Abstract
Ghana's gold exports doubled in 2025 to about US$21 billion, and small-scale mining supplied more than half of that value for the first time. The Institute of Fiscal Studies reports that mineral royalties rose only 21% and came entirely from large-scale mines; the small-scale sector paid little or no royalty or corporate tax. The state-owned Ghana Gold Board (GoldBod) now buys most artisanal output, so for the first time the government holds the transaction record for the untaxed half of the industry. This paper models the small-scale sector as a channel-choice problem, where every miner chooses between GoldBod and a smuggler according to price, payment speed and levy, and derives the revenue-maximising levy and the conditions under which it stays low. It then specifies a large-scale fiscal model for the mining-law review announced in January 2026, and an indirect-method production reconstruction for audit. The illustrative result is that a levy of 1.5 to 2% on GoldBod purchases, with same-day payment and a price at or above the informal market, raises more than a 3% levy that pushes gold back across the border; on 2025 volumes the difference is worth roughly US$130 to 160 million a year.
Introduction
Ghana is Africa's largest gold producer and, since 2025, a country where the informal half of the industry outproduces the formal half. Artisanal and small-scale mining (ASM) reached a record 104 tonnes in 2025, and GoldBod expects 2026 to match or exceed it. The revenue system was designed for a handful of large licensed mines with audited accounts. It has no working instrument for two million people digging in the Western and Ashanti regions and selling to a state buyer within days.
The research question is how to tax a sector whose output is fully mobile across a porous border, whose sellers are price-sensitive to a few percentage points, and whose transactions now run through one state counter. A second question follows for the large-scale sector: how to raise the state's share in the mining-law review without repeating the exploration collapse that followed previous rounds of fiscal tightening in West Africa.
The organising identity is the series' gap decomposition:
R^{pot} - R = \underbrace{\left(R^{pot} - R^{law}\right)}_{\text{policy gap: no workable ASM instrument}} + \underbrace{\left(R^{law} - R\right)}_{\text{compliance gap: smuggling, under-declared grade}}Current tax issues
2.1 The untaxed half
Ghana's mineral royalty regime was written for a dozen large mines. Each holds a lease, files audited accounts, sells doré through a handful of refiners and pays royalty on the value of production, corporate tax on profit and, since 2025, a sliding-scale royalty of 5 to 12% linked to price. The Minerals Income Investment Fund reports that in 2025 all of the US$442 million of mineral royalty came from those mines. Small-scale mining, which under the Minerals and Mining Act is reserved for Ghanaian citizens and operates on small concessions with limited machinery, exported gold worth about US$10.8 billion in the same year and paid, on the Institute of Fiscal Studies' reading, no royalty. Corporate income tax from the sector is also negligible, because almost none of the licensees files a return.
This is not new. Small-scale gold has been under-taxed since the sector was legalised in 1989, and every attempt to change that has run into the same wall. The 2020 experiment is the clearest case: a 3% withholding on small-scale gold exports was introduced in May 2020, small-scale output fell sharply as sellers moved to smugglers and Togo's recorded gold exports rose, and the rate was cut to 1.5% within a year and reconsidered again after that. The lesson that took hold in Accra was that small-scale gold cannot be taxed. The correct lesson was narrower: it cannot be taxed at a rate that exceeds the seller's cost of switching channel, and in 2020 that cost was low because the state buyer paid less than the smuggler and paid late.
The scale has since changed. Small-scale output reached a record 104 tonnes in 2025, overtaking large-scale for the first time, and GoldBod expects 2026 to match or exceed it. A sector that produces more than half the country's gold and contributes almost nothing to the Treasury is the largest single revenue gap in Ghana's fiscal system, larger than any plausible gain from the large-scale law review. It is also the gap the state is now best placed to close, for the reason set out next.
2.2 GoldBod changes the data, not yet the tax
The Ghana Gold Board was established by Act in 2025 and has operated since 1 May 2025 as the sole regulator and accredited buyer of gold from small-scale miners, with authority over licensed aggregators and over re-exports. It replaced the Precious Minerals Marketing Company's buying role and the licensed-buyer system that had run alongside it. GoldBod's stated purposes were foreign exchange and the suppression of smuggling, and on both it has done what it set out to do: it purchased 50 to 54 tonnes in the first half of 2026, reports near US$11 billion of foreign-exchange earnings from the sector in 2025, and the cedi's recovery in 2025 owed a good deal to those inflows.
What GoldBod has created, without setting out to, is a transaction register. Every purchase is weighed, assayed for fineness, priced against a published reference, and paid electronically to an identified seller at a dated time. That record is the custodian record of the series overview: a record of the taxpayer's activity held by a party whose interests are not the taxpayer's. Revenue authorities in most countries spend a decade trying to build a register of this quality for an informal sector and fail. Ghana has one, covering the untaxed half of its gold industry, as a by-product of a foreign-exchange policy.
No tax is yet attached to the counter in a form that reaches the Treasury as revenue. GoldBod earns a margin between its buying and selling price, part of which flows to the state as the board's surplus, but that is a trading margin, not a tax, and it is not reported as mineral revenue. The paper's argument is that the register makes a small levy administrable for the first time, provided the terms at the counter, price and payment speed, are set so that the levy does not repeat 2020. Section 3.1 gives the model; Part II works the numbers.
2.3 Large-scale fiscal terms under review
The Minerals and Mining Act of 2006 and its fiscal provisions were set when gold was below US$700 an ounce. In January 2026, with the price above US$5,000, the Minerals Commission announced a rewrite intended to increase the state's share of revenue. The 2025 sliding-scale royalty of 5 to 12% by price was the first step; the review is expected to address state participation beyond the current 10% free carry, the treatment of development agreements, local content and the fiscal terms for renewals. The Minerals Commission's acting chief executive described the 2014 fiscal policy as never having been reviewed and the change as necessary to close the gap.
