I generated ten random series of a thousand samples each with a simple Python script, and here's the result:
Apologies for the horrible colors. Notice how one of the lines is slightly below the 99.7% confidence interval for a while, before eventually climbing back up? Exactly like you are seeing in your experiment.
While 99.7% may seem like an impressive number, it only covers 997 of every 1000 cases. The remaining three fall outside it. And there are many thousands of players on each server; maybe even tens of thousands. So it's far from impossible for someone to have such a streak of bad luck.