Predictive Load Balancing prefering more busy servers than less busy servers over time

Have you seen this?

Observed (member): Specifies that the system ranks nodes based on the number of connections. Nodes that have a better balance of fewest connections receive a greater proportion of the connections. This method differs from Least Connections (member), in that the Least Connections method measures connections only at the moment of load balancing, while the Observed method tracks the number of Layer 4 connections to each node over time and creates a ratio for load balancing. This dynamic load balancing method works well in any environment, but may be particularly useful in environments where node performance varies significantly.

Predictive (member): Uses the ranking method used by the Observed (member) methods, except that the system analyzes the trend of the ranking over time, determining whether a node's performance is improving or declining. The nodes in the pool with better performance rankings that are currently improving, rather than declining, receive a higher proportion of the connections. This dynamic load balancing method works well in any environment.

Yes I have. But the help text doesn't explain how it is analyzing the trend of the ranking over time, and how I can (or can not) examine the actual trend.

Is there a stat/logfile on the server where I can examine the data its basing the trend stats on?