A decade of invoices changes the question
This is the sixth paper in the African Revenue Intelligence Research Series, instantiating ARIM (Paper 1) for Rwanda. It explores the mature e-invoicing environment after more than a decade of EBM operations, demonstrating how risk shifts from missing-trader chains (Kenya, Paper 2) to invoice-authenticity and invoice-selling schemes, while contrasting with Nigeria's proxy tier (Paper 4) and South Africa's periodic data constraints (Paper 3).
The Rwanda Revenue Authority introduced Electronic Billing Machines in March 2013 and upgraded to EBM Version 2.0 in 2017, years before Kenya's eTIMS (2023) or Nigeria's MBS (2025–2026). VAT stands at 18% and corporate income tax at 28%, filed through the E-Tax portal.
So the starting point differs: not “how do we obtain transaction-level data”, but “how do we detect fraud that exploits a mature transaction-level system”. RRA's own reporting documents operators buying or fabricating EBM invoices with no underlying goods or services, purely to support VAT-refund claims.
Rwanda's strongest asset is its data
ARIM grades each source as essential, desirable, optional or potentially unavailable. Here EBM invoicing data, E-Tax returns and past audit outcomes are essential; Rwanda Development Board (RDB) registrations and EAC transit declarations are desirable; Irembo usage data is optional; and mirror statistics with every EAC partner may simply not exist.
Seven typologies guide the models: invoice-authenticity fraud, invoice-selling networks, margin suppression, EAC transit mispricing, shell clusters, mining under-declaration and PAYE ghosting. Each can equally reflect legitimate trade, error or lawful planning.
Asking whether an invoice is real
Where Kenya's model asks whether a counterparty's VAT covers a claimed input, Model RW1 asks whether the invoice corresponds to a plausible real transaction at all. Its AuthenticityScore combines four signals: issuance frequency out of line with business scale, an implausible share of round amounts, an unusually wide set of unconnected buyers, and no evidence of physical capacity to deliver.
No single signal decides. A wholesaler serving many small retail customers has high buyer diversity but no round-amount pattern and a verifiable site, so it correctly scores low. That is why the signals must be read together.
One seller, many strangers
RDB integration lets TINs and registration numbers be cross-referenced deterministically, leaving a smaller review queue for Fellegi–Sunter matching, which is still needed for sellers who register under variant names to widen their apparent buyer base.
On the resulting graph, Model RW2 hunts a signature distinct from missing-trader carousels: one high-out-degree seller linked to many buyers who barely trade with each other. Model RW3 adds cross-border risk, comparing declared transit values with the value implied at the final EAC destination.
Six domains, weighted for maturity
ARIM scores each taxpayer as a domain-weighted sum. For Rwanda, TransactionBehaviour and HistoricalCompliance carry higher illustrative weights than in any earlier country paper, which directly tests the hypothesis that a long enforcement record is more predictive.
Tax behaviour
VAT-to-sales ratio deviation and CIT gross-margin anomalies.
Transaction behaviour
The EBM authenticity score, the strongest and most mature signal in the series.
Network behaviour
RDB director and shareholder centrality plus the invoice-selling signature.
Sector behaviour
Tin, tantalum and tungsten physical-to-fiscal reconciliation.
Historical compliance
Audit adjustments, weighted up after a decade of enforcement history.
External data
EAC transit declarations and mirror-statistics gaps where available.
A flag is a signal, never a finding
On a fully synthetic set of 30 taxpayers across eight sectors, the seeded invoice-selling profile scored 1.767, far above the next at 0.661, driven by round amounts and buyer diversity. The seeded mining margin-suppression case ranked fourth, driven instead by tax behaviour, so the two anomalies need different evidence.
Every case file carries a domain-tagged reason code, including for false positives. Under Law N° 058/2021 on personal data, a sharp separation must not tempt anyone to treat a high score as proof. No detection or recovery figure is claimed until RRA validates the model on its own data.