Delete the best trade. If the edge disappears, it was never the edge.

The battery

The same sample, computed five ways, with the outliers treated differently each time.

Realised EV is the plain mean and the number everyone quotes. EV without the best trade removes the single largest winner and recomputes — the bluntest test there is and often the most informative. Trimmed or winsorised EV pulls in the extreme tails at both ends rather than deleting them, which is the fairer version of the same idea. Median trade EV ignores magnitude entirely and reports the middle outcome, describing what a typical trade actually does. The EV lower bound goes further and asks what the sample supports at the conservative end of a confidence interval rather than at its centre. Read together they answer whether the headline figure describes the population or a tourist attraction inside it.

Reading the spread

The gap between the readings is the diagnosis; the readings themselves are not.

None of the five figures is the correct one, which is the part operators find uncomfortable. What carries information is how far apart they sit. A tight cluster means the edge is broad — most trades contribute, deleting any one changes little, and the mean is an honest summary. A wide spread means the expectancy is concentrated in the right tail, and everything downstream inherits that fragility: sizing decisions, the confidence with which a branch is deployed, and the emotional experience of a stretch in which the tail simply does not arrive. Skew dependency, the difference between mean and median, is the single number that summarises the spread.

Tail dependence is not automatically a fault

Some branches are supposed to be carried by their right tail. The test is whether the right one is.

This is where the robustness family has to be read alongside the branch lens rather than on its own. Trend No-Partial exists precisely to produce a small number of large outcomes; a TNP sample whose expectancy collapses without its best trade is behaving as designed, and 'fix the tail dependence' would be an instruction to destroy the branch. Normal is the opposite case. It is a static, controlled, high-frequency branch, and if its expectancy is carried by one outsized result then something has gone wrong — either a trade was misclassified, or the branch was quietly managed as a trend trade and nobody logged it. The same statistic reads as health in one branch and as contamination in another.

• Tail dependence in Trend No-Partial is the branch working, not a defect to repair.
• Tail dependence in Normal is usually misclassification or undeclared identity drift.
• Always ask which branch produced the outlier before deciding what the spread means.

The mirror test

Deleting the worst trade is the same test, and operators only ever run the first half.

The battery is symmetric and human attention is not. Removing the largest winner feels like a stress test; removing the largest loser feels like cheating, so it rarely gets run. It should be. If expectancy improves dramatically when the worst trade is removed, the sample contains a loss the system was not designed to take — a stop that was widened, a trade held through news, a Normal position that became an unmanaged runner. That is a governance finding rather than a statistical one, and it is invisible in the without-best-trade reading. The honest practice is to run both deletions and to treat the two results as separate reports.

What robustness cannot do

None of this makes a thin sample trustworthy; it only makes an untrustworthy one legible.

Winsorisation, trimming and lower bounds all reduce the influence of extremes, which can make a small sample look calmer than it deserves to. They are diagnostics for the shape of a distribution, not manufacturers of confidence, and applying them to twelve trades produces five stable-looking numbers that are all describing noise. The minimum sample gate governs here as everywhere else. The correct use is the one the workbook states directly: assess robustness with these tools, do not use them to hide trades you would rather not have taken.

The key idea

The question is not whether expectancy is positive but whether it would survive a boring month.

An operator sizing against a mean is implicitly betting that the sequence which produced it repeats, tail included. The robustness battery converts that implicit bet into a stated one. If the without-best-trade figure still supports deployment, the plan survives a stretch in which nothing extraordinary happens — which describes most stretches. If it does not, that is worth knowing before the boring month rather than during it, because during it the number the operator will be looking at is the one that already collapsed.