The gap is in compliance, not in rates
IMF revenue-gap analysis and the ATAF African Tax Outlook consistently find that compliance gaps, rather than statutory tax rates, are the larger constraint on domestic resource mobilization across the continent. UNCTAD and the OECD add that trade mispricing and profit-shifting risk concentrate in the extractive sectors.
At the same time, VAT invoicing, customs single-window declarations, company registries and, increasingly, mobile-money records are outstripping manual audit-selection capacity. The opportunity is real; the data-science capacity to seize it is what most administrations lack.
Data that is uneven, never assumed
ARIM grades every candidate data source as essential, desirable, optional or potentially unavailable. Taxpayer registration and VAT, CIT and payroll returns sit at the core; customs declarations and company registries add depth; beneficial-ownership registers and bank or mobile-money data are bonuses where lawfully accessible.
The framework therefore degrades gracefully: it computes the domains it can, flags the ones it cannot, and lets an administration begin with returns and payments alone before adding more.
Resolving who is who
Where taxpayer identifiers are fragmented, one beneficial owner often appears under several registered entities. ARIM therefore treats entity resolution as a prerequisite layer, combining deterministic rules with Fellegi–Sunter probabilistic matching to link records without inflating false matches.
The hypothesis is that resolved entities will reshuffle risk rankings relative to treating every registered TIN independently — because control, not registration, is what carries risk.
From single taxpayers to networks
Once identities are resolved, a taxpayer graph links directors, shareholders, addresses and transactions. Centrality and community-detection algorithms then separate organised shell-company and circular-trading clusters from legitimate group structures — patterns invisible in any one taxpayer's own returns.
Six domains, one explainable score
ARIM scores each taxpayer as a domain-weighted sum across six behavioural domains. Domain weights start from expert judgement and move to regression-calibrated weights once enough labelled audit outcomes exist, with data-poor domains carrying lower weight by default.
Tax behaviour
Patterns within filed returns: margin anomalies, VAT ratios, refund frequency.
Transaction behaviour
Filing timeliness, payment defaults and volatility in transaction values over time.
Network behaviour
Centrality, community risk and cycle involvement within the ownership graph.
Sector behaviour
Deviation from sector benchmarks, including physical-to-fiscal gaps in extractives.
Historical compliance
Prior audit adjustments, appeals and payment-default history.
External data
Customs mirror statistics, procurement payments and bank turnover versus declared sales.
A flag is a signal, never a finding
Every model output, from logistic regression and gradient boosting to Isolation Forest and Bayesian updating, is converted into an auditor-readable reason code using SHAP attributions and counterfactuals. The score tells an auditor which kind of risk was flagged, such as network risk rather than merely high risk.
Because fraud prevalence is typically well under five percent, performance is judged on precision-at-top-K, PR-AUC and calibration rather than raw accuracy — and the framework deliberately claims no real-world detection figures until an administration validates it on its own data.