Introduction

Samoa collects tax well. Its problem is that what it collects depends on things it does not control: remittances, grants and the weather. The 2026 Article IV projections make the point in one table. Revenue and grants fall from 35.6% of GDP in FY2024 to 27.0% in FY2026, tax revenue from 24.2% to 21.8%, and growth from 4.8% to 0.4%. When the shocks pass, the projections recover to a tax ratio near 22.9%, which is below the FY2024 level. A large surplus built on restrained spending and unexecuted capital projects is not fiscal strength; it is deferred investment.

The research question is how Samoa can make its revenue more its own: less tied to remittance cycles and grant timing, more tied to domestic value added, and protected against the next cyclone. The paper uses the same gap identity as the rest of the series:

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}}

Samoa's policy gap is modest; its compliance and structural gaps are what matter.

Current tax issues

2.1 Revenue that follows remittances

Remittances are among the largest inflows in the economy and finance imported consumption. Import duty, excise and VAGST at the border are collected on that consumption, so the tax base rises when Samoans abroad send more and falls when they send less. The IMF's 2023 and 2024 consultations credited buoyant tax revenue partly to remittances and tourism recovery. The FY2026 dip in the tax ratio shows the reverse.

The scale is worth stating. Remittances to Samoa run at roughly a quarter to a third of GDP, among the highest ratios in the world, sent mainly from New Zealand, Australia and the United States by a diaspora larger than the resident population. They arrive as cash to households and are spent, overwhelmingly, on imported goods: food, fuel, vehicles, building materials, consumer durables. Every one of those purchases passes the wharf at Apia or the airport at Faleolo, where import duty, excise and VAGST are collected on the landed value, and then passes through a registered retailer, where VAGST is collected again on the margin. More than half of Samoa's tax revenue is therefore collected on transactions financed by money earned in Auckland and Sydney.

The consequence is a tax base that the domestic economy does not control. When New Zealand's labour market is strong and the diaspora sends more, Samoa's revenue rises with no change in domestic output; when a recession abroad or a stronger tala cuts the flow, revenue falls with no change in domestic policy. The IMF's 2023 and 2024 consultations credited buoyant tax collection partly to remittances and tourism recovery, and the 2026 consultation, discussed in section 8, shows the reverse: a weak year for domestic demand with revenue falling by about two and a half points of GDP.

Part II works the arithmetic. If roughly 55% of tax revenue is collected at or near the border with an elasticity to remittances near 0.9, a 10% fall in remittances lowers tax revenue by about 1.1% of GDP, and a 20% fall by about 2.2%. Those are the magnitudes of the swings the IMF projections show, and they are outside the Ministry of Finance's control. The paper's first model exists to measure that dependence and to give the budget a revenue figure with the remittance cycle removed.

2.2 Grant volatility and under-execution

Grants swing by several percentage points of GDP from year to year, driven by donor project cycles. The IMF has repeatedly flagged under-execution of public investment because of capacity constraints. The surplus is therefore partly an artefact of projects not built, and the future tax base is smaller because of it.

Grants to Samoa come from Australia, New Zealand, China, Japan, the Asian Development Bank and the World Bank, most of them tied to projects: a wharf, a hospital, a road, a submarine cable. Each project has its own appraisal, procurement and disbursement cycle, and grant receipts in any budget year depend on where each project sits in that cycle rather than on anything the government decides. The IMF's tables show the result: grants of 8.3% of GDP in FY2024, 2.5% in FY2026 and 6.9% projected for FY2027. A swing of six points of GDP in grants is larger than any tax measure Samoa could take.

Under-execution compounds the volatility. A grant that is committed but not spent shows up as a surplus, and Samoa's FY2024 surplus of 10.1% of GDP reflected, in part, capital projects that were appropriated and not built. The IMF has repeatedly noted capacity constraints in project implementation: too few project managers, procurement rules that stall on donor conditions, and contractors stretched across concurrent works. The fiscal accounts record the surplus as strength. The economy records it as a hospital or a road that does not yet exist, and the tax base of 2030 records it as output that was never generated.

The paper treats under-execution as a fiscal problem in its own right, because revenue reform that funds a capital budget which is not executed changes nothing. Section 3.3 models the execution rate as a function of capacity and the number of concurrent projects, and Part II shows that at plausible congestion parameters a pipeline of twenty concurrent projects executes at a third to a half of appropriation while a pipeline of ten executes at two-thirds or better. Fewer, larger, fully executed projects raise the future tax base more than a wide pipeline that stalls.

