Introduction

Vanuatu has been a zero-direct-tax jurisdiction since 1971. That choice was affordable while tourism, aid and, after 2015, passport sales grew. It stopped being affordable in 2024. Air Vanuatu entered liquidation in May 2024, a major earthquake struck Port Vila in December 2024, and visa-free access to the Schengen area and the United Kingdom was withdrawn from Vanuatu passport holders, cutting the value of the citizenship product.

The research question is practical. Given that the state must replace a declining non-tax stream, which taxes should replace it, at what parameters, and in what order, so that revenue rises, volatility falls, and the burden lands on those able to pay? The paper answers with three models that the Department of Customs and Inland Revenue and the Department of Finance and Treasury can run with data they already hold.

The paper uses one organising identity throughout. For any tax k, the distance between potential and actual revenue splits into a policy gap and a compliance gap:

R^{pot}_{k} - R_{k} = \underbrace{\left(R^{pot}_{k} - R^{law}_{k}\right)}_{\text{policy gap}} + \underbrace{\left(R^{law}_{k} - R_{k}\right)}_{\text{compliance gap}}

Vanuatu is unusual in that its largest gap is a pure policy gap: a tax that does not exist.

Current tax issues

Vanuatu's core problem is that its second-largest revenue line is a discretionary foreign purchase, not a tax. The IMF's 2025 consultation found VAT holding up on better collection while other streams weakened, and described Economic Citizenship Program (ECP) revenue as significant but in structural decline.

2.1 No income tax base

There is no personal income tax, corporate tax, capital gains tax or withholding tax. The only charge resembling a direct tax is a 12.5% tax on rental income above VT 200,000 per half-year. The 2025 Article IV report records the authorities' intention to broaden the base by introducing personal and corporate income tax. The 2026 Budget commissions a holistic review covering direct, indirect, non-tax and administration reform, but enacts no direct tax. The system is therefore regressive by construction: households pay 15% on consumption, while profits, high salaries and capital gains pay nothing.

2.2 Citizenship revenue: volatile and hard to verify

ECP receipts have run near 7% of GDP. The IMF noted that the 2025 rebound followed a switch to invoicing through the government financial management system, and that it remains to be verified whether cash received matches invoices issued. A Commission of Inquiry report on the programme was under ministerial review at the time of the consultation. Demand depends on decisions made in Brussels and London, not Port Vila.

2.3 VAT leakage

VAT at 15% carries the system. The government issued the Vanuatu Sales Monitoring System (VSMS) Regulation Order No. 153 of 2025 and plans full rollout in 2026, aimed at under-reported sales and the grey economy. The VT 4 million registration threshold leaves a large informal retail tail outside the net, and exempt financial services and residential rents break the credit chain.

2.4 International standing

The EU Council reconfirmed Vanuatu on its list of non-cooperative tax jurisdictions on 17 February 2026, with the next review in October 2026. The 20-year exemption for international companies is the visible irritant. Listing raises correspondent-banking costs, which feeds back into trade and therefore VAT. A properly designed corporate tax with substance rules would address the listing criteria and the revenue gap with one instrument.

2.5 Fiscal drift

Current spending has averaged about 35% of GDP since 2020 against 27% in the previous five years, while capital spending fell from about 10% to 6%. New revenue that funds wages will not rebuild buffers, so revenue reform needs a companion expenditure rule.

2.6 Administrative capacity

The Department of Customs and Inland Revenue has never administered an assessment-based tax. It has no income tax return, no PAYE system, no audit method for profit and few staff with accounting qualifications. Any model proposed here must run on data the state already collects: VAT returns, customs entries, Vanuatu National Provident Fund (VNPF) contribution records, business licences and, from 2026, VSMS transaction feeds.

Models

3.1 Model A: revenue as a portfolio

Total revenue is a weighted sum of streams, and its volatility depends on each stream's own volatility and on how the streams move together. With share s-k, coefficient of variation v-k and correlation rho between streams:

CV_R = \sqrt{\sum_{k}\sum_{l} s_k\, s_l\, v_k\, v_l\, \rho_{kl}}

Each stream's contribution to total risk is its marginal share of that variance:

RC_k = \frac{s_k v_k \sum_l s_l v_l \rho_{kl}}{CV_R^{2}}, \qquad \sum_k RC_k = 1

A stream with RC-k well above its revenue share s-k is supplying more risk than money. ECP is the clear case. The design problem is then a mean-variance choice of tax mix, subject to a revenue target, a regressivity cap and an administrative-capacity cap on how much new tax can be absorbed per year:

\min_{s}\; CV_R(s) \quad \text{s.t.} \quad \sum_k s_k \bar{R} \ge R^{target}, \quad K(s) \le K^{*}, \quad \left| s_{k,t} - s_{k,t-1} \right| \le \Delta_k

K(s) is the Kakwani index of the mix, so the constraint stops the optimiser from loading everything onto VAT, which is the lowest-variance but most regressive stream.

