spb/wp46_uqo

1Four decades after ARCH, does anything beat GARCH(1,1)? We answer with a pre-registered benchmark of the conditional2volatility family on 42 instruments---ETFs, stocks, futures, exchange rates and cryptocurrencies---with one-minute and3daily data from 2000 to 2026. Sixty-four architectures are inventoried in a definition-faithful genealogy; 29 are scored4with a frozen protocol (rolling 1{,}000-day windows, weekly re-estimation, horizons of 1, 5 and 22 days, QLIKE against a5full-day realized proxy, Diebold--Mariano, model confidence set, superior-predictive-ability and joint block-bootstrap6inference, VaR/ES backtests, pre-defined regimes, failure and compute accounting), eight more as labelled extensions (FIAPARCH and Realized EGARCH among them, the latter the best one-day forecaster of the study). Within7the daily information set, nothing beats GARCH(1,1) by a material margin at one day: the best architectures (FIGARCH,8component GARCH) gain two to three percent, asymmetric models gain on equities only and nothing on average, regime9switching gains nothing at any horizon, and unconstrained EGARCH produces explosive forecasts on a third of the instruments10from well-fitting parameters. One day of intraday information is worth twelve to fifteen percent of QLIKE at one day for11every realized-measure model except the levels HAR---four times the largest daily-tier gain---and the advantage fades by one12month except for log-HAR. Innovation laws move variance forecasts by one percent but decide tail calibration: only the skewed13Student-$t$ passes 1\% and 5\% VaR backtests on most instruments. Rankings are stable across trailing-volatility regimes; what14reorders them is the speed of a variance jump, which only intraday data measure in time. Estimation failures are rare; the15operational risks are inadmissible fat-tailed fits and explosive out-of-sample dynamics. The replication package ships the16taxonomy, code, tests, manifests and aggregate forecasts.17