Definition
Ratio of actual shoplifting confirmation rate to the CV system's stated confidence within each confidence band, measuring whether security response intensity is anchored to a trustworthy signal.
The Measurement Gap
Every CV-based shoplifting detection system outputs a confidence score that drives security team behavior. High-confidence alerts get immediate attention; low-confidence alerts get deprioritized. The assumption is that an 85% confidence alert should confirm as actual theft roughly 85% of the time. That assumption is rarely tested. Most loss prevention platforms report aggregate detection rates and aggregate false positive rates but do not report calibration by confidence band. Aggregate accuracy can look acceptable while hiding severe miscalibration at specific levels. A model that is overconfident at 80-90% (actual confirmation rate is 40%) and underconfident at 60-70% (actual confirmation rate is 75%) will show reasonable aggregate statistics but systematically mislead security teams at every decision point.
Decision Supported:
Should security teams treat high-confidence alerts differently than low-confidence ones, and at what confidence threshold is intervention justified?
Formula
Formula:
SDCCI(b) = Actual_confirmation_rate(b) / Stated_confidence(b)
Variables:
- Actual_confirmation_rate(b) = confirmed_thefts(b) / total_alerts(b) for confidence band b
- Stated_confidence(b) = midpoint of confidence band b
- Confidence bands: 10-percentage-point bands (50-60%, 60-70%, 70-80%, 80-90%, 90-100%)
Units:
Ratio per band. SDCCI = 1.0 per band means perfectly calibrated. SDCCI > 1.0 means the model is underconfident (more theft occurs than the score predicts). SDCCI < 1.0 means the model is overconfident (less theft occurs than the score predicts). Report as a calibration curve, not a single number.
Edge Cases and Exclusions:
Low alert volume in high-confidence bands requires combining bands to achieve a minimum of 30 alerts per band. Standardize the "confirmed theft" definition across stores (apprehension versus evidence without apprehension). Alerts where security did not engage due to staffing or timing constraints cannot be confirmed or denied and should be reported as "untested" with the untested rate tracked separately. Seasonal variation in shoplifting patterns requires rolling rather than fixed-period calibration computation.
Proposed formula, pending validation pilots.