ENH: Andrews (1991) AR(1) plug-in bandwidth for the kernel covariance estimators #887

The kernel estimators in arch.covariance choose their bandwidth by the
Newey and West (1994) procedure, whose pilot lag n = ceil(4 (T/100)^rate)
sets how many autocovariances enter s0 and s(q). The 4 is the value
Newey and West used in their Monte Carlo (Table II), so the bandwidth of a
series depends on that choice.

Andrews (1991, Section 6) gives a plug-in bandwidth with no pilot lag. An
AR(1) fitted to each column gives (rho, sigma^2), and

alpha(1) = sum_a w_a 4 rho_a^2 sigma_a^4 / ((1 - rho_a)^6 (1 + rho_a)^2)
           / sum_a w_a sigma_a^4 / (1 - rho_a)^4
alpha(2) = sum_a w_a 4 rho_a^2 sigma_a^4 / (1 - rho_a)^8
           / sum_a w_a sigma_a^4 / (1 - rho_a)^4

(6.3)-(6.4), with the bandwidth 1.1447 (alpha(1) T)^(1/3) for the Bartlett
kernel, 2.6614 (alpha(2) T)^(1/5) for Parzen and 1.3221 (alpha(2) T)^(1/5)
for the quadratic spectral kernel ((5.3), (6.2)). The kernel constants are
derived in the paper, and everything else is estimated from the data. R's
sandwich::bwAndrews and cointReg::getBandwidthAnd compute it.

Would an option to choose the bandwidth this way be welcome, such as a
bandwidth_method argument on the kernel estimators taking "newey-west"
(the current default) or "andrews"? PhillipsPerron could then use it
for its long-run variance, in place of the 12 (T/100)^(1/4) default lags,
since it already computes the regression residuals the AR(1) would be
fitted to. I can open a pull request if so.