this is not needed.)
trans_prior	m x mp (m: factors, p: lags) prior matrix for B (the transition matrix) in the transition equation. Default is zeros. E.g., to use a random walk prior with m factors and p lags, set trans_prior = cbind(diag(1,m,m), matrix(0,m,m*(p-1))).
trans_shrink	numeric. Prior tightness on B matrix in transition equation where a value of zero is used to denote an improper (flat) prior (i.e. no shrinkage). Use to shrink forecast values towards the prior trans_prior, which may help reduce parameter uncertainty in estimation.
trans_df	numeric. Prior degrees of freedom for inverse-Wishart distribution of shocks in the transition equation, where 0 implies no shrinkage. Shrinking shocks to the transition equation will increase the magnitude of shocks to the observation equation dampening updates from observed series. High values of trans_df can lead to instability in simulations.
obs_prior	k x m (k: observed series, m: factors) prior matrix for H (loadings) in the observation equation Default is zeros.
obs_shrink	numeric. Prior tightness on H (loadings) in the observation equation where a value of zero is used to denote an improper (flat) prior (i.e. no shrinkage). A greater value will shrink estimates of loadings more aggressively towards the prior obs_prior. When the prior is zero (the default value), this is an alternative (and typically more stable) approach to dampening the impact of updates from observed series.
obs_df	named vector (see details). prior degrees of freedom for inverse chi-squared distribution in the observation equation. This is useful to give specific series a larger weight, e.g. 1. (default 0).
identification	names or index values (see details), or character. Factor identification. "pc_long" (default) identifies on principal components from series with at least the median number of observations. "pc_wide" identifies on principal components using all series, where rows of the observations matrix containing missing data are omitted. "name" uses Stock and Watson's "naming factors" identification, i.e. identifying on the first m series provided where m is the number of factors. Identification can also be done manually, by supplying names or index values from which identifying series are derived via principal components.
keep_posterior	names or index values (see details). Series of which to keep the full posterior distribution of predicted values (method "bayesian" only). This is useful for forecasting as the posterior median forecast value tends to me more accurate than forecasts using the posterior median parameter estimates, and allows for the evaluation of forecast accuracy.
interpolate	logical. Should output return intra-frequency estimates of low frequency observables? Put differently, if the model includes monthly and quarterly data, should output include estimates of quarterly data every month (where quarterly refers to an aggregate of the current and previous 2 months; for interpolate = TRUE) or just at the end of the quarter (months 3, 6, 9, and 12; for interpolate = FALSE, default)?
orthogonal_shocks	logical. Return a rotation of the model with orthogonal shocks and factors. This is used to isolate the impact of each factor on observables, allowing for a clean interpretation of how shocks (which, if TRUE are not correlated) impact observed series.
reps	integer. Number of repetitions for MCMC sampling
burn	integer. Number of iterations to burn in MCMC sampling
verbose	logical. Print status of function during evaluation. Default is TRUE in interactive mode, FALSE otherwise, so it does not appear, e.g., in reprex::reprex().
tol	numeric. Tolerance for convergence of EM algorithm (method "ml" only). The default value is 0.01 which corresponds to the convergence criteria used in Doz, Giannone, and Reichlin (2012).

Details

Specifying series: Individual series can be specified either by names (recommended) or index values. An index value refers to the position of the series in data.

Specifying parameters for specific series: Parameters for individual series can be specified using a named vector (recommended) or using a unnamed vector of the same length as as the number of series in data.

See Also

vignette("dfm"), for a more comprehensive intro to the package.

Practical Implementation of Factor Models for a comprehensive overview of dynamic factor models.

Examples

dta <- cbind(fdeaths, mdeaths)

m0 <- dfm(dta, forecast = 2) # estimation with 2 period forecast
predict(m0)                  # series with imputations and forecasts
summary(m0)                  # summary of the model
factors(m0)                  # estimated factor

# informative priors: giving 'fdeaths' a higher weight
m1 <- dfm(dta, obs_df = c("fdeaths" = 1))
summary(m1)

## Not run: 
# Forecasting U.S. GDP
m1 <- dfm(econ_us,
  pre_differenced = "A191RL1Q225SBEA",
  keep_posterior = "A191RL1Q225SBEA"
)

# interpolating low frequency series
dta_mixed <- econ_us[, c(1, 3)]
predict(dfm(dta_mixed))
predict(dfm(dta_mixed, interpolate = TRUE))

## End(Not run)

srlanalytics/bdfm documentation built on Sept. 21, 2020, 10:45 p.m.