Introduction

Tonga collects more tax relative to GDP than most Pacific neighbours, yet remains dependent on grants. Three structural facts explain this. Most taxable consumption is imported and financed by remittances, so revenue rises and falls with transfers from Tongans abroad. Labour mobility schemes move wage earners offshore, thinning the PAYE base. And every disaster, including the January 2022 Hunga Tonga-Hunga Ha'apai eruption and tsunami, produces both a collection shock and a new layer of exemptions that rarely expire.

The research question follows the IMF's constraint. If rates cannot rise, how much revenue can be recovered from the existing base, from which exemptions and which compliance channels, at what cost to poorer households, and how should part of it be set aside so that the next disaster does not force new borrowing?

The organising identity is the split of the revenue gap for tax k 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: exemptions}} + \underbrace{\left(R^{law}_{k} - R_{k}\right)}_{\text{compliance gap: border and inland}}

Current tax issues

2.1 Grant-masked deficits and debt

The FY2025 overall surplus of 5.6% of GDP was largely driven by grants. Counting only firmly committed grants while reconstruction spending stays high, the IMF projects a deficit of 7.7% of GDP in FY2026. Tonga is assessed at high risk of debt distress. Debt service reached 3.5% of GDP in FY2024 and stays above 3% until FY2027. The IMF asks the government to avoid new non-concessional borrowing, which leaves domestic revenue and grants as the only financing.

2.2 Exemptions

Consumption tax and duty concessions have accumulated through disaster relief, reconstruction, development projects, diplomatic and church privileges, and sector incentives for tourism, agriculture and fisheries. Each is defensible alone. Together they break the input-credit chain, create routes for diverting exempt goods into taxed markets, and require officer discretion at the counter. I found no regularly published tax-expenditure statement, so the cost is unmeasured. That is the first gap this paper addresses.

2.3 Border dependence

Most consumption tax and excise is collected on imports at Nuku'alofa. The IMF credited stronger consumption and income tax receipts in FY2024 to remittances and construction. The main evasion channels are therefore under-valuation, misclassification into lower-duty or exempt codes, and abuse of exemption certificates. The inland taxpayer population is small, so audit coverage inland yields less than targeting at the wharf.

2.4 A shrinking wage base

Seasonal and longer-term labour schemes in Australia and New Zealand take working-age Tongans offshore. Their income is taxed abroad or not at all, and returns as remittances that are taxed only when spent. The tax mix drifts toward the border whether or not policy intends it, which raises the stakes on customs integrity.

2.5 Disaster shocks

Tonga ranks among the most disaster-exposed countries in the world. The 2022 eruption halted commerce and collection for weeks. Tonga has led support for the Pacific Resilience Facility, and a Disaster Risk Management bill is in train, but no fiscal rule links good-year revenue to a national buffer.

2.6 Governance

The IMF flags governance and corruption vulnerabilities and notes the 2024 establishment of the Anti-Corruption Commission Office. Discretionary exemptions and manual valuation are the two points in the tax system where those vulnerabilities bite. Models that replace discretion with a published score reduce that exposure.

Models

3.1 Model A: pricing and ranking exemptions

Model A answers the IMF's instruction directly: which exemptions to remove at unchanged rates. It starts from the theoretical consumption tax liability across commodities i:

VTTL = \sum_{i} \tau_i \left[ HC_{i} + GC_{i}\,\phi^{G}_i + \sum_{s} IC_{i,s}\left(1-\rho_s\right) + GFCF_{i}\left(1-\rho^{K}_i\right) \right] - \theta

HC is household consumption, GC government consumption with taxable share phi, IC intermediate consumption by sector s with input-credit recovery rho (zero for exempt sectors), GFCF investment, and theta the revenue excluded by the registration threshold. In Tonga, where most taxable goods are imported, VTTL can be built from customs entries by HS code plus the Household Income and Expenditure Survey, without a supply-use table.

