Mature, but not yet real-time
This is the third paper in the African Revenue Intelligence Research Series, instantiating ARIM (Paper 1) for South Africa. In deliberate contrast to Kenya (Paper 2), South Africa relies on mature periodic filing rather than real-time e-invoicing, shifting the analytical focus to MNE transfer-pricing risk, Global Minimum Tax data, and mining physical-to-fiscal reconciliation.
SARS collects VAT at 15% and corporate income tax at 27% through eFiling, with bi-monthly VAT 201 returns, monthly EMP201 PAYE filings and annual ITR14 returns. Tax-to-GDP of roughly 24–26% sits well below the OECD average of about 34%, a gap National Treasury partly attributes to compliance shortfalls.
Unlike Kenya, South Africa has no real-time e-invoicing mandate; SARS’s Smart Nation roadmap signals it only for 2026–2027. The economy adds MNE-heavy finance, a significant mining sector and, since January 2024, a Global Minimum Tax top-up regime.
A data model graded by what SARS holds
ARIM grades each source. Essential: eFiling VAT 201, ITR14, EMP201/EMP501 and audit-adjustment outcomes. Desirable: the CIPC registry, Global Minimum Tax top-up filings and customs declarations. Optional: National Treasury procurement records and mining production data from the Department of Mineral Resources and Energy.
The Global Minimum Tax filing is the new, high-value source: it lets transfer-pricing risk be read against a group’s own implied effective tax rate.
Three South Africa-specific models
SA1 scores MNE transfer pricing from related-party ratio, margin deviation, royalty intensity and the Global-Minimum-Tax-implied tax rate, and scores VAT-carousel risk through cycle detection on periodic data. SA2 reconciles production, grade, recovery rate and commodity price against declared sales for platinum-group metals, gold, coal and diamonds.
SA3 applies Louvain community detection and betweenness centrality to the CIPC director and shareholder graph, scored by internal density, incorporation recency and external-trade ratio.
Resolving who is who, then who trades with whom
The SARS taxpayer reference and CIPC registration number are strong deterministic identifiers. Fellegi–Sunter probabilistic matching is still needed for MNE group entities registered under different names and jurisdictions, and for shell clusters whose directors use minor name variants.
The resulting taxpayer graph links companies, directors, shareholders, MNE group entities and customs codes. Its trading edges are inferred from periodic VAT and CIT filings, a structural limitation the models are designed around.
A sector-weighted, explainable score
Because filing is periodic, ARIM lowers the transaction-behaviour weight and raises sector behaviour. In the synthetic demonstration of 30 taxpayers across eight sectors, sector behaviour carries 0.35 against 0.20 for each of tax, network and historical compliance.
Tax behaviour
VAT-to-sales deviation, input/output VAT at bi-monthly cadence, CIT gross-margin anomaly.
Transaction behaviour
Filing timeliness and refund frequency, weighted lower without real-time verification.
Network behaviour
CIPC director and shareholder centrality, community risk and VAT-chain cycles.
Sector behaviour
Related-party ratio, margin deviation and the Global Minimum Tax implied tax rate.
Historical compliance
Prior SARS audit-adjustment count and magnitude, and objection frequency.
External data
Mining, procurement-payment and customs mirror-statistics reconciliation gaps.
The seeded MNE profile ranks first, driven by a 62% related-party ratio and a 6% implied tax rate. The seeded mining profile ranks second through its 55% physical-to-fiscal gap, which a purely transactional model would miss.
A flag is a signal, never a finding
Every case file carries domain-tagged reason codes, such as related-party transactions at 62% of turnover against a sector norm below 15%. SHAP attributions and counterfactuals apply unchanged; only the feature set differs.
POPIA governs personal data in CIPC-linked analytics, and Global Minimum Tax and mining data warrant strict, separate access control. The demonstration claims no real detection or recovery figures; only a SARS-validated deployment can.