Within the daily information set, the answer to the title question is ``nothing material at one day, a little at one
month''. The best pre-registered daily architecture, FIGARCH with GED innovations, has a pooled QLIKE of
\RatioFigarchOneOneGedHOne{} times that of GARCH(1,1), and the component model \RatioCgarchOneOneTHOne; every asymmetric,
power, in-mean, higher-order or regime-switching variant is within the two-percent band that the protocol defines as
immaterial, or worse. At twenty-two days the absolute-value and threshold models of Zakoïan and Taylor--Schwert gain five
percent, and GARCH(1,1) with fat-tailed innovations two percent. Asymmetry pays on equities (3--7\%) and costs on
futures and FX; fat tails and long memory pay on exchange rates (8--11\%) and nowhere else. Unconstrained EGARCH, which
fits as well as any model in sample, returns parameters with a negative magnitude coefficient and near-unit persistence on
a third of the instruments and then produces explosive forecasts: its pre-registered mean loss is six orders of magnitude
above GARCH(1,1)'s, a result we report as is and diagnose. Intraday information changes the picture: the previous day's
realized variance, used through a logarithmic HAR regression, a GARCH-X, a HEAVY or a Realized GARCH model, lowers the
one-day loss by 12--15\% on average, significantly on 25--31 of 42 instruments, with the largest gains on futures and FX
and none on cryptocurrencies; the gain fades by one month except for log-HAR. Tail risk is governed by the innovation law,
not the architecture; the skewed Student-$t$ is the only law that calibrates both VaR levels. Dynamic correlation models
beat constant correlations by 15\% on an equity basket and 3\% on a cross-asset basket. Rankings do not move across
trailing-volatility terciles; they move during the COVID jump, when realized-measure models gain 15--29\% and
asymmetric models lose. Estimation failures are rare after a multi-start policy, but 4--7\% of fat-tailed GARCH fits are
inadmissible and two models (EGARCH, levels HAR) produce explosive forecasts; computing cost is immaterial except for
fractional models. The eight post-registration extensions do not change these conclusions: FIAPARCH is the only daily architecture to improve on every parent (six percent at one month), Realized EGARCH is the best one-day forecaster of the study (15\% below GARCH(1,1)), and the exponential family's explosions are cured by a bounded score or a sign constraint at no gain in accuracy.