A tick arrives at 1,043 for a stock whose last 200 prints sit between 98 and 102. Give a rule that flags a bad print without flagging a genuine 8 percent gap on news, say what you publish for the flagged tick, and state the cost of the rule when it is wrong.

A tick arrives at 1,043 for a stock whose last 200 prints sit between 98 and 102. Give a rule that flags a bad print without flagging a genuine 8 percent gap on news, say what you publish for the flagged tick, and state the cost of the rule when it is wrong.

Approach: Use the fact that an erroneous print reverts on the following prints while a real move does not, then decide what the pipeline publishes during the interval when the answer is still unknown.

Flag a print whose move exceeds k times a robust dispersion estimate and whose following prints revert most of the way back, and publish the flagged tick to a quarantine stream rather than into the price series until the next few prints settle the question. A size test alone cannot separate the two cases, because an 8 percent news gap and a fat finger both look large. Use the median absolute deviation of recent returns rather than the standard deviation, since one bad print inflates the standard deviation enough to hide itself, and start with k around 10. The discriminator is the reversion test: an erroneous print is followed by prints back at the old level while a genuine move is followed by prints near the new one, so hold judgement for a small number of ticks or a few hundred milliseconds. Cross checks help, a quote that never moved, an odd venue, or a condition code marking the trade out of sequence. The false rejection cost is asymmetric: discarding a real move leaves a stale price feeding every model downstream, which harms a signal more than one outlier would, so the rule should delay and quarantine rather than delete, and every rejected tick should be written where a person can review it.

Follow-up: How do you set k when the same rule runs across equities, futures and crypto in one pipeline?

Key concepts: reversion test, robust dispersion estimate, quarantine, false rejection cost.