E-detectors in Volatility Monitoring

Volatility can jump due to market shocks or drift due to creeping changes, and an online monitor must detect both. Building on recent work on e-detectors and e-backtesting, we monitor a committed parametric (GARCH) forecast with a betting e-statistic whose validity requires only that the forecast bound the returns' conditional mean and variance, leaving the return distribution otherwise unspecified. In simulation experiments the monitor alarms two to eight weeks before the scheduled quarterly review that would otherwise first reveal the change. On S&P~500 returns (2006--2022 monitored, 2005 the burn-in year) it raises nineteen alarms that coincide with the recognized volatility events of the period, flagging the 2007 credit crunch thirteen days before the funding freeze that conventionally dates it. We compare two ways of choosing the bet: gridding the unknown post-change volatility multiple, and fitting the bet from recent evidence.