The Chamber of Mines has not opposed a higher state share in principle but has warned that the proposals as floated would undermine investment. Two structural facts shape what the review can achieve. First, Ghana's largest producers, Newmont at Ahafo and Akyem, Gold Fields at Tarkwa and Damang, AngloGold Ashanti at Obuasi and Iduapriem, and Perseus at Edikan, operate under development agreements with stability clauses ratified by Parliament. New fiscal terms bite on those mines only at renewal or by negotiated amendment. The review's revenue therefore comes first from new projects and renewals, and any estimate that applies the new terms to total output is wrong by the stabilised share. Second, Ghana competes for capital with Côte d'Ivoire, Senegal and Guinea, none of which has tightened terms in the way Mali and Burkina Faso have, and all of which share the Birimian geology.
The paper's position is that the review's success depends less on the headline rate than on the instrument. A price-linked royalty is simple and it captures price windfalls, but it treats a high-cost mine at US$4,000 the same as a low-cost one, and it raises take when prices fall relative to costs, which is when marginal mines close. A margin-linked instrument or a rent tax captures the same windfall without that flaw. Section 3.2 sets out the comparison and Part II runs it on a representative new mine across three prices.
2.4 Price windfalls and the royalty base
Gold averaged above US$4,000 an ounce for most of 2026 after a January peak above US$5,000. For a mine with an all-in cost of US$1,500, the margin at US$4,000 is US$2,500 an ounce, of which the 2006 Act's 5% royalty takes US$200 and the sliding scale at that price takes perhaps US$360. The rest of the windfall reaches the state only through corporate tax on declared profit and dividends on its 10% carry. Ghana's take share on a representative mine, worked in Part II, is about 0.41 at US$4,500 under current terms and 0.34 at US$2,500: the state's share of the rent falls as the rent rises, which is the opposite of what a rent-capturing regime does.
The sliding scale by price is a rough proxy for a profit-based instrument, and its roughness has a cost. Ahafo and Tarkwa have different costs per ounce; at the same price the scale charges them the same royalty rate, so it over-taxes the high-cost mine relative to its rent and under-taxes the low-cost one. At the bottom of the cycle, when the price falls to US$2,500, the scale's rate falls to its floor but the floor still applies to a mine that may be earning nothing above cost. The design that fixes both problems slides on margin rather than price, or taxes returns above a hurdle. Neither is exotic; Chile, Zambia and Tanzania each have a version. Section 3.2 and Part II show what the choice is worth.
2.5 Illegal mining and environmental cost
Galamsey, the illegal end of the small-scale spectrum, has destroyed river systems in the Western, Ashanti and Eastern regions, turned the Pra, Ankobra, Birim and Offin rivers turbid with silt and mercury, forced Ghana Water Company to shut treatment plants, and cleared forest reserves and cocoa farms. The state bears the cost of remediation, of water treatment, of lost cocoa output and of the health effects of mercury, and none of it is recovered from the operators who cause it. The Environmental Protection Agency, the Minerals Commission, the Forestry Commission, the Water Resources Commission and GoldBod each hold part of the data on who is mining where; none of them holds the whole, and none links its data to the tax system.
The fiscal and environmental problems are the same problem seen from two sides. The sellers who deliver to GoldBod are, in large part, the same operators whose excavators are in the rivers. A register of resolved operators with their volumes, built from GoldBod's data as section 3.4 describes, is also a register of who should pay for remediation. Satellite change detection, which identifies new excavation and its extent within days, gives a measured footprint per operator. A remediation levy per hectare disturbed, collected through the same counter as the revenue levy, would recover part of the cost and would fall most heavily on the operators whose footprint is largest relative to their delivered gold, which is the signature of the most destructive methods. Section 5 sets out the geospatial component and section 8.6 the levy design.
2.6 Audit capacity for grade and recovery
The value of a mine's output is tonnes milled times head grade times recovery times price. Price is public. Tonnes, grade and recovery are declared by the miner, and each has a margin within which a lower declaration is defensible: grade varies within the ore body, recovery varies with ore type and plant performance, and reconciliation between the mine plan and the mill is never exact. A 3% under-declaration of grade at a 400,000-ounce mine at US$4,000 is US$48 million of revenue, US$2.4 million of royalty at 5% and a larger sum of corporate tax, and it is difficult to challenge without independent evidence.
The Ghana Revenue Authority's mining audit unit is small relative to the number of leases and has limited access to the records that would give it that evidence: reagent purchases, which are imported by a few distributors whose customs entries the GRA itself holds but does not link to mines; power, which the mines buy from the grid or generate from imported fuel; refinery receipts, which the refiners hold; and pit volumes, which satellites measure. Grade and recovery under-declaration is therefore a compliance gap with a known remedy, third-party reconstruction, that has not been applied because the data has not been assembled. Section 3.3 sets out the reconstruction, section 5 the engine that runs it, and Part II a worked example in which a declared 180,000 ounces is tested against four independent sources.
Models
3.1 Model A: channel choice and the ASM Laffer curve
The model treats every small-scale seller as making a choice. Gold in hand, the seller can take it to a GoldBod counter, or to an aggregator who will carry it to Lomé, Ouagadougou or Dubai. The formal channel pays a known price, with a levy deducted, after a known delay, with no risk of seizure. The informal channel pays a price that is often higher, because the buyer avoids the levy and any margin the state buyer keeps, but it pays it with risk, with a transport cost and with whatever the aggregator takes. The seller compares the two.
The model writes the comparison as a difference in expected value. The formal channel's advantage is a constant a, which captures trust, convenience and the value of a legitimate sale; against that stand the effective discount Delta, which is the sum of the price gap, any surrender rule and the levy, scaled by a sensitivity b; the payment delay d scaled by c; and enforcement pressure rho, which raises the cost of the informal route. Sellers differ in unobserved ways, so the model is logistic: the share choosing the formal channel rises smoothly as its advantage rises. Each small-scale seller chooses the formal channel (GoldBod) or the informal one (a smuggler to Togo, Burkina Faso or a Gulf refinery) by comparing net price after levy tau, the informal price premium pi, payment delay d in days, and enforcement risk rho. The formal share s is logistic in the net advantage:
s(\tau) = \frac{1}{1 + \exp\!\left[-\left(a - b\,(\tau + \pi) - c\,d + \rho\right)\right]}Revenue from the levy is R = tau times price times volume times s(tau). The revenue-maximising levy solves:
\frac{dR}{d\tau} = 0 \;\Rightarrow\; \tau^{*} = \frac{s(\tau^{*})}{-\,s'(\tau^{*})} = \frac{1}{b\,\left(1 - s(\tau^{*})\right)}The policy content is in the parameters, not the rate. The revenue-maximising levy is the point where the extra revenue from a slightly higher rate on the sellers who stay is exactly offset by the revenue lost on the sellers who leave. It depends on how many sellers are already formal and on how sensitive they are. Raising a (trust in GoldBod), cutting d (same-day payment), holding pi at or below zero (GoldBod pays at least the informal price) and raising rho (border and refinery enforcement) all raise the share of sellers who stay at any rate and so move tau-star up. A levy set before those parameters improve fails, as 2020 showed; the same levy set after they improve works.