2.3 VAGST compliance and the informal economy

VAGST at 15% is the largest domestic tax. The registration threshold and the size of the informal trading and services economy leave a compliance tail. There is no published VAGST gap estimate that I could find, so the size of the compliance gap is unknown.

VAGST at 15% is Samoa's largest domestic tax and the one most exposed to the informal economy. Registration is required above a turnover threshold; below it, and among registered traders who under-report, the tax is lost. Samoa's informal sector is large by the standards of its income level: village-based agriculture and fishing, market vendors, small transport operators, and a service economy of workshops, guesthouses and food outlets, many of them family businesses that keep no accounts. The formal retail sector, dominated by a few large importer-wholesalers and their outlets, is the compliant core; the tail is long.

No published VAGST gap estimate exists for Samoa. That is the first thing this paper would change, because a gap that has not been measured cannot be targeted. Part II works an illustrative gap on rough national-accounts figures: with final consumption near WST 2.6 billion, exemptions removing about 18% of the base, and VAGST collections near WST 280 million, C-efficiency is about 0.8, the policy gap about 18% and the compliance gap about 12%. Those numbers are assumptions dressed as an example, and their point is that the gap is measurable from data the Ministry for Revenue and the Bureau of Statistics already hold.

The compliance gap, once measured, is addressed by the label-free audit model of section 3.2, which reconstructs each registered trader's turnover from customs, bank and provident-fund records. The policy gap is a choice about exemptions, which the Tonga paper in this series treats in detail and which Samoa could examine with the same tax-expenditure method.

2.4 Excess liquidity and the exchange-rate peg

The IMF notes a large current-account surplus, reserve build-up and excess liquidity in the banking system. On the tax side this shows up as low taxable interest income and a financial sector whose profits are not large relative to its balance sheet. Bank data is also the strongest untapped third-party source for verifying business turnover.

Samoa's banking system holds deposits well in excess of what it lends, a structural condition the IMF has noted for several years and attributes to remittance inflows, reserve accumulation and limited bankable projects. Interest rates on deposits are low, banks earn little on their excess reserves, and the financial sector's taxable profit is small relative to its balance sheet. On the tax side this is a modest loss: little taxable interest income, low corporate tax from banks. On the audit side it is an opportunity that the paper's second model is built around.

Excess liquidity means that almost every business transaction of any size passes through a bank account, because there is nowhere else for the money to go. Deposit records in Samoa are therefore an unusually complete record of business turnover, more complete than in economies where cash or mobile money dominates. A registered VAGST trader whose bank credits are well above its declared turnover has a question to answer, and the bank holds the evidence. Part II works an example in which a trader declaring WST 1.2 million shows bank credits of WST 1.5 million, provident-fund wages consistent with WST 1.3 million, and imports consistent with only WST 0.96 million; the pattern, high bank credits and high wages against low imports, is the signature of a service business under-declaring, and it is visible only because the bank record exists.

Using that record needs a legal basis, which Samoa does not yet have in the form required, and it needs governance: narrow statutory access for matching, access logging, and no bulk transfer of customer data to the revenue authority. Section 5 sets out the safeguards, and the paper's reform sequence places the legislation in FY2028.

2.5 Disaster exposure

Samoa's debt was rated at high risk of distress principally because of natural-disaster vulnerability. The rating was upgraded to moderate on the strength of surpluses, but no fiscal rule commits part of a good year's revenue to a buffer for the bad one.

The 2009 tsunami killed more than 140 people on Upolu's south coast and destroyed villages, resorts and roads; Cyclone Evan in 2012 flooded Apia and damaged the power system; Cyclone Gita in 2018 and the measles epidemic of 2019 each disrupted the economy for months. Each event cut revenue at the wharf and in the shops, raised spending on relief and reconstruction, and drew in grants whose timing the government did not control. The IMF's debt-sustainability rating of high risk of distress rested principally on this exposure, and its upgrade to moderate rested on the surpluses since.

