I was a bit surprised Wikipedia defined the Law of Averages to reference small rather than large data sets. "Averages" implies large data sets; I will not say Wikipedias usage is "wrong," but it is incredibly stupid and counterintuitive. Regardless of nomenclature though, the Law of Large Numbers is proven, and that is the concept I have invoked throughout this discussion. The more polls we have, and the more one candidate leads, the less likely the overall result is data noise.

Since February, Romney has been nationally polled against Obama 136 times: He led 23, tied 12 and trailed 101. Chance says it should be 50/50, so those results lean heavily toward Obama. As you note, polls are not random, but polls and elections are subject to nearly all the same deterministic factors (unless one believes all polls have the same internal bias, in which case there is no point discussing them.) In other words, the same factors that cause Obama to lead a disproportionate number of polls make his re-election correspondingly likely.

Obviously deterministic factors can change between a poll and election day, and Romneys decisive debate victory should alter what have always been close polls, just as subsequent events will. However, polling up until 3 October pointed to a narrow but nearly certain Obama victory.