LessonThis is a lesson restated in new words. Any figure in it is an example from the source, not this desk's record.
Stop trading a rule when its edge is no longer positive, and watch the pieces of that edge rather than waiting for the equity curve to make the point after the fact. In the lecture's base case the win probability is 0.55, the average win is 10, and the average loss is 8. That combination has a positive expectation, and simulated equity curves over 1,000 trades drift upward. Let the win probability slide from 0.55 to 0.4 across those 1,000 trades and the curve flattens. The payoffs never changed. The frequency of wins did.
Why it matters
Systematic stock rules die for ordinary reasons: a new macro regime, a different rate environment, or other traders learning the same signal. Sentiment-style rules in the lecture worked and then leveled off around 2021 and 2022. The researchers' answer was not a single patch. The regime for that signal had changed, and a different sentiment measure might work, or none might. You can see the damage in the win rate and the payoff sizes before you have a story for it. Those quantities are themselves random and can be nonstationary, so a fresh estimate is not a population constant.
The idea in plain words
Edge here is win probability times average gain, minus loss probability times average loss. Positive means the average trade earns. Negative means it loses. With a win probability of 0.55, a win of 10, and a loss of 8, the edge is positive, and repeated 1,000-trade paths rise on average even though each path differs. Leverage multiplies the whole curve. It does not create a better risk-reward ratio. The lecture deliberately skips Sharpe and Sortino ratios. The question is statistical: are the inputs that feed the edge still the inputs you estimated? A fixed stop and a fixed target make the average win and average loss constants, so only the probabilities have to be tracked. If the exit floats with the market, those averages are random processes too. Either way, the lecture's pictures change one thing at a time so you can see which lever moved. One picture is a straight decline in win probability from 0.55 to 0.4 over 1,000 trades. Early trades still build equity. Later, the curve levels off.
How to use it
Plot three series from your own fills: equity, the cumulative win rate, and the cumulative loss rate. Loss rate is one minus win rate if every trade is a win or a loss. If equity rises and then stalls, check which input moved. The win rate may still be 60 percent while the average win has fallen below the average loss. That is a different repair from a win rate that slid from 0.55 to 0.4. You can put average gain and average loss on their own axis when stops are not fixed.
The lecture would not keep trading a rule whose win probability is falling and whose loss probability is rising, if the average win and loss stay fixed. If the win probability later mean-reverts into a clearly positive edge, the same plots are the reason to look again. Moving averages of the win rate are mentioned only as a possible window. The lecture does not develop them, because that opens a separate time-series problem.
Fitting an Ornstein-Uhlenbeck process, a standard mean-reverting diffusion, to the estimated win probability is an optional model of the pictures above. It does not tell you the true law. If you knew that law, you would know whether to continue.
What usually goes wrong
Declaring the rule finished because one week of equity dipped, without looking at the win rate. The fixed positive edge also dips. The difference is whether the dip is noise around a stable mean or a mean that has moved. The opposite mistake is trusting a backtest win rate of 0.55 after the live rate has already spent hundreds of trades near 0.4. The early part of the faded curve still looks like the healthy curve.
In-sample smoothness is not a promise. The lecture says a smoother equity path is less risky while that path is being generated. It also says the next thousand trades need not look like it. Leverage makes both the smooth path and the faded path larger. It does not restore a win probability that has decayed.
Questions people ask
What edge does 0.55, 10, and 8 imply?
The loss probability is 0.45. The edge is 0.55 times 10 minus 0.45 times 8, which is positive. The lecture uses that sign, not a dollar target, as the reason the 1,000-trade curves drift up.
Does a flat equity curve always mean the win rate fell?
No. Wins can stay frequent while the average win shrinks or the average loss grows. The lecture's linear-decay picture is only one of the ways the curve can level off.
When does the lecture say to start again?
When the estimated edge is positive again, for example after a win probability that fell toward 0.42 has returned high enough that the combination with the payoff sizes is clearly above zero. A bounce that still leaves a negative drift was not enough in their example.