Copper, Certificates and Compliance

A price-banded royalty and a $800 million dispute

This is the ninth paper in the African Revenue Intelligence Research Series, instantiating ARIM (Paper 1) for Zambia. It extends South Africa's mining-sector reconciliation approach (Paper 3) to Zambia's specific price-banded copper royalty structure, while its turnover-tax graduation model draws on parallel threshold analytics from Ghana (Paper 5) and Uganda (Paper 8).

ZRA administers VAT at 16% and Corporate Income Tax at 30% for most sectors. Mining companies pay the standard CIT plus a separate Mineral Royalty Tax on copper output, banded by price — 4% on the portion up to US$4,000 per tonne, 6.5% on the next band to US$5,000, and higher rates above — and deductible against CIT since its 2022 reinstatement.

Alongside it sits a publicly documented VAT-refund dispute. Exporters, copper miners above all, have seen refunds withheld pending import certificates from destination countries, a requirement industry bodies call impractical given third-party commodity trading. The Chamber of Mines of Zambia reports close to $800 million owed.

That distinction shapes the whole design: the dispute is treated as an administrative-verification challenge, not presumed fraud on either side.

Data the model needs, and data it only hopes for

ARIM grades each source as essential, desirable, optional or potentially unavailable. VAT, CIT and turnover-tax returns, Mineral Royalty filings with copper price-band schedules, and historical audit outcomes are essential. Shipping and commodity-exchange settlement records, and the company registry, are desirable.

Destination-country import certificates stay optional — supplemented, not replaced. International price feeds are assumed to come from published exchange data such as the LME rather than a bespoke ZRA feed, and the royalty model needs transaction-time data finer than many systems hold today.

One TPIN, many entities

Deterministic matching starts from the ZRA Taxpayer Personal Identification Number (TPIN). Fellegi–Sunter probabilistic matching then links TPIN, Patents and Companies Registration Agency records and mining licences, and — for related-party copper sales — offshore marketing entities.

The emphasis follows transfer-pricing logic: a sale priced inconsistently with the price band used for the seller's own royalty assessment only becomes visible once the buyer is known to be related.

Following copper through related parties

The taxpayer graph links TPINs, companies, directors, shareholders, licence holders and offshore marketing entities through directorship, shareholding, related-party sale and, where available, shipping or settlement-linked transaction edges.

Seven Zambia-specific typologies are tracked: royalty under-declaration, royalty/CIT interaction manipulation, VAT export-refund disputes, transfer-pricing risk, turnover-threshold manipulation, shell-company networks and customs under-invoicing.

Three models, six domains

ZM1 sums volume times band rate across every sale, then takes the gap between expected and declared royalty. ZM2 replaces a yes/no certificate test with an export-plausibility score from shipping match, settlement match and export history, less a penalty for the missing certificate. ZM3 routes each taxpayer to the right threshold: ZMW 800,000 for artisanal and small-scale miners, ZMW 5 million (5% turnover tax from 1 January 2025) for everyone else.

These feed the six ARIM domains, weighted by expert judgement until labelled audit outcomes allow calibration.

A missing certificate is a gap, not a verdict

On a synthetic dataset of 30 taxpayers across eight sectors, the seeded copper miner with a large royalty gap and weak export evidence scored 0.926, far above everyone else. Copper miners with moderate gaps but plausibility of 0.55–0.62 ranked much lower, because a documentation gap with strong corroboration differs from one without.

Five hypothetical cases test this, from a royalty mismatch with weak evidence to a genuine exporter lacking only a certificate, plus an artisanal miner declaring ZMW 790,000 against the ZMW 800,000 line. Every case file carries domain-tagged reason codes, and the chain from risk score to prosecution never skips a human step.

The results are mechanical coherence on synthetic data, not real detection performance. The ~$800 million figure is industry-sourced and unverified here. Under the Data Protection Act, 2021, any shift to alternative evidence should come with published criteria for refund release, so exporters meet a predictable standard.