LessonThis is a lesson restated in new words. Any figure in it is an example from the source, not this desk's record.
A positive expected value is the drift of a wealth path, and it is not the whole story. If you add a fixed dollar win and loss, paths in the lecture stay near that drift. If you bet a fixed fraction of a growing bankroll, most paths go broke even though the average bet is favorable. The Kelly fraction is the stake that optimizes the typical path's growth. When the edge itself moves, the lecture prefers half of that fraction over the full one.
Why it matters
The lecture opens from a common saying that something like 90 percent of traders lose, then sets the saying aside for the mechanism. Edge is the expected trade: average winner times the chance of a win, plus average loser times the chance of a loss. You can try to raise the win probability or the average win, or cut the loss probability or the average loss. In a casino game the distribution stays put, so the edge stays put. Nothing the player does changes it. Poker and trading are incomplete-information games. The edge moves, and the actions you choose are part of why it moves. You never see the true expectation. You try to bet more when your proxy is higher and less, or nothing, when it is low or negative.
You also do not choose your path. The aim is a positive theoretical drift that the path you do get can track.
The idea in plain words
With $1,000, betting $10 a time is natural. With $1,000,000, a $10 bet is not. Stakes that scale with the bankroll are multiplicative. The lecture draws two positive-expectation strategies. Fixed $10 bets produce paths that hug the theoretical line. Proportional bets produce a fan in which only a few paths get rich and the rest go bankrupt. That split is the ergodicity point. On the additive side, the time average matches the ensemble average: one player's experience lines up with the group's. On the multiplicative side, the group's average is not what a typical player lives through. Positive expectation can coexist with almost every path dying.
Optimizing the time average of wealth produces the Kelly criterion. Replace the arbitrary fraction with that stake and the picture improves. The system is still not ergodic. More of the paths, in the lecture's comparison, finish ahead. In one draw, only two proportional paths end above their starting wealth. Under Kelly, significantly more do. You would rather be handed a random path from the second picture.
How to use it
Full Kelly assumes you know the edge. In the lecture's later draw, the edge flips between richer and poorer stretches, and you only notice after the fact, through a model or a history. Full Kelly is called too aggressive there. The simulated ensemble average of wealth ends below the starting stake. Half Kelly, half the optimal fraction, reduces the exposure in the stretches where the edge is weak. The lecture does not claim you can time those stretches. It claims a smaller fraction is how you avoid betting the full theoretical size on a proxy.
The practical sequence is to estimate the four pieces of the expectation, notice that a proportional stake changes the path even when those pieces are fixed, use a growth-optimal fraction rather than an arbitrary one, and shrink it because the estimate is stale. Additive betting avoids the ruin picture and stops making sense once the bankroll has changed scale, which is why the lecture treats multiplicative betting as the real problem.
What usually goes wrong
Quoting a positive expectation and then betting a large fixed fraction of wealth is the failure the side-by-side paths are drawn to show. Ensemble averages hide it, because a few enormous paths pull the group average up while typical paths hit zero.
Using full Kelly on a moving edge is the second failure. The lecture's full-Kelly cloud ends, on average, worse than the starting wealth, while the half-Kelly cloud is less punished when the edge dips. Neither picture is a promise that a formula produces gains. Both are simulations under an edge the drawer controls. Live, the edge is a proxy, the path is random, and the lecture's caution is the half-size stake.
Questions people ask
What is the difference between the two betting pictures?
Fixed $10 bets stay near the theoretical drift. A constant fraction of wealth, with the same positive expectation, leaves most paths broke and a few very large.
What does the Kelly stake change?
It is the fraction that optimizes the time average of wealth. In the lecture's draw, many more paths finish above the start than under the arbitrary fraction, where only two did.
Why half Kelly instead of the full fraction?
Because the edge moves and you do not observe it directly. Full Kelly in that simulation left the group's average below the starting stake. Half the fraction cuts the exposure.