"is plotted with official exchange timestamps" suggests that they are doing it wrong. You can't compare apples to oranges without knowing how the exchanges are timed.

For those who do latency tests, this is a very important point: you should always be on the lookout for what clock is recording the 'start' and the 'stop' and to be sure to consider clock skew.

To get a sense for how far timestamps can diverge, OATS -- the reports that are sent to the Financial Industry Regulatory Authority -- require that machines be synced to within 3 seconds of NIST (which is nearly 7.5x longer than the 400ms quoted).