[ChadyWady]

That makes sense, I was confused since the Birthday problem is more about the pigeonhole principle rather than larger samples being more likely to have individually unlikely outcomes.

But for this case, it would still be very unlikely. For any arbitrary 3-year span, the probability of consecutively getting a "1 in 500 year" flood for 3 years is (1/500)^3. If we have N cities, then the probability of none of these cities having 3 consecutive floods is (1-(1/500)^3)^N. For even a vast overestimate such as N=20000, it is still a significantly improbable event. Of course, this doesn't account for "3 consecutive years or more within some year range" and it is a gross simplification, but I think there are probably better explanations than selection bias, such as the inaccuracy of the model or the fact that these events might be temporally dependent on each other.