Two refinements matter in practice, and both are worked in Part II. First, the state's objective is not levy revenue alone. An ounce that stays in the formal channel also yields income tax on the seller, foreign exchange, reserves and the integrity of the supply chain; an ounce that leaves yields nothing and often finances something worse. That wider value, nu per unit, belongs in the objective and pulls the optimal levy down. Second, the state may care about formal share for its own sake, as a floor below which it will not go regardless of revenue. With a floor of 75%, the highest admissible levy on assumed parameters is about 1.25%; with a floor of 70%, about 2%. The choice between them is a choice about how much gold the state will let leave for an extra US$50 million, and the model's job is to make that trade explicit rather than to hide it in a rate.
3.2 Model B: large-scale fiscal regime and investment response
The large-scale model is the fiscal regime simulator of the series overview, section 4, applied to Ghana's review. A mine is described by output, all-in cost, capital and life from its technical report; a regime by its royalty function, corporate rate, depreciation rule, state equity and any rent tax. The simulator runs the mine's cash flows across a distribution of prices and reports two numbers per regime: the state's take, meaning the present value of everything it receives divided by the present value of the mine's pre-tax cash flow, and the investor's post-tax internal rate of return. For a mine with price p, output q, all-in cost c, royalty r on revenue (sliding scale r(p)), corporate rate t, and state equity e, government take G and investor cash flow V over life T at discount rate delta are:
G = \sum_{t} \frac{r(p_t)\,p_t q_t + t\,\Pi_t + e\,(1-t)\,\Pi_t}{(1+\delta)^{t}}, \qquad \Pi_t = p_t q_t (1 - r(p_t)) - c_t q_t - K_tThe design objective is not to maximise take. It is to maximise expected take across the price distribution subject to the constraint that the marginal project, the highest-cost project Ghana hopes to attract, still gets built: its post-tax internal rate of return must stay above the investor's hurdle at a price near the low end of the distribution. If the constraint fails, the project is not built and take on it is zero, which no rate can fix. The design objective is to maximise expected take across a price distribution subject to an investment constraint: the marginal project's post-tax internal rate of return must stay above the hurdle h, or the project is not built and take is zero:
\max_{r(\cdot),\,t,\,e}\; \mathbb{E}_{p}[G] \quad \text{s.t.} \quad IRR^{post}(\text{marginal project}) \ge h, \qquad \text{take share} \le \bar{g}The resource rent tax satisfies this constraint by construction, because it taxes only returns above the hurdle and nothing below: at low prices it is zero, at high prices it takes a share of the excess. A sliding royalty approximates it only if the scale is set on margin, not price. The comparison in Part II shows the difference on a representative 350,000-ounce mine. At US$4,500 the rent tax at 25% above a 15% return lifts take from 0.41 to 0.51 while leaving the investor at 27%; the price-linked sliding scale lifts take to 0.46 but also lifts it at US$2,500, where the project earns only 9% and would not be built under any regime. The review should choose its instrument on that comparison, run on the technical reports of the mines it actually hopes to attract, at the 25th-percentile price it expects rather than at spot.
3.3 Model C: indirect-method production reconstruction
The reconstruction model gives the GRA the independent evidence that section 2.6 says it lacks. Gold cannot be produced without inputs, and the inputs leave records with other parties. Cyanide is imported by a few distributors under customs entries the GRA holds. Electricity is bought from the grid or generated from imported fuel, both of which are metered by third parties. Doré is weighed and assayed on receipt at a refinery. Open-pit mines move rock, and the volume moved is measured by satellite stereo imagery to within a few percent. Each of these gives an estimate of ounces produced that the mine did not write. Declared gold output is tested against physical records held by parties other than the miner: reagent purchases, grid electricity, doré weights at the refinery, tailings volume, and satellite-measured pit and dump growth. For mine i:
\hat{Q}^{(1)}_i = \frac{\text{NaCN}_i}{\nu_s\, g_i^{-1}}, \quad \hat{Q}^{(2)}_i = \frac{E_i}{\eta_s}\, g_i\, \gamma_i, \quad \hat{Q}^{(3)}_i = W^{dor\acute{e}}_i \cdot f_i, \quad \hat{Q}^{(4)}_i = \Delta V^{pit}_i\, \rho\, g_i\, \gamma_iNaCN is cyanide consumption with sector consumption per tonne nu, E electricity with intensity eta, g head grade, gamma recovery, W doré weight with fineness f, and delta-V the pit volume change from satellite stereo imagery with density rho. The coefficients nu and eta are not constants; they vary by ore type, and the model estimates them by ore type from the mines whose declarations pass every identity check, so that a mine is compared with its peers rather than with a textbook figure. Each estimate carries noise sigma, measured as the dispersion of the estimate around the declaration across compliant mines; a source that is noisy for a given ore type gets a lower weight. The residual score and hard flags follow the series' standard form:
A_i = \sum_m \omega_m \max\!\left(0, \frac{\ln \hat{Q}^{(m)}_i - \ln Q^{dec}_i}{\sigma_{m,s}}\right) + \delta H_iA hard flag fires when an accounting identity, rather than a statistical benchmark, is breached: declared output exceeds what the electricity and reagent records could physically have processed, or refinery doré receipts exceed declared production for the period after allowing for opening stock. A breach is not a probability; it is a query the mine must answer. The worked example in Part II shows the difference. A declared 180,000 ounces sits below what throughput, cyanide and power imply, which ranks the mine for audit and gives the auditor a specific question about 400,000 tonnes of ore; it matches refinery receipts, so no identity is breached and no automatic assessment follows. The distinction between ranking and flagging is what keeps the model defensible: the ranking is a judgement the auditor exercises, the flag is a fact the taxpayer explains.
