Anomaly Detection
Written by Michelle Henley
Updated at September 16th, 2024
Table of Contents
| Description | Identifies anomalous transactions according to the transaction's value relative to a historical period of related transactions (1 year). Related transactions are those of the same type (e.g., consulting, promotional speaker, sales/marketing) and nature (e.g., provided lodging, HCP in-office lunch, payment: check). |
| Domain(s) | HCX |
| Analysis Type | Context Flag |
| Focus Area | Value |
| Score Methodology | The 'z-score' (number of standard deviations from the mean) for the transaction relative to its peer group. By definition, 68% of transactions are within 1 standard deviation of the mean, while 95% are within 2 standard deviations, and finally 99.7% of transactions are within 3 standard deviations. It is recommended to avoid having a criteria with less than 2. |
Default Scoring Criteria
Importance: 5 (default)
Enabled: True (default)
| Risk Result | Default Value | Notation |
|---|---|---|
| Weak | 2.5 |
The 'z-score' (number of standard deviations from the mean) for the transaction relative to its peer group
|
| Moderate | 3 | |
| Strong | 4 | |
| Auto | Not Set |
Unique Configuration
- The historical period for anomaly context is 365 days
- Configured to detect anomalies for items based on the combination of nature and spend purpose
-
All values are log-transformed, which allows us to mitigate the undesirable effects of a “long tail” distribution. This enhances the occurrence of our anomaly detection
Exclusions
- Items less than $40 cannot receive a risk result