Classifying Limit Breaches And Near Breaches

Why Classification Rules Matter

A monitoring result is useful only if its meaning is clear. A value just below a limit, a validated value above a limit, and a statistical signal that has moved unusually far from its baseline are different findings. If they are all labelled simply “red” or “breach,” readers may misunderstand what the metric actually shows.

Classification starts with the metric definition and the limit rule: what is measured, over what period, in what units, and whether equality with the limit counts as a breach. Data validation and duration also matter. A value awaiting validation should not be reported as a confirmed breach, while a short-lived exceedance should not be erased merely because the next observation is lower.

These terms describe related but distinct dimensions. Threshold status tells whether the limit was crossed. A data-driven alert tells whether the result is unusual relative to a baseline. A duration tag describes how long a confirmed exceedance lasted. Keeping those dimensions separate avoids treating “temporary” as a reason to disregard a breach or treating an alert as proof that a limit was crossed.

A Consistent Classification Method

For a metric with an upper limit, define the direction of concern and boundary rule before examining results. If the limit is “must not exceed \(L\),” values strictly above \(L\) breach; if a policy instead says “must remain below \(L\),” equality may breach. Record the rule in words and use it consistently. For a lower limit, reverse the inequalities.

A near-breach band also needs an explicit definition. One common form sets its lower boundary to a fraction \(c\) of the limit, where \(0<c<1\). For an upper limit, a value in the interval from \(cL\) through \(L\) is near the limit but not in breach, assuming the breach rule is \(x>L\). The chosen fraction is a monitoring convention, not a universal statistical constant.

Before assigning a final classification, check whether the observation is valid and belongs to the defined measurement period. If validation is incomplete, record it as pending validation rather than as a confirmed breach or as a clean result. For validated observations, compare the value with the threshold and near band. Then assess data-driven alerts and duration separately, if the monitoring design uses them.

Near Breaches and Data-Driven Alerts

A near-breach band is a threshold-based classification. It describes how close a value is to a specified limit. A data-driven alert instead compares a result with a baseline or expected pattern. For example, a metric may be far from its hard limit but still be unusually high compared with its recent history. Conversely, it may be near the limit without being statistically unusual if the baseline is already close to that limit.

If a monitoring design standardizes an observation using a historical mean \(\mu\) and standard deviation \(\sigma>0\), the standardized value is \(z=(x-\mu)/\sigma\). An alert rule might require \(z\geq 2\) for an adverse upward movement. Such a rule is a signal convention; by itself it does not establish a probability of loss, prove a breach, or show that the baseline assumptions remain suitable.

The standardization follows directly from the purpose of a z-score: subtracting \(\mu\) expresses the observation’s distance from the baseline in the original units, and dividing by \(\sigma\) expresses that distance in standard-deviation units. Thus \(x=\mu+2\sigma\) gives \(z=2\). This conversion allows a monitoring rule to be stated in comparable units, but the rule still needs a defined baseline period and a stated alert threshold.

Temporary Exceedances: A Duration Tag, Not an Eraser

A temporary-exceedance rule needs an observation frequency, a maximum duration, and a recovery condition. For example, a hypothetical policy might define a temporary exceedance as a single validated monthly breach followed by a validated observation at or below the limit in the next scheduled month. That definition cannot be applied to a missing next observation, an unvalidated value, or a period longer than the stated maximum.

The classification should preserve both facts: the limit was breached, and the breach met the rule for being temporary. This distinction matters for reporting and trend review. A subsequent return below the limit does not change the historical value that crossed it. The label describes duration; it does not retroactively change the breach boundary.

Worked Classification Across Threshold and Baseline Signals

A complete monitoring record may include both a hard limit and a data-driven alert. Consider a hypothetical monthly service metric with an upper limit of 60 delayed cases. Its alert baseline has mean \(\mu=40\) cases and standard deviation \(\sigma=5\) cases, calculated from a defined historical period. The stated upward-alert rule is \(z\geq2\). Assume all observations are validated.

Common Classification Errors

• Using the wrong boundary. If “90% of an 8.0% limit” defines the near-band start, calculate \(0.90\times8.0\%=7.2\%\). Subtracting 0.9 percentage points to obtain 7.1% applies a different rule.
• Ignoring the equality convention. A value equal to the limit is a breach only if the stated rule includes equality. Do not infer the boundary convention from a chart color.
• Calling every alert a breach. A baseline comparison and a hard-limit comparison are separate tests. Report each result separately.
• Calling an unvalidated value confirmed. Record its status as pending validation until the required checks are complete.
• Using “temporary” to dismiss the breach. Preserve the confirmed breach classification and add the duration tag only when its definition is met.
• Comparing incompatible measurements. Verify that units, reporting periods, denominators, and aggregation methods match the metric definition before comparing a value with its limit.

Consistent labels make monitoring results interpretable across reporting periods and readers. The essential discipline is to state the threshold rule, validation status, alert rule, and duration rule explicitly, then apply each without allowing one classification to substitute for another. This provides a reliable description of what the data show; decisions about what follows are a separate part of risk management.