3.4 Model D: GoldBod transaction analytics
The last model turns GoldBod's register into a map of who is actually mining. Small-scale licences are held by individuals and cooperatives, and the sellers at the counter are often nominees: a single financier, sometimes foreign, sometimes a local aggregator, sponsors dozens of pits and sells through dozens of names. A levy applied per seller therefore falls on the digger, who is the least able to pay and the most sensitive to a discount, while the financier who takes the margin is invisible. Resolving the sellers into operators is the precondition for a levy that spares the digger and reaches the concession holder.
The method is entity resolution followed by graph analysis. Sellers who share a phone number, a bank account, a vehicle registration or a delivery pattern (arriving within an hour of one another at the same counter, week after week) are linked; probabilistic record linkage handles variant spellings and missing identifiers; community detection on the resulting graph finds the clusters that behave as single operators. Each cluster's aggregate volume is the base for a graduated levy: a low rate on small clusters, higher on large ones. The same register supports assay anomaly detection (a district whose sellers' fineness is systematically high may be laundering gold from elsewhere), geographic plausibility checks (a seller whose volume exceeds what its licensed area's satellite footprint could yield), and the remediation levy of section 2.5. With every ASM purchase on record, the seller population can be analysed for aggregation, grade anomalies and geographic implausibility, and the resolved operators are the base for a presumptive or graduated levy that spares the artisanal digger and reaches the concession holder.
Algorithms and illustrative results
4.1 Algorithm: setting the ASM levy
INPUT GoldBod purchases (date, seller, weight, assay, price, payment time),
informal price proxies (Lomé, Dubai spot less transport), border seizures
ESTIMATE a, b, c, rho by fitting formal share to monthly variation in premium and delay
FOR tau in 0..8%:
s <- s(tau); R <- tau * P * Q * s
SET tau_star <- argmax R; publish it with the parameter estimates
MONTHLY: re-estimate; if s drops below target, cut delay or raise price before cutting tau4.2 Illustrative results
Parameters are assumed: a = 1.6, b = 0.35 per percentage point, c = 0.08 per day, delay 3 days, informal premium 2%. Formal share at zero levy is 0.66.
| Levy | Formal share | Revenue index | Approximate revenue on 2025 ASM base (US$ m) |
|---|---|---|---|
| 1.0% | 0.58 | 0.58 | 95 |
| 1.5% | 0.53 | 0.80 | 131 |
| 3.0% | 0.40 | 1.21 | 198 |
| 5.0% | 0.25 | 1.26 | 205 |
| 8.0% | 0.11 | 0.84 | 137 |
Revenue peaks near 4% but at that rate two-thirds of gold has left the formal channel, with the foreign-exchange and integrity losses that implies. If GoldBod eliminates the price discount and pays same day, the same model gives a formal share above 0.8 at a 2% levy, and revenue near US$170 million with most gold still on the books. The right objective is revenue subject to a floor on the formal share, not revenue alone.
Large-scale. For a 400,000-ounce mine at US$4,000 with all-in cost of US$1,500, moving from a 5% royalty, 35% corporate rate and 10% free carry to a 7% average sliding royalty, 35% rate and 15% carry raises the present value of government take by about a fifth while keeping post-tax IRR above 20%. A rent tax at 25% above a 15% return does the same with less distortion. Figures are illustrative.
4.4 Reading the results
The levy table in 4.3 is a Laffer curve with a twist. Levy revenue rises with the rate until about 4% and then falls, which is the familiar shape. The twist is the formal-share column: at the revenue-maximising rate, only a quarter to a third of small-scale gold is still passing through GoldBod. The other two-thirds has gone to Lomé and Dubai, taking with it the foreign exchange that the cedi's 2025 recovery depended on, the income tax that could be assessed on identified sellers, and the supply-chain integrity that European and American refiners now demand. Maximising levy revenue would destroy the thing that makes the levy collectable.
That is why the paper's objective is revenue subject to a floor on formal share, and why the buyer's terms come before the rate. On the same parameters, eliminating the 2% price discount and paying same-day moves the formal share at a 2% levy from about 0.5 to above 0.7 and the levy yield to about US$150 million with most gold still on the books. The 2020 episode, read through the model, was a levy set at 3% while the discount and delay were both large; the share fell, the revenue disappointed, and the rate was cut. The rate was never the problem.
The large-scale paragraph in 4.3 makes a different point. On a representative 400,000-ounce mine, moving from Ghana's 2006 terms to a higher sliding royalty and larger carry raises the state's present-value take by about a fifth while leaving the investor above 20%; a rent tax achieves the same take with less distortion at the bottom of the cycle. Part II runs the comparison on a more conservative 350,000-ounce mine with heavier capital and shows the same ordering: the rent tax takes most at high prices and nothing extra at low ones, the price-linked scale takes more at both. The review's choice of instrument matters more than its choice of rate.
4.5 What the illustrative numbers do not show
The channel-choice parameters are assumed, and Ghana is the one country in this series where they need not be. GoldBod's weekly purchase series since May 2025, its published buying price against the LBMA reference, its payment-lag data, and the observable spreads in Lomé and Dubai identify a, b, c and rho within a few quarters of data; Part II's algorithm D.1 sets out the state-space estimation. Until that is done, the table shows the shape of the answer, not its level. The base of US$10.8 billion is 2025's export value; at a lower gold price the same tonnage yields less, and at a lower formal share the base itself is smaller.
The large-scale numbers assume a single price and a representative cost. Ahafo, Tarkwa and Obuasi differ in cost, remaining life and stability status; the review's revenue depends on which of them are unshielded and when, and the mine-by-mine runs proposed in section 8.3 are the only way to know. The 25th-percentile price at which the review should be judged is also a choice; the paper suggests it because a regime that only works at spot is a regime that fails in the next downturn.
