LessonThis is a lesson restated in new words. Any figure in it is an example from the source, not this desk's record.
A profitable backtest says the average trade was positive in the sample you already have. A tradable strategy, in this lecture, is one whose win distribution and loss distribution still look like that sample when you step forward. An Nvidia example is profitable from 2020 through 2023, with a win rate near 57 percent and an expected value of 89 per trade, then fails from 2024 through the end of 2025 because the averages change shape. The win rate barely moves. The payoff sizes do.
Why it matters
Asking whether a trader "made money last year," or quoting a Sharpe or Sortino ratio, is called a backward question. Edge is the forward one, and edge is defined as a positive expected value, the way a casino game is positive for the house across many plays, not on each play. Risk is the expected deviation from that value, which is what variance measures. The lecture sets ergodicity and bet sizing aside and assumes the book is operating in a reasonably ergodic way so that the expectation is allowed to describe the path's drift.
The law of total expectation splits a trade into the average winner times the win probability, plus the average loser times the loss probability. Winners and losers partition the outcomes. If a strategy is tradable, a walk-forward or a strict out-of-sample window should show similar win and loss distributions and similar probabilities. The 2020-2023 Nvidia run has a win rate around 57 percent and an average winner larger than the average loser, so the edge prints as 89 and positive. The later window keeps a similar win rate and changes the averages enough that the loser dominates. Expectation turns negative. Same rules, different distributions.
The idea in plain words
Stability is not magic, and it is not what you get by resimulating until a curve looks good. The lecture calls it a data problem: you are trying to trade a structure that means something economically or statistically. Discretionary trading and quantitative trading are said to live on that same mechanism. The same entry does not have to draw from one gain law at every time. Hidden and visible processes move the shape.
One attempt to pin the shape down uses a volatility regime: rolling quadratic variation, squared returns over the previous 20 days, cut at the 33rd percentile and the 66th, with the top group running up toward the 99th. Low, mid, and high are the three buckets. The hypothesis is that one entry rule will show different gain distributions in each bucket. From 2020 through the end of 2023, the low and high buckets have negative expectancy and the mid bucket has positive expectancy. From 2024 onward, the low bucket stays negative, the high bucket flips from negative to positive, and the mid bucket stays in a similar positive range. The lecture cites mid-regime figures of about 0.66 and 0.53.
How to use it
Treat regime stability as upstream of the gain chart. If the feature that defines the entry is unstable, the gain law downstream has little chance. The lecture's question to small funds, in the 30-to-100-million range, is to take the feature or the entry and look at its distribution, or a principal-component view, across time. Stability there is offered as the reason a gain distribution might also repeat.
What usually goes wrong
Optimizing a Sharpe ratio on the full sample, then going live, skips the stability check. So does a percentile regime that looks principled and quietly dumps crisis volatility into the low bucket. The mid bucket's calmer handoff, 0.66 versus 0.53 in the lecture's figures, is not a recommendation to trade only medium volatility. It is evidence that this particular splitter was kindest to the middle.
Questions people ask
What broke in the later Nvidia window?
Not the win rate, which stayed roughly the same. The average winner and average loser changed enough that expectancy turned negative.
Why was the middle volatility bucket calmer?
Because terciles on a rolling window shove extreme periods into the outer buckets. The lecture finds the smallest distribution distance, and the steadier expectancy, in the middle.
Does a positive backtest mean the strategy is tradable?
Only if the win law, the loss law, and their probabilities still resemble the backtest out of sample. In this example they did not.