LessonThis is a lesson restated in new words. Any figure in it is an example from the source, not this desk's record.
A sample average has its own distribution. When every draw comes from one fixed source and the sample is large, that distribution is a normal, or Gaussian, bell curve, even if each raw observation is not. Equity returns do not stay in one source, so a bell curve fit to old block averages can treat a common loss as almost impossible.
Why it matters
Market questions are often questions about an average: a window of returns, or a sentiment score. Those averages are random. Another sample from the same source would have produced a different number, so you need the distribution of the average before you call a reading ordinary.
A Poisson count can stand for trades in a 100-millisecond window with a mean of four. One simulation records counts across 25,000 milliseconds until the histogram nears the theoretical mass function. Repeated averages landed near 4, then 3.4, 3.2, and 4.2. They wander, then settle. The law of large numbers says that if every draw shares one unchanging source, averages and empirical frequencies settle on the population values.
Markets drop that premise. Fit the same Poisson model on a 10,000-millisecond window and the mean can be four, while the next window averages two. The useful question is narrower. If the source really had mean four, how surprising is a sample mean of two?
The idea in plain words
Poisson assigns a probability to each count. In the lecture's picture, about a tenth of the mass sits on six trades in one window. A continuous density does not work that way. Height at a point is a relative likelihood, and probabilities are areas. The standard normal, mean zero and variance one, puts half its area left of zero and roughly two percent left of negative two. Moving the mean or the variance moves those areas, for every normal curve.
The central limit theorem names the law of the standardized sample mean. Standardization subtracts the population mean and divides by a standard deviation that has been rescaled with the square root of the sample size. A sketch uses characteristic functions, the transforms that identify a distribution. Independence turns a sum into a product of those transforms. A Taylor step and a large-sample limit leave the standard-normal transform. Lévy's continuity result turns that into convergence of the distribution itself. The parent law can be Poisson, Bernoulli, or binomial. The mean's law is still normal, centered on the population mean, with spread that shrinks as the sample grows.
More draws from a fixed source pull the observed average toward the population mean. Center a normal on that average, the placement that makes it as likely as possible, and read probabilities off areas. Confidence intervals and tests use the same bridge.
How to use it
Use the curve when the question is about today's average and the window is short enough to treat the source as roughly steady. A sentiment check from 8:00 to 9:20 in the morning has only 50 documents. A cross-sectional book would buy high scores and sell low scores. Before a name joins the long bucket, the lecture asks, under a conservative null curve, for the chance of a sentiment mean at least this large. That area was under one percent, so the name was dropped.
The same curve was fit to disjoint 90-day mean returns for Nvidia. The average of the block means was 0.21 percent, with a standard deviation of 0.049 percent. A normal centered there put essentially no mass on a negative block mean, which the lecture turned into a wait on the order of 42,290 years.
What usually goes wrong
The blocks did not wait. Of 74, twenty were negative, about 27 percent. The figure for the negatives already seen was a wait of 845,915 years. One normal, built from every block mean, treated a moving source like a Poisson experiment that never changes.
The theorem still describes the sample mean while the source is steady. After that, the mean, the variance, and the probabilities move, and the model has to be rebuilt. A mean from three months ago, reused as today's law, is a weaker use than asking whether a small sample today looks extreme. Curves centered far away make the observed average almost impossible, so plausible ones sit nearby and still drift.
Questions people ask
Does each daily return have to be normal?
The theorem is about the average. Daily returns can stay skewed and heavy-tailed while averages of many draws from one fixed source look bell-shaped.
Why did the Nvidia curve look so safe?
It used one curve for every 90-day block. Twenty of 74 blocks had a negative mean return, so the curve was far too tight for a source that moves.
Can 50 documents support a signal?
They can support a filter. Under the lecture's conservative null, the chance of the observed sentiment mean, or anything higher, was under one percent, and the name stayed out of the long bucket.