Solutions, software, data science and AI models
| Issue | Solution | Data science or AI model | Software component |
|---|---|---|---|
| Untaxed ASM | Low levy withheld at GoldBod counter, conditional on same-day payment and market price | Channel-choice logit; monthly Laffer re-estimation | Levy module inside the GoldBod purchase system |
| Aggregators behind small sellers | Graduated levy by resolved operator | Entity resolution (Splink) and graph community detection on sellers, licences, phones, accounts | Operator graph and register |
| Smuggling | Border and refinery enforcement targeted by flow model | Mirror-data analysis of Togo, UAE and Swiss gold imports against Ghana exports; anomaly detection | Trade-mirror dashboard shared with customs |
| Large-scale fiscal review | Rent-based instrument; margin-linked sliding scale | Project cash-flow simulator under price distributions; investment-response constraint | Fiscal regime simulator for the Minerals Commission and Finance |
| Grade and recovery under-declaration | Indirect-method audit | Reagent, power and satellite reconstruction; MCD anomaly scoring | Mining audit workbench (ONYX-class physical-balance engine) |
| Environmental cost recovery | Remediation levy by measured footprint | Satellite change detection (Sentinel-2 imagery, random-forest land-cover classifier) | Geospatial monitoring service shared with EPA |
| Audit capacity | Explainable, label-free selection | Custodian reconstruction; Bayesian reweighting as audits close | Risk engine and case management |
Software stack. GoldBod purchase system and GRA's tax system as sources; Python with Airflow; PostgreSQL bitemporal warehouse; scikit-learn, PyOD, LightGBM, PyMC; OR-Tools; Splink and Neo4j Community; Sentinel-2 via open Earth-observation APIs with rasterio and scikit-learn; Superset dashboards; Docker and MLflow; hosted in a government or sovereign cloud region.
Governance of AI. Levy parameters are published monthly so miners can see why the rate is what it is. Scores select audits; officers assess. Satellite evidence supports inspection, never assessment on its own. Seller-level data stays inside GoldBod and GRA under a data-sharing agreement.
Team. A joint GoldBod and GRA analytics unit of eight, with 12 to 18 months of external support. Against an ASM export base above US$10 billion, the unit pays for itself at a levy yield of a few basis points.
5.7 The solution described as a system
GoldBod's purchase system is the foundation. It already exists. Every purchase is recorded with seller, weight, assay, price and payment time, and the record is signed by the state buyer. What the paper adds is a levy module that computes the withholding at the counter from a published rate and a published set of parameters, and a data-sharing agreement with the Ghana Revenue Authority under which the seller-level record reaches the GRA weekly. The levy module is small; the agreement is the work, because GoldBod's data was collected for foreign-exchange and integrity purposes and its use for tax needs a legal basis and a governance rule.
The operator register is built from that record. Entity resolution on names, identity numbers, phones, bank accounts, vehicles and delivery patterns, followed by community detection, resolves the sellers into operators. The register is the base for three things: a graduated levy on operator volume, so that a financier who sells through forty names pays as a financier; assay anomaly detection by district, which flags gold that may not be Ghanaian; and the link to satellite footprint that section 2.5 describes. It is rebuilt monthly and reviewed by a small team, because record linkage produces false matches that a person must check.
The levy calibration runs monthly and is published. The channel-choice model is re-estimated on the latest purchase data, the constrained levy is recomputed, and the rate with its parameters and intervals is published. Publication is not a courtesy: sellers who can see why the rate is what it is, and who can see that it falls when the formal share falls, have a reason to stay. A rate set in a ministry and announced by gazette has the opposite effect, as 2020 showed.
The fiscal regime simulator serves the law review. It encodes the current Act and each proposal, loads the technical reports of the large mines and of the projects Ghana hopes to attract, and runs take and investor return across a price distribution mine by mine. Its output to the Minerals Commission and the Ministry of Finance is a table like Part II's B, for real mines: which are shielded and until when, what each proposal yields on the unshielded base, and at what price each proposal stops the marginal project. It is the same code as the Mali paper's simulator; Ghana's version needs Ghana's regimes and mines.
The mining audit workbench tests the large mines. It links three data sets the GRA already holds or can obtain: cyanide and reagent imports by distributor, matched to mine deliveries; power by mine from ECG and GRIDCo, with self-generation fuel from customs; and refinery receipts by shipment from the Royal Ghana Gold Refinery and export refiners. It adds Sentinel-2 volumetrics for the open pits. The reconstruction, scoring and audit allocation of Part II's algorithm D.4 run on these annually; the workbench presents each mine's decomposition to the auditor in a form the mine can be shown.
The geospatial service serves both revenue and environment. Sentinel-2 change detection, run monthly with a land-cover classifier trained on known galamsey sites, produces polygons of new excavation with dates. Matched to the operator register, it gives footprint per operator; matched to the EPA's and Forestry Commission's data, it gives a shared operational picture that none of the agencies has today. The remediation levy of section 8.6 is priced from it.
The trade-mirror dashboard measures what escapes. Monthly imports of gold by origin into the UAE, Switzerland, India and Türkiye, from UN Comtrade and national customs, are compared with Ghana's declared exports and GoldBod's purchases. The residual bounds the smuggling that remains, and its trend tells the GRA and GoldBod whether the levy and the buyer's terms are working.
5.8 Cost, staffing and sequencing of the build
The joint GoldBod and GRA analytics unit is eight people: a head, two data engineers, two analysts, a geospatial specialist, an entity-resolution reviewer and a liaison to the GRA's mining audit unit. External support of 12 to 18 months covers the levy module, the register, the simulator, the workbench and the geospatial pipeline. The software is open source; the cost is people, data agreements and satellite processing, and it is small against a small-scale export base above US$10 billion: the unit pays for itself at a levy yield of a few basis points, or at a single recovered under-declaration at a large mine.
Sequencing follows the reform table in section 6: the data-sharing agreement and the levy module first, because the levy cannot start without them; the register in parallel, because it needs only GoldBod's data; the simulator early, because the law review is under way; the workbench and geospatial service through 2027; the mirror dashboard whenever Comtrade data allows, which is with a lag of several months.
