Scale collapse and outlier robustness
Keep a standardized displacement distinct from tail probability when the prior scale is small, unstable or changed.
The denominator is part of the feature
A standardized displacement divides a price difference by an explicitly defined prior scale. Population and sample standard deviation are different conventions, and fit scope determines which observations can enter either. The original lab uses population standard deviation from preceding observations. A constant window has zero scale and an undefined z-score, so the lab records its exclusion rather than assigning a convenient zero. A very small nonzero scale can produce a large score for modest price movement. Record that denominator beside the score before interpreting the event as unusually informative.
Large score does not mean calibrated odds
Normal-tail probabilities require an appropriate distributional and parameter model. Rolling price z-scores, returns and fitted residuals do not automatically satisfy one merely because their values are standardized. Dependence, skew, outliers and estimation all matter. The new path lens has no scale fit or probability output; its absolute band is an entered tolerance. Changing that tolerance alters endpoints without estimating statistical rarity. Do not use an authored recovery example to calibrate conviction grades. A genuine probability assessment needs actual source, frozen forecasts or labels and its own honest evaluation process.
A robust alternative is a new specification
A researcher might investigate another scale or location estimator when outliers dominate. That requires a named procedure, causal fit scope and candidate-selection record. Do not retroactively replace a failed denominator with the favorable alternative while retaining the original validation badge. The new guides are not a benchmark of robust estimators, and the site fits none to real data. Keep predecessor and successor identity. The complete native research method determines whether the revised hypothesis reaches independent completion, not a visually cleaner chart on an inspected sample.
Exclusions can alter the apparent success rate
Dropping zero-scale, missing or extreme windows changes the eligible population. Preserve counts and reasons, including whether exclusion was predeclared or introduced after outcomes were known. A threshold subset cannot be compared with all raw observations while hiding warmup and final-label exclusions. The original lab's displayed eligible count is descriptive, not an audited account population. Authored fixtures exercise zero and finite-scale behavior only. Financial claims additionally need actual entries, exits, quantities and costs. A statistical denominator does not supply those missing economic objects or independent master evidence.
Authored reference and excursion case
An invented prior window is constant at 100, so its scale is zero and a current 100.1 has no defined prior-window z-score. In a different authored window the scale is 0.01, giving displacement ten for the same 0.1 price difference. Neither case establishes a normal probability or a real trading event. The example makes zero-scale exclusion and small-denominator interpretation explicit, rather than claiming that a large standardized number predicts recovery.
| Object | Record | Bounded finding |
|---|---|---|
| Constant window | Scale 0 | Undefined score |
| Small-scale scenario | Scale 0.01 / difference 0.1 | Score 10, not calibrated odds |
| Changed estimator | New procedure needed | No inherited validation |

Primary references and scope
Source review: 14 October 2026. Official mechanism descriptions are not this desk's paid-account test, independently captured fills or vendor-return audit.