I've been following the discussion about this tracking poll for a bit and so far I've mostly followed the 538 opinion. This poll is useful as long as you use it to track trends, not as an accurate indicator of the race. However, today inspired by "the formula" I've been thinking about what the daily fluctuations of this poll can tell us. Looking at these, it appears that this poll is very elastic, likely exaggerating swings in both directions. (Apologies if anyone remarked on this previously, I didn't read through the complete thread yet.)
Let's take a look at today's numbers. Today the results are C:42.3%, T:46.7%. Yesterday it was C:41.7%, T:47.3%. Assuming that the "daily cohorts" are all of an equal size, we see the following:
x_{t-7} / 7 + (6/7)y = 41.7%
x_t / 7 + (6/7)y = 42.3%
with x_t the % Clinton supporters in the Saturday cohort today, x_{t-7} the % Clinton supports in the Saturday cohort last week and y the % Clinton supports in the other cohorts polled over the previous 6 days.
Solving this (and doing the same for Trump) we see that among the Saturday cohort, there has been a massive 8.4 percentage point swing for Clinton over the last week. Doing the same for the Friday cohort, there has been an even larger 11.2 percentage point swing for Trump. In fact, during September there have been 7 swings of over 10 points (including one 23.1 points swing) among daily cohorts in a weeks time. Even though the daily cohorts are fairly small and have high MOE, such regular large swings should probably be much rarer.
The methodology of this poll can explain the elasticity. Remember that this poll weights based on self-reported likelihood to vote. Good or bad news for a candidate thus doesn't only influence the amount of supporters, but through increased enthusiasm further amplifies swings.
Great analysis! From my non-mathematical lay observations I've noticed that the poll seems to lag about a week behind the consensus of other pollsters/trackers. That said, I've been getting more skeptical of the underlying methodology (self-reporting, only using 2012 voters).
Also, for those of us who also don't have the time to read through the entire threat, could someone (re)explain the 'formula' and how it comes to its parallel numbers each day?