Surpluses are the right response, but surpluses without a rule are discretionary, and discretion is spent. Samoa has no fiscal rule that commits a share of a good year's revenue to a fund for the bad one, and no published estimate of what the bad year costs. The paper's fourth model supplies both: a distribution of annual revenue loss fitted to Samoa's own event history, a buffer target at the 1-in-20-year loss, and a contribution rule that fills the buffer from surplus years. Part II's illustrative run puts the 1-in-20-year loss near 15% of annual revenue and shows the buffer filling in about five years at a contribution of 2.5% of revenue plus half of any windfall. The FY2024 surplus alone would have filled most of it.

2.6 Labour mobility and the PAYE base

Seasonal worker schemes to Australia and New Zealand move wage earners offshore. Their income is taxed abroad or not at all and returns as remittances, taxed only when spent. The base migrates toward the border, which raises the value of customs integrity and of taxing the assets and rents that remittances buy, principally land and rental income.

The Recognised Seasonal Employer scheme in New Zealand and the Pacific Australia Labour Mobility scheme have become, for many Samoan households, the main source of cash income. Several thousand Samoans work abroad under them at any time, and the numbers have grown steadily. The fiscal effect has two signs. Income earned abroad is taxed abroad or not at all, so every worker who leaves a formal Samoan job removes a PAYE payer; the same worker's remittances raise household spending at home, which raises VAGST and border taxes. The net is probably positive today, because most departing workers were in low-wage or informal employment and their remittances are large relative to what they earned at home.

The balance shifts as the schemes draw more skilled workers, nurses, tradespeople and teachers, whose home wages were taxable and whose departure leaves gaps that the public sector must fill at a cost. The base migrates from wages to consumption, and from the Ministry for Revenue's assessment function to its customs function. That migration has two implications the paper takes up: customs integrity becomes the central compliance problem, as in Tonga; and the assets that remittances buy, principally land and rental property, become the untaxed residual. Section 8.6 proposes a labour-mobility fiscal model to track the crossing point and a property and rental register to reach the residual.

Models

3.1 Model A: remittance-driven revenue nowcast and buoyancy

Revenue for tax k responds to remittances M, tourism arrivals A and domestic non-remittance income Y with separate elasticities, estimated on quarterly data:

\ln R_{k,t} = \alpha_k + \beta_{k,M} \ln M_t + \beta_{k,A} \ln A_t + \beta_{k,Y} \ln Y_t + \gamma_k D_t + \varepsilon_{k,t}

D is a disaster indicator. Structural revenue removes the cyclical remittance and tourism components, so the budget is set on what the domestic economy supports:

R^{struct}_{k,t} = R_{k,t} \left(\frac{\bar{M}}{M_t}\right)^{\beta_{k,M}} \left(\frac{\bar{A}}{A_t}\right)^{\beta_{k,A}}

Buoyancy above 1 in booms and below 1 in busts is the signature of a remittance-driven base; the model measures it so that the fiscal rule can target the structural figure.

What the model does and why. The nowcast has two jobs. The first is forecasting: to tell the Ministry of Finance, within weeks rather than quarters, where revenue is heading, using the monthly remittance series from the Central Bank, weekly customs entries and monthly arrivals. The second is decomposition: to separate the part of revenue that the domestic economy generates from the part that remittance and tourism cycles generate, so that the budget can be set on the first. The elasticities beta are the model's substance. Estimated on quarterly data since 2010, they answer questions the budget currently cannot: how much of border revenue moves with remittances, how much of VAGST moves with arrivals, and how much of each moves with domestic income.

Interpreting the structural figure. Structural revenue is what revenue would have been had remittances and arrivals been at trend. If remittances are 8% above trend and arrivals 5% above, and the elasticities are as assumed, Part II shows structural revenue about 4% below actual on illustrative figures: a windfall of about WST 30 million that the budget should treat as temporary. The reverse holds in a weak year, when actual revenue understates the structural position and the right response is to run down buffers rather than cut spending. The FY2026 dip, which section 8 attributes to domestic demand and fuel prices rather than remittances, is a case where the decomposition changes the response: an energy-price shock is temporary and is met from buffers; a remittance shift is persistent and is met by adjusting the base.

Buoyancy as a diagnostic. Revenue buoyancy is the percentage change in revenue per percentage change in GDP. A base driven by remittances shows buoyancy above one in booms and below one in busts, because remittances move more than GDP. Measuring buoyancy year by year against the structural figure tells the Ministry whether a good revenue year is domestic growth or foreign transfers, which is exactly the distinction the fiscal rule needs.

