LessonThis is a lesson restated in new words. Any figure in it is an example from the source, not this desk's record.
Probability problems keep the same skeleton while the object gets richer: summarize data, count, condition, name a distribution, then estimate, then index the randomness by time. On five daily log returns the lecture’s arithmetic sums to 0.05, so the sample mean is 0.01, and the ordered sample makes the median and the mode 0.01 as well. At the far end, geometric Brownian motion is solved by taking a logarithm and applying Ito’s lemma, and a research-shaped question about the Hurst exponent of intraday volatility is left open.
Why it matters
Adding genuine randomness to a math problem is what makes the subject feel new. The technique is usually an old one with a probability symbol in front. Counting arrangements is a pipeline of four execution modules: 4 factorial orders, and a one-in-four chance that the lowest-latency module is placed first, because 3 factorial over 4 factorial is 1/4. Bayes’ theorem is a weighted update. A Poisson count of trading errors is a formula you can plug into a calculator. If those feel unrelated to a stochastic differential equation, it is because the notation hid the repetition.
The idea in plain words
A conditional-probability diagram in the lecture puts P(S) at 0.2, P(O given S) at 0.85, and P(O given the complement) at 0.3. Total probability makes P(O) equal 0.41. Bayes then puts P(S given O) near 0.4146, or 17/41.
A Poisson count with intensity 3 gives a probability near 0.224 of exactly two events. The density 3t squared on the unit interval has cumulative distribution 0 below 0, x cubed on the interval, and 1 above 1. Outcomes negative 1, 0, and 2, with probabilities 0.2, 0.5, and 0.3, have expectation 0.4. The second moment is 1.4, and subtracting the square of the mean leaves variance 1.24.
On the lecture’s triangular joint density, the mean of X is 1/3, the mean of Y is 2/3, and the mean of the product is 1/4, so the pair is dependent. Minus the log of a standard uniform draw is exponential with rate 1. For 100 independent exponentials with mean and variance 1, the sum has mean 100 and variance 100, and the chance it exceeds 110 is about 0.159 by a normal approximation.
The maximum-likelihood intensity of a Poisson sample is the sample mean. A 95 percent interval around a daily-return mean of 0.001, with standard deviation 0.02 and 250 observations, runs roughly from negative 0.0015 to positive 0.0035. In the frequentist reading, about 95 of 100 such intervals would cover the fixed parameter. It is not a 95 percent probability that the parameter sits in this one interval. A beta prior updated with K successes in N Bernoulli trials stays beta, with the success count added to the first parameter and the failure count added to the second.
An autoregression with coefficient 0.6 and noise variance 4 is stationary. Lag-1 autocorrelation is 0.6. Stationary variance is 6.25. One two-state chain in the lecture has long-run weights 3/5 and 2/5. A Brownian increment from time 1 to time 4 is normal with mean 0 and variance 3.
Ito’s formula turns geometric Brownian motion into an additive equation for the log. Exponentiating gives the usual lognormal price: initial price times an exponential of (mu minus half sigma squared) times t, plus sigma times the Brownian value. The lecture’s last question is open: estimate a Hurst exponent below one half from intraday log-volatility increments.
How to use it
Do the problems in that order, and do not skip the sentence under the number. A confidence interval has a reading. A covariance has a one-directional test. A density value at a point is not a probability until you integrate. When you reach Ito, write the partial derivatives before you substitute the price dynamics, or the algebra will look like a memorized final answer.
What usually goes wrong
Quoting a density as a probability. Reading a confidence interval as a posterior. Treating zero covariance as independence. Forgetting the piecewise cumulative distribution outside the support and losing the exam points the lecture warns about. And applying Ito to the price when the workable equation is the log.
Questions people ask
What is the maximum-likelihood Poisson intensity?
The sample mean. Differentiate the log likelihood, set the derivative to zero, and the intensity estimate is the average count.
When are two continuous variables not independent?
When the expectation of the product differs from the product of expectations. In the lecture’s example those were 1/4 versus (1/3) times (2/3).
What does Ito buy you on a geometric price?
A translation into additive noise for the logarithm, and then the explicit lognormal formula for the price after you exponentiate.