3.2 Model B: income-tax microsimulation

Model B computes revenue, number of payers and distribution for any candidate personal income tax (PIT) schedule. The population is VNPF contribution records, which give monthly wages for every formal employee, reweighted to the latest household survey for non-wage income. For individual i with annual income y-i, threshold b-0 and brackets j with rates t-j:

T_i = \sum_{j=1}^{J} t_j \cdot \max\!\left(0,\; \min\!\left(y_i, b_j\right) - b_{j-1}\right), \qquad R^{PIT} = \sum_i w_i\, c_i\, T_i\!\left(y_i \left(1 - e\,\Delta\!\ln(1-t^{m}_i)\right)\right)

w-i is the sampling weight, c-i the compliance ratio (near 1 for PAYE wages, 0.3 to 0.6 for self-employment), and e the elasticity of taxable income applied to the change in the net-of-marginal-tax rate. Progressivity is measured by the Kakwani index, the concentration of tax minus the Gini of pre-tax income:

K = C_T - G_Y, \qquad C_T = 1 - 2\int_0^1 L_T(p)\, dp

The threshold is chosen, not assumed. Raising b-0 loses revenue but removes low-value payers whose administration costs more than they pay. The optimal threshold sets the marginal revenue lost equal to the marginal administrative and compliance cost saved:

b_0^{*} = \arg\max_{b_0} \left[ R^{PIT}(b_0) - \left(a + \kappa\right) N(b_0) \right]

a is the annual administration cost per payer, kappa the payer's compliance cost, and N the number of payers above the threshold.

Corporate tax is modelled from the VAT returns the department already holds. Value added declared for VAT, less VNPF-recorded wages, approximates gross operating surplus for firm f:

\hat{\Pi}_f = \left(S_f - P_f\right) - \frac{W_f}{0.12} - \delta K_f - I_f, \qquad R^{CIT} = t_c \sum_f c_f \max\!\left(0, \hat{\Pi}_f - L_f\right)

S is sales and P purchases from VAT returns, W is VNPF contributions, grossed up to wages at the 12% combined contribution rate, delta-K depreciation, I interest and L carried-forward losses. This gives a firm-level revenue estimate before a single income tax return exists. It also gives the audit benchmark on day one, because a firm's first corporate return can be tested against its own VAT history.

3.3 Model C: VAT gap and VSMS-based audit selection

C-efficiency needs three national numbers and tells the government how far VAT is from its ceiling:

CE = \frac{V}{\tau \cdot (C - V)}, \qquad \text{Compliance gap} = \frac{VTTL - V}{VTTL}

V is net VAT, tau the 15% rate, C final consumption, and VTTL the theoretical liability built from household, government and exempt-sector consumption by commodity. Audit selection cannot rely on supervised learning, because Vanuatu has too few completed audits to train a classifier. Model C is label-free: it rebuilds each trader's turnover from records held by third parties and scores the shortfall. For trader i:

\hat{Q}^{(1)}_i = M_i(1+\mu_s), \quad \hat{Q}^{(2)}_i = S^{VSMS}_i, \quad \hat{Q}^{(3)}_i = \frac{W_i}{\lambda_s}, \quad \hat{Q}^{(4)}_i = B_i

These are customs imports with a sector mark-up, VSMS recorded sales, VNPF wages over the sector labour share, and bank credits where available. The anomaly score weights each positive shortfall by source reliability omega and adds a hard flag H for breaches of accounting identities, such as sales below imports for a trader with no stock build:

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_i

Audits are then allocated by integer programme to maximise expected recovery within available auditor hours H-k, with a random-audit share alpha held back to keep the gap estimate unbiased:

\max_{x \in \{0,1\}} \sum_i x_i\, \pi_i\, \tau \max\!\left(0, \textstyle\sum_m \omega_m \hat{Q}^{(m)}_i - Q^{dec}_i\right) \quad \text{s.t.} \quad \sum_i h_{ik} x_i \le H_k, \quad \sum_{i \in rand} x_i \ge \alpha \sum_i x_i