Reform sequence
| Phase | Period | Measure | Model |
|---|---|---|---|
| 1 | Q4 2026 | Same-day payment and market-price guarantee at GoldBod; publish purchase statistics | A |
| 2 | 2027 H1 | Levy at 1.5% withheld at the counter; monthly Laffer report | A, D |
| 3 | 2027 | Mining-law review adopts margin-linked royalty or rent tax; stability agreements respected | B |
| 4 | 2027 to 2028 | Indirect-method audit of the ten largest mines; satellite monitoring live | C |
| 5 | 2028 | Graduated levy on resolved operators; remediation levy by footprint | D |
6.1 The phases explained
Phase 1, the buyer's terms (Q4 2026). Before any levy, GoldBod guarantees same-day payment at every counter and a buying price at or above the reference price the informal market pays, and it publishes its purchase statistics weekly. This is the precondition the model identifies and the step 2020 skipped. Deliverable: the payment guarantee, the price rule and the published series. Owner: GoldBod. Risk: cash positioning fails at peak seasons; mitigated by the forecasting model of Part II and by pre-positioned liquidity.
Phase 2, the levy (2027 H1). A levy of 1.5% is withheld at the counter, with the monthly calibration published. The rate is chosen from the model's constrained optimum at a formal-share floor of 0.70 and is expected to yield of the order of US$120 to 150 million a year on 2025 volumes if the formal share holds; if it falls, the published rule cuts the discount and the delay before it touches the rate. Deliverable: the levy module, the data-sharing agreement, the first monthly Laffer report. Owner: GoldBod and the GRA. Risk: the formal share falls faster than the model predicts; the pre-committed response rule limits the damage, and the monthly re-estimation catches it within a quarter.
Phase 3, the law review (2027). The review adopts a margin-linked royalty or a resource rent tax for new projects and renewals, encoded and tested on the simulator, and leaves stabilised development agreements in place until expiry. Deliverable: the amended Act and the published simulator runs that justified its parameters. Owner: the Minerals Commission and the Ministry of Finance with Parliament. Risk: a price-linked scale is adopted because it is simpler; the simulator's runs at the 25th-percentile price are the argument against it.
Phase 4, the audit workbench (2027 to 2028). The reconstruction runs on the ten largest mines with the data sets named in 5.7; the satellite service goes live for the open pits. Deliverable: audit rankings and any hard flags for the 2027 returns; a benchmark of reagent, power and recovery coefficients by ore type. Owner: the GRA's mining audit unit. Risk: refiners and utilities resist data sharing; the statutory provision in the amended Act is the remedy.
Phase 5, graduated and remediation levies (2028). With the operator register mature, the counter levy becomes graduated by operator volume, and a remediation levy priced per hectare disturbed is collected through the same counter. Deliverable: the graduated schedule and the footprint-based remediation charge. Owner: GoldBod, the GRA and the EPA. Risk: operators fragment to stay below tiers; the register's monthly re-run flags sudden fragmentation.
6.2 What can go wrong in the sequence
The sequence fails at Phase 1 if GoldBod's cash positioning fails, which is an operational problem with an operational fix. It fails at Phase 2 if the levy is set before Phase 1 is complete, which is a political temptation because the revenue is visible and the discount is not. It fails at Phase 3 if the review is judged at spot prices, which makes every regime look affordable. It fails at Phase 4 if the data agreements are not in the Act. The paper's ordering is designed so that each phase's data makes the next phase possible, and so that the visible revenue, the levy, arrives only after the invisible precondition, the buyer's terms, is in place.
Limitations
7.1 The limitations discussed
Assumed channel-choice parameters. Every number in the levy table depends on a, b, c and nu, and all four are assumed. Ghana holds the data to estimate them, and Part II sets out how; until it does, the table's shape is reliable and its levels are not. The paper's recommendation of 1.5% at a 0.70 floor is a recommendation to start the estimation at a rate the model says is safe, not a claim that 1.5% is optimal.
Noisy informal price proxies. The informal price is inferred from Lomé and Dubai spreads, which move with enforcement in Togo and the UAE as well as with conditions in Ghana. A period in which Togo tightens its own border will look, in Ghana's data, like a fall in the informal premium. The state-space treatment in Part II, which infers the premium from purchase volumes and seizures jointly, is designed to reduce that error but not to remove it.
Stability agreements. The review's revenue is computed on the unshielded base, and the paper does not know that base precisely, because the development agreements' expiry dates and renewal terms are not all public. The simulator must model renewals and new projects separately from stabilised production, and its national revenue figure will be wrong by the stabilised share until the register of agreements is complete.
Open pits only. Satellite volumetrics measure surface excavation. Obuasi is underground, and the reconstruction there relies on hoist, power and reagent records rather than imagery; alluvial small-scale operations move little rock per ounce and are poorly measured by any physical method. The reconstruction model is strongest where Ghana's largest mines are, in open pits, and weakest where its most destructive mining is, in rivers.
The register's false matches. Entity resolution links records probabilistically, and a graduated levy applied to a wrongly merged cluster falls on sellers who are not the operator. The review step in the register's monthly rebuild is essential, and the levy tiers should be wide enough that a single false link does not move a seller across one.
Political economy. The 2020 levy was cut within a year under pressure from the small-scale sector, which is politically organised and electorally significant in the Western and Ashanti regions. A levy that the sector experiences as a discount will meet the same response. The paper's insistence on the buyer's terms first is as much a political design as an economic one: a levy that arrives with same-day payment and a market price is a different proposition from one that arrives alone.
Extended modelling: improvements and further research
8.1 Research update, September 2026
GoldBod has been sole regulator and accredited buyer of small-scale gold since 1 May 2025, with a purchase target of at least 3 tonnes a week. Burkina Faso's SONASP runs a parallel model and lifted recorded artisanal output from 8.1 tonnes in 2024 to about 42 tonnes in 2025, which is direct evidence that a state buyer paying promptly at market prices can pull large volumes into the formal channel. Gold prices fell from a January 2026 peak above US$5,000 to the low US$4,000s by mid-year; GoldBod based its 2026 plan on about US$5,000. Both facts sharpen the paper's argument: the formalisation effect is real and large, and the levy design must survive a price decline, when the informal premium and miner sensitivity both rise.
