With today's projection, 270towin claims Obama-Biden has a win probability of >99%, 29 days out from the election¹ and with 115 EVs labelled toss-up. They do so by using poll² numbers³, then take the poll's margin of sampling error⁴ as a sample standard deviation⁵ to find the joint probability (confidence interval) -- the probability that for polls showing 48%-44% the candidate with 48 is actually ahead.

Then they take those numbers for each state and effectively MULTIPLY THEM TOGETHER⁶. Maybe you won't read the footnote, so I repeat: THIS IS COMPLETELY UNJUSTIFIABLE AND DOESN'T EVEN PASS THE LAUGH TEST. This "multiply the probabilities" only works for independent outcomes. If I roll two dice at an honest craps table, a 6 on one die says nothing about the result of the other; there are 6x6 = 36 equally-likely outcomes. The fuzzy dice on my car's rear-view mirror have a string connecting two corners that will never land opposite each other: there is a strong dependency of one die's outcome on the other's and I'd like to meet the chump who will play me at craps with them.

So: let's say things go well for Obama in most of the country: his ground game edge yields FL, NH, MN, CO, NV. Hooray! That's 262 votes, and with any single one of the remaining states it's bicycle time! But what if McCain stops slinging mud long enough to find some issue (or finds just the right flavor of mud) to win voters in the Rust Belt and mid-Atlantic? If any of those states flip there's a strong dependent possibility they all flip. If the above flawed analysis gives about a 75% chance of Obama victory in each of IN OH MI PA VA NC, then the probability of a McCain victory is not .25*.25*.25*.25*.25*.25=99.98%, it's more like 25%. And that's before you include the other 900 mDFW⁷ of neglected uncertainty.

============================================

Now everything I've described above is rigorously handled by exactly and only one source, Nate Silver at fivethirtyeight.com. But even those neglect the most crucial source of uncertainty: Model Risk⁸. Silver's, and all others', use the average poll behavior of every past election to project the range of outcomes for this one.

That's fair, of course, because most past elections have been like the other past elections, and so will future elections, until one election is completely and unexpectedly different than any other election. In no past election has Osama bin Laden been captured⁹, or GM gone bankrupt¹⁰, or Homeless Yuppies riots broken out in San Jose¹¹. By definition, there's no way to enumerate these unforseeable events, let alone estimate their probabilities, and so no model includes them.

When unforseen stuff like that does happen, we have an ingrained positivist reaction to say "Well, of course our estimates were wrong, but nobody could forsee that happening". And we say that every single time something unforseeable happens, refusing to forsee that the unforseeable is clearly forseeable; it just isn't measurable¹².

If you'd like a picaresque and mostly-non-technical tour of Model Risk, I highly recommend the book "The Black Swan" by Nassim Taleb; or to at least, the next time you're in a book store, to pull it off the shelf and read the chapter about the hero who prevented the 9/11 attacks. Also enjoy the recurring predictions of an economic meltdown caused by correlated risk in the global financial markets and the dangers of a housing bubble¹⁴.

============================================

What I'm trying to say is, quit hitting refresh on TPM or writing overlong exegeses on statistical uncertainty and make some calls during your lunch break.

============================================

1. The election is not being held today: past elections have exhibited clear trends approaching election day.
2. Do you have a land line? I don't have a land line. People without land lines have clear demographic similarities.
3. Poll numbers have to be adjusted by both demographic and likely voter proportions. Can you estimate within 10% the voter turnout in a historically followed election, in which record numbers from the young and black voter segments -- two groups with historically low turnout -- are expected to turn out in droves? Can you then adjust that figure for Voter Caging, Forward and Reverse Bradley effects, reported reticence by people of color to reveal their race to pollsters, and the tendency of young voters to underreport on election day?
4. This is not the "margin of error", it's one particular kind of error calles the "sampling error", which gives only the uncertainty from using a census of say 3,000 people to estimate the opinion of say 500,000 people.
5. Non-mathies -- when you head 'standard deviation' please read as "Fuzziness in Result, but Defined Rigorously. Roughly speaking, repeated measurements will be +/- one Fuzziness two-thirds of the time, and +/- two Fuzzinesses 95% of the time.
6. THIS IS COMPLETELY UNJUSTIFIABLE AND DOESN'T EVEN PASS THE LAUGH TEST.
7. a mDFW is one-thousanth of a Davidfosterwallace, the standard measurement for footnote density.
8. A topic introduced on the last class day with devastating clarity by my Computational Finance professor. A mathy way of seeing it: as you introduce more and more dimensions to produce an ever more sophisticated model, mapping "what I can measure" onto "input variables" compresses and stretches the phase space of solutions. A neglected variables or hidden dependency gives a near-degenerate (and thus massive) eigenvector, and so a tiny displacement can have an enormous and unaccounted-for effect.
   
   A non-mathy way of seeing it: the professor's takeaway message was "Everything you just learned in this class is mostly right, the balance of 'mostly' being where it is inexorably and fundamentally wrong."
9. Praise be to Allah
10. God forbid
11. Granfallon be to Bokonon
12. Though much maligned for it, this forms the thesis for one of philosopher-poet D. Rumsfeld's classic perorations: "There are known knowns. There are things we know that we know. There are known unknowns. That is to say, there are things that we now know we don’t know. But there are also unknown unknowns. There are things we do not know we don’t know." I'm not being sarcastic: get past the dadaism and this is a really deep concept correctly put.
13. In a rant against positivism I'm allowed to be superstitious
14. Model risk is the one and only root cause of the current Global Financial Meltdown; the rest is commentary.
15. BTW: To lose at Risk, try for Asia but fail to secure Kamchatka, or write off your opponent as lightweight because she's a model.