Algorithms and illustrative results

4.1 Algorithm 1: PIT schedule search

INPUT  VNPF wage records y[i], survey weights w[i], compliance c[i],
       elasticity e, admin cost a, compliance cost kappa,
       candidate thresholds B0, candidate rate schedules S
FOR each threshold b0 in B0:
  FOR each schedule s in S:
    FOR each person i:
      mtr[i]  <- marginal rate of s at y[i]
      y_adj   <- y[i] * (1 + e * ln(1 - mtr[i]))      # behavioural response
      T[i]    <- tax(y_adj, b0, s) * c[i]
    R      <- sum(w * T)
    N      <- sum(w * (T > 0))
    K      <- concentration(T, y) - gini(y)
    net    <- R - (a + kappa) * N
    STORE (b0, s, R, N, K, net)
RETURN Pareto frontier over (net revenue, K, N)

4.2 Algorithm 2: label-free VAT audit selection

INPUT  VAT returns, customs entries, VSMS feeds, VNPF records, auditor hours H[k]
FOR each sector s:
  estimate mark-up mu[s], labour share lambda[s] by median of compliant-looking core
  estimate noise sigma[m,s] by median absolute deviation
FOR each trader i:
  build Qhat[m,i] for each available source m
  z[m,i] <- (ln Qhat[m,i] - ln Qdec[i]) / sigma[m,s]
  H[i]   <- 1 if any accounting identity is breached else 0
  A[i]   <- sum_m omega[m] * max(0, z[m,i]) + delta * H[i]
  EY[i]  <- pi[i] * 0.15 * max(0, sum_m omega[m]*Qhat[m,i] - Qdec[i])
SOLVE integer programme: max sum x[i]*EY[i] s.t. hours, random share alpha
ROUTE  hard-flag cases under value cap -> automatic letter
       ranked cases                    -> officer approval
       large or sensitive entities     -> senior referral
AFTER each closed audit: update omega[m] by Bayesian reweighting on observed Qaudit

4.3 Illustrative results

The runs below use synthetic inputs, because VNPF microdata and stream-level revenue series are not public. They show the shape of the answer and must be re-run on official data before any figure is quoted.

Portfolio. Assumed current shares: VAT 40%, ECP 25%, trade taxes 20%, other 15%, with coefficients of variation of 0.08, 0.35, 0.12 and 0.10. Total revenue CV is 0.108, and ECP supplies most of that risk from a quarter of the money. A reform mix that halves the ECP share and adds income taxes at 17% of revenue (CV 0.07) lowers total CV to 0.077, a fall of about 29%.

Personal income tax. Synthetic population of 32,000 formal employees, lognormal wages with median VT 950,000, GDP VT 135 billion, rates of 10%, 20% and 30% with bracket steps at VT 3 million and VT 6 million, full PAYE compliance.

Moving the threshold from VT 1.0 million to VT 1.5 million gives up a third of the revenue but removes 42% of payers. For an administration building PAYE from nothing, that trade is worth making in the first three years.

Corporate tax. Assuming formal corporate operating surplus of 18% of GDP, 45% of it taxable after depreciation, interest and losses, and 60% first-years compliance, yield is about 0.73% of GDP at a 15% rate, 0.97% at 20% and 1.21% at 25%.

Reading. PIT at the VT 1.5 million threshold plus corporate tax at 20% yields about 1.9% of GDP, against ECP receipts near 7%. Income tax is therefore a stabiliser and a fairness measure, not a full replacement. The remaining gap has to come from VAT compliance through VSMS, excise review, and spending restraint.

Solutions, software, data science and AI models

Every issue in section 2 maps to one policy solution, one analytical model and one software component. Nothing here needs proprietary licences or a large data team; the binding constraint is data access agreements, not technology.

5.1 Issue-to-solution map

5.2 Model catalogue

5.3 Reference architecture

flowchart TD
  A[Sources<br/>VAT returns, customs, VSMS, VNPF, licences, FMIS] --> B[Ingestion and data quality<br/>validation, deduplication]
  B --> C[Warehouse<br/>bitemporal, single taxpayer view]
  C --> D[Policy lab<br/>microsimulation, portfolio, VAT gap]
  C --> E[Risk engine<br/>reconstruction, anomaly, graph]
  E --> F[Audit allocator<br/>integer programme]
  F --> G[Case management<br/>auto letter, recommend, refer]
  G --> H[Audit outcomes]
  H --> E
  D --> I[Treasury and Minister dashboards]
  C --> J[Tax-law assistant<br/>retrieval grounded]
  J --> G

Data flows one way from operational systems into a warehouse, two analytical engines read from it, and audit outcomes feed back to reweight the risk engine.

