Methodology & Mathematical Foundation

1. Official Data Sources & Provenance

The monitoring system collects data exclusively from official, publicly published government sources:

• MDULS (Ministry of Public Administration and Local Self-Government): Official voter totals by municipality and weekly municipal roll modification summaries.
• RGZ (Republic Geodetic Authority / GeoSrbija): Official national Address Register containing 2,488,964 geocoded house numbers and 96,048 street names.

Every collected raw HTML or CSV snapshot is cryptographically hashed (SHA-256) and archived in append-only storage. observed_at (when our collector fetched the data) is strictly distinguished from source_date (the official point-in-time timestamp assigned by MDULS).

2. Per-1,000 Normalization & Percentile Ranks

Serbian municipalities vary drastically in electorate size — from Crna Trava (~900 voters) to Novi Sad (~340,000 voters). Absolute counts cannot be compared directly.

All change categories are normalized per 1,000 registered voters using the municipality's baseline voter total:

Municipalities are then ranked cross-sectionally to obtain national percentile ranks from 0.0% (lowest activity) to 100.0% (highest activity).

3. Provisional Anomaly Score V1 & Non-Overlapping Components

The Anomaly Score (0–100) measures how unusual a municipality's weekly roll activity is relative to national peers. In the current single-snapshot release, scores are strictly provisional cross-sectional indicators because historical voter delta, multi-snapshot deviation, and roll reconciliation cannot yet contribute.

Three Non-Overlapping Count Components & Derived Inflow Modifier: To prevent double-counting correlated activity, components are strictly partitioned:

• Residence Registrations (Max 40 pts): Exact residence-change ADD count category normalized per 1,000 voters.
• Other Additions (Max 25 pts): All remaining ADD categories (total ADD minus residence registrations) per 1,000 voters.
• Removals (Max 20 pts): Total removals count category per 1,000 voters.
• Inflow Imbalance Modifier (Max 15 pts): A derived interaction modifier based on the relationship and volume ratio between total additions and removals (≥ 15 adds), not an independent event category.

The Small-Number Guard & Strong Evidence Rule: In small municipalities, tiny absolute counts (e.g. 2 residence registrations in Crna Trava) produce mathematically high per-1,000 rates. To prevent false alarms, our scoring engine applies an evidence dampening multiplier based on raw event count:

Furthermore, an EXTREME (80–100) score strictly requires substantial absolute scale (≥ 25 residence registrations or ≥ 40 total events) OR multiple independent strong components (≥ 18 pts each), preventing isolated small counts from breaching extreme bands.

4. Roll Reconciliation Mathematical Model

When multiple official voter snapshots exist, the system verifies internal arithmetic consistency:

Any non-zero reconciliation gap is flagged for audit, as it indicates additions or removals occurring outside standard published categories.

5. Address-Level Matching & Anomaly Signals

Aggregate address-voter data is deterministically matched against the complete 2.48M RGZ Address Register. Signals include:

• House Number 0 (HOUSE_NUMBER_ZERO): Addresses with numeric house number "0" (ghost / unassigned cadastral numbers).
• Address Not in RGZ (ADDRESS_NOT_IN_RGZ): Addresses that cannot be located in the official national address register.
• Ambiguous Address Match (ADDRESS_MATCH_AMBIGUOUS): Addresses matching multiple conflicting records requiring manual inspection.
• High Voter Count at Address (HIGH_VOTERS_PER_ADDRESS): Unusually high voter concentration (>50, >100, >250 registered voters at a single house number).
• Sudden Voter Spike at Address (SUDDEN_ADDRESS_SPIKE): Rapid addition of voters at an address between snapshots.

6. Neutral Anomaly Terminology

All signals produced by this system represent statistical anomalies and data-quality indicators. They serve as objective pointers for public oversight and institutional verification. An anomaly flag does not automatically establish or infer fraudulent conduct.