Somehow, some way, he is able to figure the one day samples. No one on this board knows how he figures it out, but I'd bet money his numbers are dead on.
+/- 0.3% actually (Rounding error) but close enough
Actually even at 1000 interviews per night, the amount of random "noise" is fairly high.
Each individual day will have a MOE of 3.1%, however since each candidate is subject to error, the actual error on the reported lead is 1.414 * 3.1 => +/- 4.4% on the "lead" In a tracking poll you also have two error inputs, because the random error of the sample dropping off is also a factor.
For example, if pro-Kerry "random error" drops off the 3 day sample, the same day some pro-Bush random error rolls one, you a have two blips operating together. Sometimes the errors cancel, but sometimes they add.
On average the daily difference in error (between the sample rolling on and off) will be +/- 6.2% 19/20 times.
At the 50% confidence level, this is about +/- 2.1%
Bottom line:
If the 3 day average moves 0.7% there is a 50/50 chance it is just pure random noise.
A 1% shift has a 1/3rd chance of being pure noise.
Once or twice a month, we will get a 1 day shift over 2% that is just nothing but random noise.
If you look at anything in Rasmussen, look at the 7 day average.