5.4 Software stack

5.5 Governance of AI use

• Human decision. Models rank and explain; officers decide. Automatic action is limited to letters on hard identity breaches under a value cap.
• Explainability by construction. The primary score is a sum of named, record-based terms. Black-box models are used only as cross-checks.
• Random audit share. At least 10% of audits are random, to measure the gap without bias and to test whether the model is drifting.
• Fairness monitoring. Hit rates and selection rates are reported by island, sector and firm size, so the system does not simply concentrate on Port Vila importers because their data is best.
• Legal basis. Data-sharing agreements with VNPF, the Reserve Bank, utilities and the Financial Services Commission need statutory footing before the first extract.

5.6 Team and indicative effort

A core unit of six can run this: one lead, two data engineers, two analysts and one audit liaison, with 12 to 18 months of external support for build and training. Indicative build cost is in the low single-digit millions of US dollars, which the VAT compliance gain would repay within the first two years if the compliance gap closes by even 2 percentage points. Cost figures are my estimates and need market testing.

Limitations

• All quantitative results use synthetic or assumed inputs. Shares, volatilities, the wage distribution and the corporate surplus ratio must be replaced with Treasury, VNPF and national accounts data.
• The elasticity of taxable income for Vanuatu is unknown. International estimates of 0.2 to 0.4 are a starting range; avoidance through incorporation will be the main response and argues for aligning the top PIT rate with the corporate rate plus dividend treatment.
• Model C depends on VSMS coverage. Until coverage is broad, customs and VNPF reconstructions carry the score, which favours detection among importers over service providers.
• The portfolio model treats correlations as stable. A cyclone lowers VAT, trade tax and tourism together, so tail correlation is higher than the average used here.
• Political feasibility is outside the models. Income tax has been proposed and withdrawn before.

Extended modelling: improvements and further research

8.1 Research update, September 2026

VSMS is live. The system was launched on 10 March 2026, is described by the Department of Customs and Inland Revenue as fully operational, and large and very large taxpayers were required to be accredited and issuing invoices through it by 1 July 2026, with an extension process for those still integrating. About 3,000 VAT-registered businesses are in scope. The rollout is province by province: Shefa first, then Sanma, Malampa, Penama, Torba and Tafea through 2027 and 2028. Each registered business gets a smart card device that signs transactions from its point-of-sale software. For the models in this paper the consequence is direct: from mid-2026 Vanuatu has a real-time, third-party-signed record of B2C sales for its largest taxpayers, which is the custodian record Model C was written to use. The staged provincial rollout also creates a natural experiment, discussed below.

8.2 Model A, revenue portfolio: improvements

• Replace static correlations with a regime-switching covariance. Cyclone and earthquake years show high tail correlation across VAT, trade tax and tourism. Estimate a two-regime Markov-switching model on monthly revenue by stream, and report portfolio CV separately for calm and shock regimes. The reform mix should be chosen on the shock-regime covariance, because that is when diversification is needed.
• Model ECP as a stochastic process with jumps. Application flows drop discretely on visa-access decisions. A compound Poisson jump-diffusion on monthly ECP receipts, with jump intensity as a function of an index of visa-access events, gives forecast bands with realistic tails and a value for each additional visa-waiver loss.
• Add a fiscal-rule constraint. Extend the mean-variance problem with a chance constraint that total revenue stays above the wage bill plus debt service with probability 0.95, so that the optimiser cannot recommend a mix that is low-variance on average but fails in the years that matter.

8.3 Model B, income-tax microsimulation: improvements

• Build the population from linked administrative data, not a synthetic distribution. VNPF contribution records give monthly wages for every formal employee; business-licence and VAT records give proprietors; the 2019 census and the latest HIES supply weights for the informal population. Link with probabilistic record matching on name, date of birth and employer.
• Estimate the elasticity of taxable income with a bunching design. Once a PIT exists, the threshold creates an incentive to bunch below it. Before that, borrow the design from the VAT threshold: the density of declared turnover just below VT 4 million reveals how responsive Vanuatu's firms are to a notch, and gives a first estimate of the behavioural parameter the microsimulation needs.
• Model incorporation as an avoidance channel. With a corporate rate below the top PIT rate, high earners incorporate. Add a discrete-choice module where each individual above the threshold chooses employee, sole trader or company status, with switching cost and a tax-differential term. This determines the revenue-maximising alignment of the two rates.
• Add a compliance-cost survey. The optimal-threshold formula needs administration cost per payer and compliance cost per payer. Neither is known for Vanuatu; a 300-firm survey in Port Vila and Luganville would supply both within a quarter.
• Distributional analysis with in-kind income. A large share of rural households' income is subsistence production. The Kakwani index should be computed on total income including imputed subsistence, or the tax will look more progressive than it is.

