No matter what reward method is used, when a pool is down to zero buffer on a long round, it must either underpay or create debt for the pool owner. That is an inevitable fact of pooled mining. Nothing can stop it. *MPPS attempts to mitigate it by offering to make up for the loss later.
Your mistake is assuming that the goal of a pool is to reward each share with the statistical average, exactly.
The real goal of a pool is, however, to reward each share with the statistical average, on average (while reducing variance as much as possible). On the one hand you have PPS, which is 0 variance for the miner (with maximal risk to the operator); on the other hand, solo mining which has the highest variance; and a variety of methods in between.
PS. If MPPS is what I think it is, then it is just as bad as the *SMPPS.