A z-score expresses how far a value lies from a reference mean in standard-deviation units. In production, that turns raw data into an alarm rule: compare current missingness or feature statistics with cleaned historical behavior. section 14.5.1
Two practical roles stand out:
- Monitor data quality by setting alarms at three z-scores for important features and four for other features. Use comparable thresholds for expected feature ranges, alongside domain checks such as nonnegative sales values. section 14.5.1
- Plan experiments with Zα for the desired Type I error rate and Zβ for desired power. Combined with estimated variance and minimum detectable change, these values contribute to the required sample size per group. section 12.2.1
A data-quality alarm
Suppose missing values for an important feature rise three standard deviations beyond cleaned historical behavior. Trigger an alarm, then check schema compliance, expected ranges, and invalid statistics before treating the change as a pipeline or data problem. section 14.5.1