8.4 Model C, VAT gap and audit selection: improvements

• Exploit the provincial rollout as a difference-in-differences experiment. Shefa firms integrate in 2026; Sanma and the others in 2027 and 2028. Comparing declared VAT of treated and not-yet-treated firms, within sector and size, identifies the causal effect of VSMS on compliance before the rollout is complete. This is the cleanest estimate of the compliance gap the country will ever get, and it costs nothing to collect.
• Use VSMS data to estimate sector mark-ups directly. Model C used a sector mark-up on imports to reconstruct turnover. With signed sales data for the compliant core, mark-ups and their dispersion can be estimated by sector and firm size, replacing assumed parameters with measured ones.
• Detect under-issuance, not just under-declaration. VSMS records what is issued through the device; the evasion channel moves to sales made off-device. Add a model of expected transaction count and value by hour and day, fitted to each firm's own history, that flags gaps (a restaurant issuing no invoices on Friday evenings). Extreme-value theory on inter-transaction times gives the significance test.
• Semi-supervised learning with propensity-corrected labels. As audits close, outcomes are only observed for selected firms. Train the supervised layer with inverse-probability weighting on the selection score, or the model will learn the selector's biases rather than the evasion pattern.
• Formal fairness constraint in the allocator. Add constraints that selection rates by province and by firm size lie within a band of their base rates, so that data-rich Port Vila importers do not absorb the entire audit programme.

8.5 New models proposed

• EU-listing impact model. Estimate the effect of the February 2026 reconfirmation on correspondent-banking fees and trade finance costs using bank-reported spreads, then on import volumes and VAT. This gives a revenue value for the reforms needed to exit the list, which is currently argued qualitatively.
• Presumptive tax for the informal tail. A turnover-proxy model (electricity, mobile-money receipts, import records) predicts turnover for unregistered traders; a simple presumptive charge collected with the business licence is then designed with the same Laffer logic as the artisanal-gold papers in the companion series: set the rate where the formalisation response allows.
• Land and rent register. Rental income is the only taxed income in Vanuatu today. A hedonic model on lease registrations and rental listings would identify under-declared rents and give a base for folding rent tax into PIT.

8.6 Further research

• A published VAT gap estimate for Vanuatu (none exists).
• Effect of the 2020 to 2025 visa-access losses on ECP demand, estimated as an event study.
• Firm-level evidence on the VAT threshold notch.
• Evaluation design for VSMS, registered before the 2027 provincial rollout so that the estimate cannot be accused of hindsight.
• Political-economy analysis of previous income-tax proposals (2017 onward) and the coalition conditions under which one could pass.

Sources for this section: DCIR VSMS awareness presentation; DCIR VSMS extension process; Vanuatu Business Review, 11 March 2026; Daily Post on the VT300m system and 3,000 registrants.

References

• International Monetary Fund. Vanuatu: 2025 Article IV Consultation, Staff Report. September 2025.
• International Monetary Fund. Press Release 25/314: Executive Board Concludes 2025 Article IV Consultation with Vanuatu. 25 September 2025.
• Government of the Republic of Vanuatu. Budget 2026, Volume 1.
• IMF: Vanuatu citizenship revenues reach 7% of GDP. Citizenship by Investment News, 26 September 2025 (secondary).
• Vanuatu tax rates and EU list status, 2026 (secondary; confirm against the EU Council conclusions of 17 February 2026).
• Methods: Keen, M. (2013), The Anatomy of the VAT, IMF Working Paper 13/111; Hutton, E. (2017), The Revenue Administration Gap Analysis Program, IMF Technical Notes and Manuals 17/04; Kakwani, N. (1977), Measurement of tax progressivity, Economic Journal 87; Rousseeuw, P. and Van Driessen, K. (1999), A fast algorithm for the minimum covariance determinant estimator, Technometrics 41. Method citations are from memory; verify before publication.