3.2 Model B: VAGST gap and label-free audit selection

C-efficiency and the theoretical liability are built from imports by HS code, the household survey and VAGST returns by sector, exactly as in the Tonga paper. Audit selection reconstructs each trader's turnover from third-party records: customs imports with a sector mark-up, bank credits (the excess-liquidity banking system holds unusually complete deposit records), national provident fund wages over a sector labour share, and utility consumption:

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 allocated by integer programme subject to auditor hours, with a random share for gap measurement.

Two independent estimates of the gap. The top-down estimate uses national accounts and the household survey to compute the theoretical VAGST liability across commodities, as in the Tonga paper, and compares it with collections. The bottom-up estimate audits a stratified random sample of about 150 registrants and scales the findings. The two should agree within their error bands, and where they do not, the reconciliation itself is informative: a top-down gap larger than the bottom-up one points to unregistered traders below the threshold; the reverse points to under-reporting within the register. Part II's illustrative gap of about 12% is a top-down figure and needs its bottom-up check.

How the custodian reconstruction works in Samoa. For each registered trader, three or four independent estimates of turnover are built. Customs imports with a sector mark-up: a retailer that imported WST 600,000 of goods at a typical mark-up of 1.6 should be selling about WST 960,000. Bank credits: the sum of deposits to the trader's business accounts, net of transfers between its own accounts. Provident-fund wages over a sector labour share: a business paying WST 180,000 in wages in a sector where wages are 14% of turnover is running about WST 1.3 million. Utility consumption over a sector intensity, where available. Each estimate is compared with the declaration in log terms and scaled by that estimate's known noise, so that a source that is unreliable for a given sector counts for less. Only positive shortfalls count, because over-declaration is not a revenue risk.

What the score means to an officer. The score is a sum of named terms, each of which says which record disagrees with the return and by how much. In Part II's example, the imports estimate sits below the declaration, the wages estimate slightly above, and the bank estimate well above; the pattern is a service business whose revenue is not import-driven, and the audit question is specific: reconcile WST 1.5 million of bank credits with WST 1.2 million of declared sales. An officer can put that question to the trader; the trader can answer it from its own bank statements. Nothing in the model is a black box, and the random audit share keeps the gap estimate honest and tests whether the model is drifting.

3.3 Model C: public-investment execution and the future base

Revenue reform in Samoa fails if the money is not invested. Let the capital budget be K, execution rate e, and the marginal product of public capital theta. The tax base next period rises with executed investment:

Y^{N}_{t+1} = Y^{N}_t (1+g) + \theta\, e_t K_t, \qquad \Delta R_{t+1} = \tau\, b\, \theta\, e_t K_t

Execution rate is modelled as a function of project-management capacity C and the number of concurrent projects n, with congestion:

e_t = \frac{C_t}{C_t + \kappa\, n_t}

The practical output is a project-pipeline cap: the number of concurrent projects that keeps execution above a target, given the ministry's staffing. Fewer, larger, fully executed projects raise the future tax base more than a wide pipeline that stalls.

Why execution belongs in a revenue paper. Revenue reform is justified by what it funds. If the capital budget it funds is executed at half its appropriation, half the justification is void, and the future tax base that executed investment would have generated does not appear. The model links the two: executed investment raises non-oil output through the marginal product of public capital, and higher output raises revenue through the tax ratio. Part II's example puts the marginal product at 0.15 and a capital budget at WST 300 million: raising execution from 55% to 75% adds about WST 9 million of output a year and WST 2 million of tax, every year, from a single year's budget. Over a decade of budgets the cumulative effect is the difference between a tax base that grows and one that does not.

The congestion function. Execution falls as concurrent projects rise, because the same project managers, procurement officers and contractors are spread across them. The function e = C/(C + kappa n) says that with capacity C and congestion parameter kappa, adding projects lowers the execution rate of all of them. Part II tabulates it: at C = 10 and kappa = 1, five concurrent projects execute at 67%, ten at 50%, twenty at 33%. The policy instrument is the pipeline cap, the number of concurrent projects that keeps execution above a target, and it is a decision the Cabinet can take without new money. Raising C, by hiring or by a project-management unit, moves the whole curve; Part II shows that C = 14 lifts execution at ten projects from 50% to 58%.