The cost of exemption j is found by re-running VTTL with j switched off, allowing demand to respond through own-price elasticity epsilon and netting off the input credits that become claimable:

TE_j = \tau \cdot B_j \cdot \left(1 + \varepsilon_j \cdot \frac{\tau}{1+\tau}\right) - \Delta\rho_j

Each exemption then gets a targeting ratio: the share of its benefit that reaches the bottom 40% of households, using survey spending x and weights w:

\eta_j = \frac{\sum_{h \in B40} w_h\, x_{h,j}}{\sum_{h} w_h\, x_{h,j}}

A leakage term is added for exemptions granted by certificate, estimated as the share of exempt imports by certificate holders that exceeds a plausible own-use benchmark U:

\ell_j = \frac{\sum_{c} \max\!\left(0,\; M^{exempt}_{c,j} - U_{c,j}\right)}{\sum_c M^{exempt}_{c,j}}

The decision rule removes exemption j when its targeting is poor or its leakage high, and returns part of the revenue to poor households as a transfer:

\text{remove } j \;\; \text{if} \;\; \eta_j < \eta^{*} \;\text{or}\; \ell_j > \ell^{*}; \qquad \Gamma = \sum_{j \in removed} TE_j \cdot \eta_j; \qquad \Delta R^{net} = \sum_{j \in removed} TE_j \left(1 - \eta_j\right)

Because eta is below 1 for every exemption, net revenue is positive for every removal, and the bottom 40% are held harmless in aggregate.

3.2 Model B: label-free customs valuation

Model B scores every import entry for under-valuation and misclassification without needing past seizure data. For entry e of HS code g from origin o in month t, the declared unit value u is compared with a robust reference built from three independent sources: the peer distribution of other importers, mirror export data from the partner country, and published international prices where they exist:

z^{peer}_e = \frac{\ln \tilde{u}_{g,o,t} - \ln u_e}{1.4826 \cdot MAD_{g,o,t}}, \qquad z^{mirror}_e = \frac{\ln u^{X}_{g,o,t} - \ln u_e}{\sigma^{X}_g}

The tilde denotes the median and MAD the median absolute deviation, so a cluster of under-valuers cannot pull the benchmark down. Physical constraints give hard flags: declared weight must be consistent with container count and shipping manifest; declared quantity of vehicles must equal registrations. Misclassification is scored by the duty saved if the entry moved to its nearest plausible neighbour code g-prime, weighted by the similarity of the goods description d:

m_e = \max_{g'} \left[ \left(\tau_{g'} - \tau_g\right) \cdot \mathrm{sim}\!\left(d_e, g'\right) \right]

The entry score and the importer score, which accumulates over the importer's entries, are:

A_e = \omega_1 \max(0, z^{peer}_e) + \omega_2 \max(0, z^{mirror}_e) + \omega_3 m_e + \delta H_e, \qquad A_i = \frac{\sum_{e \in i} V_e A_e}{\sum_{e \in i} V_e}

Expected revenue at stake converts the score to pa'anga, and determines whether the response is inspection at the wharf or post-clearance audit of the importer:

\mathbb{E}[Y_e] = \left(\tau^{duty}_g + \tau^{excise}_g + \tau^{CT}\left(1 + \tau^{duty}_g + \tau^{excise}_g\right)\right) \cdot q_e \cdot \max\!\left(0, \tilde{u}_{g,o,t} - u_e\right)

3.3 Model C: disaster revenue buffer

Model C sets how much revenue to save in normal years. Disaster events arrive as a Poisson process with rate lambda. Each event destroys a share X of annual revenue, with X lognormal and capped. Annual revenue loss L is the compound sum:

N \sim \mathrm{Poisson}(\lambda), \qquad X_n \sim \mathrm{LogNormal}(\mu_X, \sigma_X^2), \qquad L = \min\!\left(\sum_{n=1}^{N} X_n,\; L^{max}\right)

The buffer target is the loss quantile at confidence q, net of expected insurance and facility payouts I:

B^{*} = VaR_q(L) \cdot \bar{R} - \mathbb{E}\left[I \mid L \ge VaR_q\right]

The contribution rule saves a fixed share c of revenue plus a share of any revenue surprise above forecast, until the buffer is full, and permits drawdown only on a declared disaster:

B_{t+1} = B_t (1 + r) + \mathbb{1}\!\left[B_t < B^{*}\right]\left( c R_t + \psi \max\!\left(0, R_t - \hat{R}_t\right) \right) - \mathbb{1}\!\left[\text{disaster}_t\right] \min\!\left(B_t, L_t R_t\right)

The rule links Models A and B to Model C: the first use of revenue recovered from exemptions and customs is filling the buffer, which is a use the IMF's call to rebuild buffers supports and which does not raise recurrent spending.

