LessonThis is a lesson restated in new words. Any figure in it is an example from the source, not this desk's record.
A trading edge, in this lecture, is the expected gain and loss of one bet. Multiply how often you win by the size of a typical win. Multiply how often you lose by the size of a typical loss. Add those products. You do not observe those four inputs. You estimate them from a history that will not repeat, and the inputs can change with time. A positive estimate can still produce a painful path. It is not evidence that the next sample will make money.
Why it matters
If the true probabilities behind a market were known, modeling would be unnecessary. They are not. The lecture starts with a coin that really has heads probability 0.5. Ten flips came up two heads and eight tails, so a person who saw only that series would estimate 0.2. The previous ten sessions of a market also happen once. You cannot rerun them to average the luck out. Confidence intervals and variance-reduction tricks narrow a statement about a fixed parameter. They do not create a parameter that stayed fixed.
The idea in plain words
Call the strategy S. Its expectation splits into a win piece and a loss piece. The split is more informative than the raw average, because it shows whether a high win rate is being cancelled by a large average loss. The lecture's worked sample uses an average win of 115, an average loss of minus 150, a win probability near 0.57, and a loss probability near 0.43. Multiply and add, and the unconditional average is about 1.42 per trade. Four wins and three losses are what those probabilities describe. The spoken list of raw gains is garbled, so the summary figures are the ones to keep. That 1.42 assumes the estimates are the truth and that they stay the truth. A second demonstration lets the coin's heads probability wander. In one window the estimate was about 0.7 while the probability in force was 0.46. Another window estimated 0.5 against 0.51. Another estimated 0.2 against 0.4.
How to use it
Compute the four inputs from a recent window of trades, live or backtested, and form the sum. Then simulate the estimated edge forward before treating it as a description of the future. With the 115 and minus 150 averages and the 0.57 win rate, one 252-day path in the lecture finished about 3,000 down and another about 750 up. Stretch the horizon to 2,000 steps, then 100,000, then a million, and the upward drift of a truly positive edge becomes easier to see. The short path can still ruin an account before that drift shows. That is a survival problem, not a contradiction of the mean. A slider in the lecture makes the same point without a long simulation. Drop the win probability and the edge goes negative. With the same 115 and 150 averages, a win rate near 80 percent showed an edge around 67. Keep the 80 percent win rate but cut the average win to 10 against an average loss of 150, and the slider read about minus 17. An 80 percent hit rate with a tiny winner is a losing rule. Shrink the average loss to 27 and the lecture's edge was about five per trade. The sign of the sum is the object to watch. Recompute it on a schedule you choose, such as weekly or monthly, because the inputs are not still. Backtests add their own damage: looking ahead of the signal, and fitting a curve to the same history you score.
What usually goes wrong
Treating the win probability as a constant 0.57 because one sample said so. The coin demo is the warning. A second mistake is ignoring costs. The lecture notes that transaction costs make a thin edge worse, and the 1.42 figure did not include them. A third is reading one equity curve. The same positive edge produced a large loss over 252 steps and a gain over another 252. The mean is a property of many repetitions, not of the path you happened to live through.
There is also no reason the estimated 20 percent from ten flips should be used as a win rate. The lecture reruns the equity curve under that bad estimate and the path is poor. Old data after a regime change, a new regulation, or the post-pandemic market can do the same thing quietly.
Questions people ask
Is a positive edge the same as a smooth equity curve?
No. The lecture's positive case still lost about 3,000 over one simulated year and made about 750 over another. Smoothness showed up more clearly at very long horizons, which a trading account may not survive.
What should be recomputed when the market changes?
The win probability, the loss probability, the average winner, and the average loser. Any one of them can move. The edge is the combination, so a stable win rate can hide a winner that has shrunk.
Does the lecture say these estimation problems are impossible?
It says they are unavoidable and that the literature has many partial answers. It does not offer a sample length that makes the estimates safe.