8.2 Model A, channel choice: improvements
8.3 Model B, large-scale fiscal regime: improvements
8.4 Model C, indirect-method audit: improvements
8.5 Model D, GoldBod transaction analytics: improvements
8.6 New models proposed
8.7 Further research
Sources for this section: Ecofin Agency, reforms lift artisanal gold's share in Ghana and Burkina Faso, 5 February 2026; Reuters via Mining Weekly, Ghana's 2026 artisanal output, 15 July 2026.
A. Worked example 1: the ASM levy with a formal-share floor
The channel-choice model of section 3.1 is run here with the objective of section 3 of the series overview: levy revenue plus the wider fiscal value of formal gold, subject to a floor on formal share. Parameters are assumed for illustration: a = 1.6, b = 0.35 per percentage point of effective discount, c = 0.08 per day of payment delay, nu = 2% of the value of gold that leaves the formal channel (lost income tax, foreign exchange and integrity), base Q = US$10.8 billion (2025 ASM exports at 100% formality).
| Levy | Price gap to world | Payment delay (days) | Formal share | Levy revenue (US$ m) | Value lost on informal gold (US$ m) | Net (US$ m) |
|---|---|---|---|---|---|---|
| 0% | 2% | 3 | 0.66 | 0 | 74 | -74 |
| 1.5% | 2% | 3 | 0.53 | 86 | 101 | -14 |
| 3% | 2% | 3 | 0.40 | 131 | 129 | 2 |
| 5% | 2% | 3 | 0.25 | 136 | 162 | -26 |
| 1.5% | 0% | 0 | 0.75 | 121 | 55 | 66 |
| 2% | 0% | 0 | 0.71 | 154 | 62 | 91 |
| 3% | 0% | 0 | 0.63 | 205 | 79 | 126 |
Three readings. First, with a 2% price discount and three-day payment, no levy produces a positive net result; the buyer's terms are the binding problem, not the rate. Second, once the discount is zero and payment is same-day, every levy up to 3% is net positive. Third, the floor decides the rate: with s-bar = 0.75, the highest admissible levy is about 1.25%, yielding about US$103 million; with s-bar = 0.70, about 2%, yielding about US$154 million. The choice between them is a choice about how much gold the state is willing to see leave the formal channel for an extra US$50 million, which is a political decision the model makes explicit.
B. Worked example 2: the large-scale fiscal review
A representative new mine: 350,000 ounces a year, all-in sustaining cost US$1,500, capital US$1,500 million over three years, 12-year life, tax depreciation straight-line, 8% discount rate. Four regimes are compared at three prices. Take is the present value of government receipts over the present value of pre-tax cash flow; IRR is the investor's post-tax return.
| Price (US$/oz) | Current: 5% royalty, 35% CIT, 10% carry | Sliding 5 to 12% royalty by price | Margin band 5 to 10% | 5% royalty plus 25% rent tax above 15% |
|---|---|---|---|---|
| 2,500 | Take 0.34, IRR 9% | Take 0.36, IRR 9% | Take 0.36, IRR 9% | Take 0.34, IRR 9% |
| 3,500 | Take 0.39, IRR 21% | Take 0.43, IRR 19% | Take 0.43, IRR 19% | Take 0.46, IRR 19% |
| 4,500 | Take 0.41, IRR 30% | Take 0.46, IRR 27% | Take 0.45, IRR 28% | Take 0.51, IRR 27% |
The rent tax raises take most at high prices (0.51 against 0.41) while leaving the investor's IRR at 27%, and at US$2,500 it takes nothing extra because the project earns below the 15% hurdle. The price-linked sliding scale reaches 0.46 at US$4,500 but also raises take at US$2,500, where the project barely clears 9%; that is the design flaw. The margin band behaves like the sliding scale at this cost level because price and margin move together for a single mine; the difference appears across mines with different costs, which the mine-by-mine runs in section 8.3 would show.
Investment constraint. At US$2,500 no regime keeps IRR above 15%, so a new 350,000-ounce mine at US$1,500 cost would not be built at that price under any of them. The review should therefore be judged on the 25th-percentile price of the distribution it expects, not the current spot.
C. Worked example 3: indirect-method reconstruction of a declared 180,000 ounces
A mine declares 180,000 ounces from 4.5 million tonnes at 1.4 g/t and 92% recovery. Third-party records: cyanide purchases 5,200 tonnes (sector norm 1.1 kg per tonne of ore for this ore type), power 520 GWh (norm 105 kWh per tonne), refinery doré receipts 6,800 kg at 82% fineness. Each estimator's noise sigma is taken as 0.08 in log terms.
| Source | Implied ore tonnes | Implied ounces | z-score against 180,000 |
|---|---|---|---|
| Declared throughput, grade, recovery | 4,500,000 | 186,346 | 0.43 |
| Cyanide | 4,727,000 | 195,757 | 1.05 |
| Power | 4,952,000 | 205,079 | 1.63 |
| Refinery doré | not applicable | 179,272 | -0.05 |
The declared ounces are below what the mine's own throughput implies (186,000), and well below what reagents and power imply (196,000 to 205,000), but match refinery receipts. Two explanations compete: the mine processed more ore than declared and the extra gold went somewhere other than the refinery (a compliance problem), or the mine processed lower-grade ore than the norms assume and its reagent and power intensity is high (a parameter problem). The score A = sum of omega times positive z, with equal weights, is 3.1, which ranks the mine for audit; the audit question is specific: reconcile 4.5 million declared tonnes with power and reagents consistent with about 4.9 million. Note that the doré identity is not breached; a breach (refinery receipts above declared production) would have been a hard flag rather than a ranking.
