Kenya starts from a stronger data position
This is the second paper in the African Revenue Intelligence Research Series, instantiating the ARIM framework (Paper 1) for Kenya. It investigates how real-time eTIMS invoice feeds and mobile-money signals can transform missing-trader and shell-network detection, establishing foundational transaction-level methods for subsequent papers in the series.
KRA collects VAT at a 16% standard rate (8% on petroleum), corporate income tax at 30% for resident companies and PAYE on a progressive scale, all through iTax. Since 1 September 2023, VAT-registered taxpayers must also issue invoices through eTIMS, which transmits the data to KRA in real time and is still expanding to smaller businesses.
That is a materially different starting point from many peer administrations. It lets ARIM put its weight on real-time VAT-chain and network analytics rather than on basic data consolidation. The central question is whether eTIMS invoice data, and mobile-money signals where lawful, find missing traders and shell networks that return-level data and manual audit selection miss.
Data graded by what Kenya can actually reach
iTax registration and returns, together with eTIMS invoice data, are essential. Customs declarations, Business Registration Service (BRS) company and director data, and past audit outcomes are desirable. M-Pesa data and the land registry are optional, and EAC cross-border mirror statistics are potentially unavailable.
No model that leans on optional or unavailable sources is a precondition for deploying the essential-tier models, so KRA can begin with what it already holds.
Resolving who is who: PINs and BRS records
Matching starts deterministically on the KRA PIN, supplemented by the BRS registration number for companies. Where PINs diverge but other evidence suggests common control, probabilistic Fellegi–Sunter matching weighs name similarity, a shared address, a shared director, a shared eTIMS device or, where accessible, a shared mobile-money merchant ID.
Some people hold several PINs or trade under varied business names, so medium-confidence matches go to manual KRA review rather than being linked automatically. The hypothesis is that resolving across PIN and BRS records will surface materially more common-control relationships than PIN matching alone.
From invoice chains to shell networks
For each input-VAT invoice claimed through eTIMS, the first Kenya model checks whether the issuing counterparty declared enough output VAT on the same invoice. The gap between claimed and verified input VAT marks a missing-trader candidate. Because eTIMS is real-time, the check can run near-continuously instead of only when returns are filed.
A Kenya taxpayer graph then links PINs, BRS directors and shareholders, eTIMS device IDs, addresses and buyer–seller invoice edges. Louvain community detection and betweenness centrality surface clusters of newly registered companies that share directors or an address and trade heavily with each other, a pattern documented in Nairobi’s trading and professional services. Five hypothetical case studies illustrate this; none identifies a real taxpayer.
Six domains, one explainable score
Kenya uses the same domain-weighted sum as the wider series, with two adjustments. The transaction domain carries a higher illustrative weight because of eTIMS, and the external-data weight stays at zero until mobile-money access has a proper legal basis.
Tax behaviour
VAT-to-sales deviation, input/output VAT ratio and margin anomalies against KRA sector benchmarks.
Transaction behaviour
The eTIMS invoice-chain gap, invoice-issuance timing anomalies and iTax filing timeliness.
Network behaviour
Director centrality from BRS records, community risk and cycles in the eTIMS transaction graph.
Sector behaviour
Margins and payroll benchmarked separately for trade, agri-processing, services and construction.
Historical compliance
Prior KRA audit adjustments and how often assessments are objected to.
External data
M-Pesa turnover versus declared sales where lawful; customs mirror-statistics gaps for importers.
On a fully synthetic set of 30 taxpayers, the three seeded anomalies (a missing-trader profile, a network hub and margin suppression) rank in the top four. The top-ranked case rises mainly on the eTIMS gap and network risk. Rank three, a high-margin professional practice with a low gap, is plausibly legitimate, which is exactly why every flag needs explaining before a case opens.
A flag is a signal, never a finding
Every Kenya case file carries domain-tagged reason codes, for example that 55% of claimed input VAT could not be verified against counterparty eTIMS filings, or that a taxpayer is linked by a shared director into a dense transaction cluster. SHAP attributions supply the detail, and KRA’s existing objection and appeal rights under the Tax Procedures Act stay unchanged.
eTIMS concentrates near-real-time data on the whole VAT-registered base, so it needs strong segmentation and an independent security assessment (ISO/IEC 27001 or an equivalent Kenyan standard) before production use. The results are mechanical, not real-world: no detection or recovery figure is claimed until KRA validates the model on its own data.