Here is an attempt at an explanation.
Imagine we are in track and there is someone who ran under 13 minutes last year. This year the guy has only been in smaller races, maybe he had a race or two that was bad. He has been running 13:30s, but hes been winning them. Then he runs a 12:50. Together with the sub 13 last year and this year, its safe to say hes been in sub 13 shape this whole time, so we use the information from last years PR instead of his tempo races from this year.
Now imagine we have another guy who ran 13:00 last year and has been running 14:00s all of this year without a single good race. We should count both the 13 and the 14s to represent the fact that he is slower now, but could potentially get faster.
Third case: imagine someone who ran 14:00 last year, but now has run 3 races at 13:30, 13:20, 14:10. What fitness would you say that guy has? Id say around 13:25 because the 14:10 was probably a workout. In this case it makes sense to ignore last seasons 5k.
The weighting is dependent on what races are most likely to predict someones fitness. That turns out to be the group of fastest times, but only if at least one run from that group is recent.