D. Algorithms in full
D.1 Monthly levy calibration
ALGORITHM GoldBod_Levy_Calibration
INPUT:
purchases[t] weekly GoldBod purchases (oz) since May 2025
P_world[t], P_gb[t] LBMA and GoldBod prices
delay[t] mean days from delivery to payment
tau[t] levy in force
Q_hat[t] latent total ASM output; observations: mirror imports M[t], footprint F[t], seizures Z[t]
s_floor formal-share floor set by policy
nu fiscal value per unit of formal gold beyond the levy
STEP 1 State-space estimation
state: log Q_hat[t] = log Q_hat[t-1] + drift + w[t]
obs 1: log(purchases[t] / (Q_hat[t] - purchases[t])) = a - b*Delta[t] - c*delay[t] + rho*E[t] + e1
obs 2: log M[t] = log Q_hat[t] + kappa_M + e2
obs 3: log F[t] = log Q_hat[t] + kappa_F + e3
fit by MCMC (PyMC); retain posterior draws of a, b, c, rho, Q_hat
STEP 2 Constrained levy
for each posterior draw:
for tau in 0 .. 8 step 0.25:
s = logistic(a - b*(tau + gap_target) - c*delay_target + rho*E)
objective = tau*P*Q_hat*s + nu*Q_hat*s
feasible if s >= s_floor
tau_draw = argmax objective over feasible tau
tau_star = median of tau_draw; report 10th and 90th percentiles
STEP 3 Publish
publish tau_star, s_floor, posterior means of a, b, c, and the current effective discount
if realised s falls below s_floor for two consecutive months: cut delay and gap first; cut tau only if still belowD.2 Operator resolution and graduated levy
ALGORITHM Operator_Resolution
INPUT: seller records (name, ID number, phone, bank account, licence, delivery times, location)
STEP 1 Blocking on phone, account and licence; pairwise comparison with Fellegi-Sunter match weights (Splink)
STEP 2 Threshold match probability 0.9 to form seller clusters; manual review of 0.6 to 0.9
STEP 3 Build graph: nodes = resolved sellers, licences, phones, accounts, vehicles; edges = shared attributes and co-delivery within 1 hour
STEP 4 Leiden community detection; each community = candidate operator
STEP 5 Operator volume V_op = sum of purchases across community per year
STEP 6 Graduated levy: tau(V_op) = tau_1 if V_op < V_1; tau_2 if V_1 <= V_op < V_2; tau_3 above
with tau_1 = counter levy, tau_2 and tau_3 set so that expected revenue per operator class is maximised under class-specific b (mixed logit)
STEP 7 Monthly re-run; operators whose community structure changes sharply are flagged for review (fragmentation to avoid the tier)D.3 Fiscal regime simulator (mine-by-mine)
ALGORITHM Regime_Simulator
INPUT: mine parameters from technical reports (reserves, grade, recovery, throughput, AISC path, capex, life)
regime R = {royalty function, CIT rate, depreciation rule, free carry, purchase option, rent tax}
price paths from two-factor model (N = 10,000)
FOR each mine, each regime, each price path:
simulate annual revenue, costs, capital, depreciation, royalty, CIT, rent tax, dividends
government cash G[t]; investor cash I[t]
take = PV(G) / PV(pre-tax cash flow); IRR = irr(I)
REPORT per mine and regime: median and 10th/90th percentile take; Pr(IRR < hurdle)
AGGREGATE: national take and revenue by year = sum over mines weighted by production
CONSTRAINT CHECK: regime admissible if Pr(IRR < hurdle) <= 0.25 for the marginal projectD.4 Reconstruction and audit allocation
ALGORITHM Reconstruct_And_Allocate
FOR each ore type s: estimate nu_s, eta_s, sigma_m,s from mines passing all identities (median and MAD)
FOR each mine i:
Q_hat[m,i] for m in {throughput, cyanide, power, dore, satellite}
z[m,i] = (ln Q_hat[m,i] - ln Q_dec[i]) / sigma_m,s
H[i] = 1 if dore receipts > declared production OR declared tonnes > power-implied tonnes * 1.15
A[i] = sum_m omega_m * max(0, z[m,i]) + gamma * d_MCD(ratios_i) + delta * H[i]
EY[i] = pi_i * (royalty rate + CIT rate * margin) * P * max(0, weighted Q_hat - Q_dec)
SOLVE: max sum x_i EY[i] s.t. sum h_ik x_i <= H_k for each skill k; sum over random pool x_i >= alpha * sum x_i;
selection rate by region within [0.5, 2] times base rate
OUTPUT: audit list with per-mine decomposition of A
AFTER audit: update omega by Bayesian reweighting on observed Q_auditE. Data specification for Ghana
| Data | Holder | Status | Action |
|---|---|---|---|
| GoldBod purchases by seller, weight, assay, price, payment date | GoldBod | Held since May 2025 | Data-sharing agreement with GRA; weekly extract |
| Large-mine monthly returns: tonnes, grade, recovery, ounces, cost | Minerals Commission, GRA | Held | Standardise categories to technical-report definitions |
| Cyanide and reagent imports by importer | GRA Customs | Held | Link importer sales to mine deliveries |
| Power by mine | ECG, GRIDCo, self-generation logs | Partly held | Statutory access; self-generation fuel via customs |
| Refinery receipts by shipment | Royal Ghana Gold Refinery, export refiners | Partly held | Mandatory shipment reporting |
| Satellite imagery | Sentinel-2 (free), radar (commercial) | Available | Annual volumetrics for ten largest pits |
| Destination imports: UAE, Switzerland, India | UN Comtrade, national customs | Public with lag | Monthly mirror dashboard |
| Company and beneficial ownership | Registrar-General, ORC | Held | Graph build |
F. Sensitivity analysis
| Parameter | Range tested | Effect on levy recommendation | Effect on large-scale result |
|---|---|---|---|
| b (discount sensitivity) | 0.20 to 0.50 | tau-star from 3.0% to 1.0% at s-bar 0.70 | none |
| c (delay sensitivity) | 0.04 to 0.12 | same-day payment worth 0.2 to 0.6 of formal share at 3-day delay | none |
| nu (fiscal value of formal gold) | 1% to 4% | higher nu lowers tau-star by up to 1 point | none |
| Price | US$2,500 to 4,500 | levy yield scales linearly; share roughly unchanged if gap is zero | take 0.34 to 0.51 depending on regime |
| AISC | US$1,200 to 1,800 | none | at US$1,800 no regime clears 15% below US$3,200 |
| Hurdle rate | 12% to 18% | none | rent-tax threshold moves; take at US$4,500 from 0.53 to 0.49 |