have a 94.8% chance to make the playoffs with no margin of error is not correct. The basic premise that hollinger works off of is that the strength of each team can be quantified by a single number, or rating, and that the probability that a given team will win a single game is determined by the ratings of the two teams.
 In other words, for any two rankings, the scoring margins and other statistics should have a distribution. i.e. if one team is much better than the other, then we would expect the distribution to favor blowouts by the better team more than say close games. If the two teams were equal, we would expect a symmetric distribution. The simplest way to model this in basketball terms would be with a normal distribution around the expected margin, and a standard deviation based on historical data.
There is no problem with this theory. In fact, I rather like it. However, what Hollinger, and anyone else, must try to do is get estimates on what a teams “true” rating is based on the observed results of a limited number of trials, the games. As such there is an associated error between the best estimate of the rating and the actual value. For a simple example, think of polling. This is an attempt to determine the true percentage of voters on either side of an issue or candidate by taking a limited number of samples. Such polls are ALWAYS accompanied by margins of error (which is the inherent mathematical error, and doesn’t account for sampling errors, question bias or other methodological errors). In polling, and most other disciplines, prognosticators are obligated to provide some idea of how accurate their estimations are.

The problem with the playoff predictor is that it presumes that the Hollinger Rankings accurately reflect teams relative strengths, and thus give an effective prediction of the likelihood that, say, the Blazers beat the Sixers tonight. If the probabilities could truly be calculated, then the playoff predictor would be accurate and no confidence interval would be necessary. However, we don’t know exactly how likely it is that the Blazers beat the Sixers, although we could, using past data, probably show that it is likely better than 50%. However, any calculations based on this percentage would also be accompanied by some variance, which should be calculated and shared.

The idea of an error on a likelihood is a little unintuative, and it might help to think of it in terms of polling again. The percentages given, such as 53% for Obama, are really the likelihood that any one person will vote for obama. However, we don’t know that this is really the right precentage, so we include a 95% confidence interval of +/-5%, to give an idea of how accurate this prediction should be. There is a difference between polling 2000 people to get 50% with a +/-2% and 20 people to get 50% with a +/-20%. The second poll wouldn’t be given too much credence. I think that any honest calculation of variance for Hollinger’s playoff odds would look more like the 20 person sample. This is not a knock on his mathematical or statistical abilities, simply that the game of basketball is highly variable and should reflect that.