Estimating rather than assuming. The Ministry of Finance holds appropriation and disbursement by project since 2015. A regression of execution on concurrent count, donor, procurement route and implementing agency gives kappa and the agency effects; a survival analysis of the time from appropriation to first disbursement gives the delay distribution. Section 8.4 sets both out. Until they are estimated, the pipeline cap should be set conservatively, on the illustrative curve, and revised.

3.4 Model D: disaster revenue buffer

As in the Tonga paper, annual revenue loss is compound Poisson with lognormal severity, and the buffer target is the loss quantile net of expected insurance and facility payouts:

L = \min\!\left(\sum_{n=1}^{N} X_n,\, L^{max}\right), \qquad B^{*} = VaR_{q}(L)\,\bar{R} - \mathbb{E}\left[I \mid L \ge VaR_q\right]

Samoa's surplus years are exactly when the buffer should fill. A contribution rule of a fixed share of revenue plus a share of any revenue above the structural forecast converts windfalls into protection instead of into an idle reserve pile.

The distribution, not the average. Budgeting for the average disaster loss, about 2.4% of revenue a year in Part II's run, leaves the government short in exactly the years that matter, because the average is made up of many years with no loss and a few with a large one. The buffer must be sized on a quantile: the 1-in-10-year loss (about 9% of revenue), the 1-in-20-year loss (about 15%) or the 1-in-100-year loss (about 31%). The paper proposes the 1-in-20-year level as the cash buffer and transfers the tail beyond it to parametric insurance and regional facilities, because self-insuring a 1-in-100-year event ties up more cash than the event's expected cost justifies.

Fitting to Samoa's own events. The compound Poisson model has three parameters: event rate, severity and its dispersion. Samoa's monthly revenue series around the 2009 tsunami, Cyclone Evan, Cyclone Gita and the 2019 measles epidemic gives four observed shocks with which to calibrate severity and a recovery curve, the months it takes revenue to return to trend. The event rate should be treated as trending rather than constant, given climate projections, and re-estimated every three years. Section 8.5 sets out the calibration; the paper's figures are assumptions until it is done.

Windfalls fill the buffer. The contribution rule is the link between Models A and D. Structural revenue defines the windfall, the excess of actual over structural; the rule saves a fixed share of revenue plus a share of the windfall until the buffer reaches target. Part II shows the buffer reaching 15% of revenue in about five years at 2.5% of revenue plus half of periodic windfalls. In a surplus year like FY2024 the windfall alone would have gone most of the way, and the saving would have been invisible in the headline balance, which already showed a surplus. The rule converts a good year's discretion into a bad year's protection.

Algorithms and illustrative results

4.1 Algorithm: structural revenue and buffer rule

QUARTERLY:
  fit elasticities beta[k,M], beta[k,A], beta[k,Y] on rolling 10-year window
  compute R_struct[k] using trend remittances and arrivals
  windfall <- max(0, R_actual - R_struct)
  contribution <- c * R_actual + psi * windfall, while buffer < B_star
ANNUALLY:
  refit disaster loss distribution; recompute B_star at q = 0.95
  publish structural balance alongside headline balance in the budget

4.2 Illustrative results

Inputs are assumptions, not estimates. With an event rate of 0.25 a year, median loss per event of 7% of annual revenue and log standard deviation 0.8, the 1-in-10-year revenue loss is about 8.7% of annual revenue and the 1-in-20-year loss about 15%. A buffer of 15% of revenue, filled at 2 to 3% of revenue a year from surplus years, is reachable within the current projection horizon.

On the base structure, if half of border tax revenue is remittance-elastic with an elasticity near 1, a 10% fall in remittances lowers total tax revenue by roughly 1% of GDP. That is the scale of the FY2026 dip in the IMF projections, which suggests the remittance channel explains most of it.

4.3 Reading the results

The two illustrative results in 4.2 point in the same direction. The remittance arithmetic says that most of the FY2026 revenue dip could be explained by a remittance fall of the size the projections imply, if the border share and elasticity are as assumed; section 8 adds that the IMF's own account attributes the year to domestic demand and fuel prices instead, which is a reminder that the decomposition must be estimated rather than assumed. Either way, the swing is of the order of one to two points of GDP, and it arrived without any change in Samoan policy. A budget that does not separate that swing from the structural position will over-spend in the good years and over-cut in the bad ones.