Algorithms and illustrative results

4.1 Algorithm 1: exemption ranking

INPUT  customs entries with exemption codes, HIES microdata, CT returns by sector,
       elasticities eps[j], thresholds eta_star, ell_star
BUILD  VTTL_base from imports by HS code + domestic taxable supply
FOR each exemption j:
  VTTL_j   <- rebuild with exemption j removed, demand adjusted by eps[j]
  TE[j]    <- VTTL_j - VTTL_base - new input credits
  eta[j]   <- bottom-40% share of spending on j from HIES
  ell[j]   <- share of certificate imports above own-use benchmark
  score[j] <- TE[j] * (1 - eta[j]) * (1 + ell[j])
SORT exemptions by score descending
SELECT j where eta[j] < eta_star OR ell[j] > ell_star
COMPUTE transfer Gamma = sum TE[j]*eta[j]; net gain = sum TE[j]*(1 - eta[j])
PUBLISH annual tax-expenditure statement with TE, eta, ell per exemption

4.2 Algorithm 2: customs entry scoring

FOR each (HS code g, origin o, rolling 12-month window):
  u_med, MAD <- robust centre and spread of declared unit values
  u_mirror   <- partner-country export unit value where available
FOR each new entry e at lodgement:
  z_peer, z_mirror <- standardised shortfalls (positive side only)
  m_e   <- duty saved under nearest plausible code * description similarity
  H_e   <- 1 if weight, container or registration identities are breached
  A_e   <- w1*z_peer + w2*z_mirror + w3*m_e + delta*H_e
  EY_e  <- full tax rate stack * quantity * max(0, u_med - u_e)
  IF H_e = 1 OR EY_e > inspect_threshold: hold for documentary or physical check
  ELSE release, and add A_e to importer running score A_i
MONTHLY: rank importers by value-weighted A_i -> post-clearance audit list
AFTER each finding: update weights w by Bayesian reweighting

4.3 Illustrative results: disaster buffer

Inputs are assumed, not estimated: event rate 0.3 a year, median loss per event 6% of annual revenue, log standard deviation 0.9, annual loss capped at 80%. The run uses 200,000 simulated years.

Budgeting for the mean loss leaves the government short in exactly the years that matter. A buffer sized to the 1-in-20-year loss is about 16% of annual revenue. A contribution of 3% of revenue a year fills it in about five years; 2% takes about eight. Losses beyond the 95th percentile are better transferred to parametric insurance and the Pacific Resilience Facility than self-insured.

4.4 Illustrative arithmetic: exemptions

No Tongan tax-expenditure estimate is public, so this is arithmetic, not a finding. If exemptions cost 3% of GDP, a range typical for small states, and the exemptions failing the targeting and leakage tests account for half of that with an average eta of 0.25, then removal yields 1.5% of GDP gross, 0.4% returns to poorer households as transfers, and 1.1% of GDP is net new revenue with no rate increase. That is about one-seventh of the projected FY2026 deficit.

Solutions, software, data science and AI models

Tonga's solution is mostly at the border and in the exemption register, so the software centres on customs entries, exemption certificates and a published tax-expenditure statement. Each issue maps to a policy measure, a model and a component.

5.1 Issue-to-solution map

5.2 Model catalogue

5.3 Reference architecture

flowchart TD
  A[Sources<br/>customs entries, CT and PAYE returns, exemption certificates, HIES, remittance data] --> B[Ingestion and data quality]
  B --> C[Warehouse<br/>bitemporal, single taxpayer view]
  C --> D[Entry scoring service<br/>unit value, HS classifier, hard flags]
  D --> E[Customs selectivity<br/>release, document check, inspect]
  C --> F[Importer risk and certificate leakage]
  F --> G[Post-clearance audit cases]
  G --> H[Findings]
  H --> D
  C --> I[Policy lab<br/>tax expenditures, forecasts, buffer rule]
  I --> J[Published TE statement and dashboards]

The entry scoring service is the only component that must respond in real time; everything else runs nightly or monthly.

