I did a test implementation. The following problems exists:
We don't count successful messages. This is hard since a message can be dropped anywhere else much later in the call chain. We count the total messages and the dropped. The good messages would be =total-dropped.

We drop received and sent queue messages. Both works quite differently. Received messages dropped are mostly because of parsing problems, spam or unsupported messages. They come once in a while between good messages.
Sent messages are most of the time dropped in large chunks. This is because of the send queue and message priority. First important messages are transferred, low priority messages are queued for the next rotation. This continues until the live time of queued messages is over, then they are dropped.
This might cause large chunks of messages (lets say 50-200) be dropped at once. And will immediately let your algorithm drop down below the threshold, even though its only a short few seconds lasting low bandwidth period.

__________________