The probability that a roll of your die will be a 6 is 1/6 th, REGARDLESS of any past series of rolls. That is what random MEANS, that there is no correlation from try to try, the past does not predict the future, AT ALL.
It is precisely that which justifies these metapoll studies, that the polls are each separate tries, and hence (by their randomness) independent as sample data. (This is quite questionable too. Pollsters are poll junkies.)
The technical usage has a technical definition, which is correct. That does not mean the choice of the word "confidence" has the correct nontechnical connotations, and as I said, in fact it's wrong, 95% probability is not 95% confidence about any try in particular, each try is ASSUMED independent, nonrepeatable, and without replacement, "confidence" suggests "assurance", and that misleads, you have no such thing, you are still gambling. The entire system is built on the premise that each try is independent, that every time you ask the question, it is new.
A 95% confidence interval is a range in which the sample is expected to fall 95% of the time, the other 5% of the time it is expected to fall outside, and you are just as "confident" about the accuracy of the 5% as about the 95%, no more and no less.
At best, what polls provide is a crude heuristic that (done well) a sophisticated observer can interpret with some "confidence". Note that "confidence" is something someone has, an opinion, something one could sample too. I'm pretty "confident" at this point that Barak Obama will be the next President, barring "accidents". But I also know the present does not compel the future, it's just an educated guess based on a structured investigation.
I can go on. This is just scratching the surface, the easy stuff, we can get into Mr. Taleb's bootstrap issues with distributions, which are obvious to anyone with mathematical training and the will to look, for example. They annoyed me when I studied the subject 30 years ago, and they still do. There is nothing normal about the "normal distribution". The real world does not in fact exhibit zeroes or infinities or continuity, it is discrete, quantized, finite, and always something. Even "empty space" is far from empty. Velocity is relative (no absolute rest) and finite (less than the speed of light relative to any particular inertial frame). The uses of such notions in the physical world are all based in the laws of large numbers, probability and combinatorics, and all the physical constants that we know with great precision likewise.
Most of the more responsible pollsters seem to know that some of these issues exist, their rhetoric shows, but all the financial incentives run the other way, toward claims of certainty or some near approximation, or to show some desired "trend", a notion which you will see wanders off completely into the brush of cause and effect thinking.
There is a fundamental contradiction between assuming in one place that people form random populations you can sample randomly, by language no less, and in another, that people's behavior exhibits serial correlation (cause and effect, "confidence" about the future) over time, too. This I have seen mentioned, but I've never seen it really addressed at all.
People want confidence, assurance, security if you will, but that is wishing, we have no such thing and we need to pay attention, not be running around fat, dumb, happy, and full of ourselves because we think we know what's coming next. We don't.
I hope this at least clarifies what I base my views on.
