Two very different things look identical in a file: a bad print from an exchange feed, and a genuine 20% gap on a takeover. Deleting the first is cleaning; deleting the second is deleting the risk you are trying to measure.
The distinguishing evidence is external. Check whether other venues show the same move, whether volume accompanied it, whether a corporate action or news event explains it, and whether the price reverted within a tick or two. A one-tick spike that reverts instantly on a single venue is a bad-tick.
A backtest whose result depends heavily on a handful of extreme bars is fragile regardless of which kind they are. Re-run with those bars neutralised: if the edge disappears, you have a story about three days, not a strategy.
Related: bad-tick, winsorising, data-quality-check, fat-tails