5.4 Software stack

5.5 Governance of AI use

• Shadow mode first. Run entry scoring alongside officer selection for six months and publish the comparison of hit rates before switching over.
• Explainable flags. Each held entry shows the reference price, its source, the number of peer entries and the duty at stake. Importers can contest the reference, which keeps the system inside WTO valuation rules.
• No automated penalties. Scores select for checking; assessments remain officer decisions under the Customs and Revenue Acts.
• Random inspection share. A fixed share of green-lane entries is inspected at random to estimate the true leakage rate and detect drift.
• Integrity safeguards. Officer-pairing analysis is reported to an integrity unit outside the line of command, with due-process rules set in advance.

5.6 Team and indicative effort

A joint customs and revenue analytics cell of five is enough: one lead, one data engineer, two analysts and one customs valuation specialist. External build support of about 12 months is realistic, and regional pooling with other Pacific administrations would cut cost, since unit-value benchmarks improve when several small islands share reference prices. Cost and staffing figures are my estimates.

Limitations

• The exemption and buffer figures are assumptions. The first deliverable is measurement: a tax-expenditure statement and a fitted loss distribution from Ministry of Finance revenue series and disaster records.
• Mirror trade data for small islands is noisy, with transhipment through Fiji and New Zealand. Peer and price benchmarks should carry more weight than mirror data until its error is measured.
• For thinly traded HS codes there are too few entries to form a peer distribution. The model should pool across origins or fall back to the WTO valuation methods in sequence.
• Compensating transfers assume a delivery channel that reaches the bottom 40%. Coverage in the outer islands needs checking.
• The Poisson assumption treats events as independent. Climate trends imply a rising rate, so the buffer target should be re-estimated every three years.

Extended modelling: improvements and further research

8.1 Research update, September 2026

The IMF Executive Board concluded the 2025 Article IV on 6 November 2025 and the staff report (Country Report 25/293) is published. Growth was 2.7% in FY2025 on reconstruction, remittances, tourism and stronger-than-expected grants, and is projected at 2.3% for FY2026. Headline inflation is below the 5% reference rate but core inflation is near 10%, which matters for a consumption-tax base because it signals demand pressure in exactly the imported goods the tax reaches. The Board endorsed a contractionary FY2026 stance to rebuild buffers, and reiterated that medium-term consolidation plus additional grants are needed to put debt on a downward path. Reserves cover about 10.5 months of imports. Nothing in the published material indicates that a tax-expenditure statement has been produced, so the measurement gap this paper leads with remains open.

8.2 Model A, tax-expenditure ranking: improvements

• Estimate demand elasticities from Tonga's own data. The QUAIDS system can be estimated on the HIES with the standard spatial-price-variation identification (prices differ between Tongatapu and the outer islands). Where that fails, pool with Samoa and Fiji HIES data and impose Pacific-common parameters, with a Tonga fixed effect.
• Model second-round effects on the CPI and on the wage bill. Removing an exemption raises the price of the exempt good, which feeds the CPI, which feeds indexed public wages and pensions. Add a small input-output pass-through block so the net fiscal effect includes the expenditure side; some exemptions cost less to keep than to remove once indexation is counted.
• Treat exemption certificates as a network. Certificate holders, their suppliers and the retailers who end up with the goods form a graph. Leakage is detected as exempt import volume flowing to nodes that also sell at retail. The leakage estimate becomes a graph-flow calculation rather than a per-holder benchmark.
• Sunset-clause simulation. For each exemption, simulate revenue under automatic expiry with renewal on evidence versus indefinite continuation; report the option value of the sunset. This gives Parliament a number for a rule it is otherwise asked to adopt on principle.

