LessonThis is a lesson restated in new words. Any figure in it is an example from the source, not this desk's record.
A probability model writes a random experiment in symbols: which outcomes you allow, and what weights the events carry. It answers which results can occur, how likely an event is, how one event changes another, and what the expected value is. It does not tell you the next roll. If your best price sits on the right side of that expected value, a long run of bets can have a positive average even though each bet is uncertain. If a disaster is missing from the list of outcomes, the model will price it as impossible until it happens.
Why it matters
Introductory classes stop at “a fair coin is one half” and never say what the one half is for. The lecture’s die example is the missing paragraph. Each face has probability one sixth, so the expected value is 3.5. No roll equals 3.5. A market maker who buys below that value and sells above it still has a positive expectation. A simulated quote that buys at 2.5 and sells at 4.5 accumulates faster than a tighter quote around the same center, because the width captured on each fill is larger. Shrink the width and the wealth path flattens. That is the business use of the model: a stable expectation and a spread, not a prediction of the face.
The idea in plain words
A random experiment is a repeatable procedure you cannot call in advance: a coin, a die, a roulette wheel. The sample space is the set of outcomes. Sets ignore order and duplicates, so a die is just the six faces. Two dice produce 36 ordered pairs.
An event is a subset, such as the even faces. Union means either, intersection means both, and the complement is everything else in the sample space. Add probabilities only after subtracting the overlap, or the overlap is counted twice. Disjoint events have no overlap. The complement’s probability is one minus the event’s probability.
Three axioms constrain a probability measure. Weights are not negative. The whole sample space has probability one. A union of disjoint events has probability equal to the sum of the pieces. Anything you build from those rules lands between zero and one.
When every atom is equally likely, probability is a count. Three even faces out of six is one half. At least one six with two dice is 11 favorable pairs out of 36. The complement is cleaner: the chance of no six is 25/36, so at least one six is 11/36. Three ordered slots from six choices, without putting a choice back, number 120. If order is irrelevant, divide by the 6 ways to rearrange three items and 20 groups remain.
Conditional probability shrinks the denominator to what you now know. A partition cuts the sample space into disjoint slices that cover every outcome, and total probability weights those slices. Bayes’ theorem runs the conditioning backward. Volatility regimes are one such cut.
The list of outcomes is part of the model. Early option pricing already used Brownian motion. If a mass default is absent from the list, the model treats it as impossible until the loss arrives. That hole is the risk.
How to use it
Write the sample space before the formula. If you cannot list the outcomes, or honestly describe them, you are not ready to average them. Use complements when the direct count is messy. Use a partition when the unconditional probability is awkward and the slices are natural. Check disjointness before you add probabilities without a subtraction.
What usually goes wrong
Predicting the next outcome and calling that a probability course. Double-counting overlaps. Confusing a set union, which drops duplicates, with a probability sum, which must subtract them. Building regime buckets that leave a gap or overlap, so they are not a partition. And shipping a risk system whose sample space has no room for the loss that actually ruins the book.
Questions people ask
What is the expected value of a fair die, and why is it useful?
It is 3.5. You will never roll 3.5. You can still buy below it and sell above it and have a positive average result over many rolls.
How do you get 11/36 for at least one six?
Either count 11 pairs in a 36-pair space, or compute the chance of no six, 25/36, and subtract from 1.
What does a model say about an outcome it omitted?
Inside the model, that outcome has probability zero. The lecture’s credit example is the warning: the omission is the risk, not a theorem that the outcome cannot occur.