LessonThis is a lesson restated in new words. Any figure in it is an example from the source, not this desk's record.
The law of large numbers says a sample average gets arbitrarily close to the population mean, in probability, as the sample size grows. The central limit theorem says something stranger: the histogram of those averages turns bell-shaped even when each sample came from a flat, Poisson, or already-bell-shaped source. Together they are why a partial scrape of posts can still support a normal-based test.
Why it matters
A standard uniform draw is equally likely anywhere between 0 and 1, and any single point in that interval has probability zero. Ten thousand of them, in 20 histogram bins, look like a flat bar. A standard normal draw, mean 0 and standard deviation 1, is not limited to that interval. Ten thousand of them look like a bell, and about 68 percent fall between negative 1 and 1. A Poisson draw is a different shape again. The surprise is not those pictures. The surprise is what happens to the average of each picture when you repeat the averaging.
Without that surprise, a trading rule that can only read a sample of posts, rather than every post in a day, has no sampling distribution to test. The lecture’s motive is exactly that constraint: a cap on how much text an API will return, so the day’s sentiment is a sample mean instead of a census.
The idea in plain words
The sample average is the ordinary mean: add the draws and divide by how many there are. For 10,000 standard normal draws the average sits near zero, and a fresh 10,000 sits near zero again, not on the same digits. The formal line is a limit. As the sample size goes to infinity, the probability that the sample mean is more than some positive distance epsilon away from the true mean goes to zero. Epsilon is an arbitrary gap you pick in advance. The lecture picks 0.01.
You cannot simulate infinity, so they simulate the probability. One thousand times, draw a normal sample, average it, and record a 1 if the absolute average is under 0.01, else a 0. The share of 1s is the estimated probability. At 10 draws that share was 0.021. At 100 draws it was 0.091. At 50,000 draws it was 0.98. At 100,000 it moved closer to 1. Same epsilon, larger samples, the average is more often inside the gap. That is the law they are pointing at.
The central limit theorem looks at a different collection: not the draws, but the averages. Build 1,000 averages, each the mean of 10,000 standard uniforms. The histogram of those 1,000 averages is bell-shaped, not flat. Repeat the recipe with normal draws and the averages are bell-shaped too, which is less shocking. Repeat it with Poisson draws and the averages are still bell-shaped. The lecture’s claim is that the parent distribution is not what the averages inherit, once each sample is large. A realized average is a known number after you compute it. The theorem is about the distribution you would see if you computed it over and over.
How to use it
Use the epsilon experiment when you need to believe the law rather than recite it. Fix epsilon, fix the number of repetitions, and only then increase the inner sample size. If the share of “close enough” averages does not rise, the inner sample is not growing or you are subtracting the wrong true mean. For a standard normal the true mean is zero, so the absolute average is already the distance.
What usually goes wrong
Quoting the 0.021 or the 0.98 as constants ignores the design. Those shares came from 1,000 trials at a specific epsilon of 0.01. A different epsilon moves them. The pattern to keep is the rise toward 1 as the inner sample grows.
Questions people ask
What did the epsilon check print?
With epsilon at 0.01 and 1,000 trials, the share of standard-normal averages inside that gap was 0.021 at 10 draws, 0.091 at 100, and 0.98 at 50,000. At 100,000 the share was closer to 1.
Do Poisson averages really look normal?
In the lecture’s histogram, yes. So did averages of uniforms and averages of normals. The parent shape was flat, bell, or Poisson. The shape of the averages was the bell.
How does this apply to a partial day of posts?
The day’s score is a sample average. The theorem says that average’s sampling distribution is approximately normal, so a test of sign or of an extreme bucket has a distribution to refer to. A very small or very noisy sample can still be a bad bucket.