8.3 Model B, customs valuation: improvements

• Hierarchical Bayesian reference prices. Thin HS codes have too few entries for a robust median. Fit a hierarchical model where each code's reference price shrinks toward its chapter and origin group, so thinly traded lines borrow strength from related ones. Report the posterior interval, not just the point, and hold an entry only when the declared value falls outside it.
• Mirror-data error model. Rather than treat partner-country export data as truth, estimate its measurement error by comparing it with Tongan entries for goods with no incentive to under-value (zero-duty, zero-CT lines). Use the estimated error to weight the mirror z-score.
• Learn the HS classifier from officer corrections. Every reclassification an officer makes is a labelled example. A multilingual sentence-embedding model fine-tuned on those corrections outperforms a rules-based classifier within a few thousand examples; log corrections from day one.
• Officer-importer pairing as a formal test. Fit a bipartite exponential random graph model to clearance assignments; pairings that occur far more often than the model predicts, and that coincide with low inspection rates, go to the integrity unit. This turns a governance concern into a statistic with a p-value.
• Post-clearance audit as a bandit problem. With limited audit capacity and unknown hit rates by importer segment, allocate audits with a Thompson-sampling rule that balances exploiting known high-yield segments against exploring unknown ones. This is better suited than a fixed knapsack when the yield distribution is being learned.

8.4 Model C, disaster buffer: improvements

• Fit the loss distribution to Tonga's actual events. Revenue by month around Gita (2018), Harold (2020) and the 2022 eruption gives three observed shocks. Use them to calibrate the severity distribution and a recovery curve (months to return to trend), rather than the assumed lognormal.
• Non-stationary event rate. Replace the constant Poisson rate with a rate that trends with a climate index, and re-estimate every three years. Report the buffer target as a path, not a number.
• Layered financing optimisation. Choose the split among a cash buffer, contingent credit, parametric insurance and the Pacific Resilience Facility to minimise expected cost for a target coverage level. The cash buffer covers frequent small losses; insurance covers the tail; the optimum depends on premium loadings that can be obtained from PCRIC quotes.
• Revenue-side continuity model. Model days of collection lost as a function of damage to the wharf, the customs system and connectivity, and cost the offline-capable collection nodes proposed in section 5 against the revenue they preserve.

8.5 New models proposed

• Remittance-to-revenue pass-through. Estimate the elasticity of border tax and CT revenue to remittances with monthly data, controlling for reconstruction imports. This is the single most useful forecasting parameter for Tonga's budget and does not appear in published work.
• Labour-mobility fiscal model. Track the PAYE base lost per departing worker against the CT gained from remittances spent, by sector and household type. The net is probably positive today and will turn negative as departures shift from low-wage to skilled workers; the model gives the crossing point.
• State-owned financial institutions. The IMF flags governance at state-owned financial institutions. A model of their dividend and tax contributions, and of contingent liabilities to the budget, belongs beside the tax models.

8.6 Further research

• A first tax-expenditure statement for Tonga, using customs and HIES data, published with the FY2027 budget.
• An event-study evaluation of the mineral trading and buying-centre reforms in Pacific and African peers as a template for the exemption reform.
• Regional pooling of customs reference prices among Tonga, Samoa, Vanuatu and Fiji; a shared benchmark database is cheaper and more accurate than four separate ones.
• The distributional incidence of core inflation near 10% on CT revenue and on the compensating-transfer cost.

Sources for this section: IMF Press Release 25/365, 10 November 2025; IMF Country Report 25/293, Tonga 2025 Article IV staff report.

References

• International Monetary Fund. Tonga: Staff Concluding Statement of the 2025 Article IV Mission. 25 August 2025.
• International Monetary Fund. Tonga: Staff Concluding Statement of the 2024 Article IV Mission. August 2024.
• International Monetary Fund. Tonga: 2024 Article IV Consultation, Staff Report. Country Report 24/326.
• International Monetary Fund. Tonga: 2025 Article IV Consultation. Country Report 25/293 (not opened for this draft; consult for final figures).
• 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; Cummins, J. D. and Mahul, O. (2009), Catastrophe Risk Financing in Developing Countries, World Bank; WTO Agreement on Customs Valuation. Method citations are from memory; verify before publication.