The buffer arithmetic says that the protection against the bad year is affordable. A 1-in-20-year revenue loss near 15% of annual revenue, roughly WST 110 million on current collections, is covered by a buffer that fills in about five years from contributions the surplus years can easily bear. Set beside the FY2024 surplus of 10.1% of GDP, the buffer target is not large; what has been missing is the rule that turns discretion into saving.

4.4 What the illustrative numbers do not show

All figures in 4.2 and Part II are assumptions with Samoa's scale attached. The elasticities in the remittance calculation, the border share of revenue, the disaster event rate and severity, the congestion parameter and the marginal product of public capital are each estimable from data the Central Bank, the Ministry of Finance and the Bureau of Statistics hold, and none has been estimated here. The paper's contribution is the specification: which parameters matter, which data identify them, and what decision each feeds. The numbers show what the answer would look like if the parameters were near the values assumed, and they should be replaced before any figure is quoted.

Solutions, software, data science and AI models

Software stack. Revenue management system and ASYCUDA World (confirm versions); Python with Prefect or Airflow; PostgreSQL with bitemporal tables; scikit-learn, statsmodels, PyOD, PyMC; OR-Tools; Splink for record linkage; Apache Superset or Metabase; Docker and MLflow; hosting in the government data centre with an off-island encrypted replica. All open source apart from existing vendor systems.

Governance of AI. Officers decide; models rank and explain. A random audit share measures the gap. Bank-data matching needs a legal basis and access logging. Selection and hit rates are reported by island and sector so that Savai'i is not under-covered because its data is thinner.

Team. A four-person analytics cell across the Ministry for Revenue and the Ministry of Finance, with about 12 months of external support. Regional pooling with Tonga and Vanuatu on reference prices and audit methods would lower cost for all three.

5.1 The solution described as a system

The forecasting workbench is the Ministry of Finance's tool. It holds monthly remittances from the Central Bank, weekly customs entries, monthly arrivals, the fuel price and the freight index, and quarterly revenue by stream. It runs the elasticity regressions, the unobserved-components decomposition and the mixed-frequency nowcast, and it reports three figures each month: the nowcast for the quarter with a band, the structural revenue figure, and the windfall or shortfall against it. The budget's revenue chapter is written from it, and the headline and structural balances are published side by side.

The public investment dashboard links the FMIS to the execution model. It reads appropriation and disbursement by project, computes execution rates and delays, fits the congestion function, and reports the current pipeline against the cap. Its purpose is to make the Cabinet's project decisions visible in execution terms: adding a project shows its predicted effect on every other project's execution.

The risk engine and case management serve the Ministry for Revenue. The engine links VAGST returns to customs entries, provident-fund contributions, utility consumption and, once the law allows, bank credits, all through the single taxpayer view built by probabilistic record linkage. It reconstructs turnover, scores shortfalls, flags identity breaches, and allocates audits by integer programme with a random share and fairness bands by island. Cases route to automatic letters for identity breaches under a value cap, to officers for ranked cases, and to senior review for large or sensitive entities. Audit outcomes feed back to reweight the sources.

The secure matching service is how bank data enters. It does not transfer customer data to the revenue authority. It receives from each bank a monthly file of credits by business account, matched to a VAGST registration number, under a statutory provision that limits use to turnover verification, and it logs every access. The revenue authority sees a number per registrant, not a statement.

The buffer calculator and the medium-term framework implement the rule. The calculator fits the loss distribution, sets the target, computes the contribution from structural revenue and windfall, and reports the balance quarterly. The medium-term fiscal framework carries the target, the contribution rule and the drawdown conditions.

The property and rental register reaches the residual base. Lease registrations, rental listings and building permits are linked to the tax file; a hedonic model gives an expected rent per property; declared rental income is tested against it. This is the component that follows the base as it migrates from wages to assets.

Hosting and control. Everything runs in the government data centre in Apia with an encrypted replica off-island, on open-source software; no taxpayer-identifying data leaves Samoa; the structural balance, the buffer balance and the VAGST gap are published.

5.2 Cost, staffing and sequencing of the build

A four-person analytics cell across the Ministry for Revenue and the Ministry of Finance: a lead, a data engineer, an analyst with revenue-forecasting experience and an analyst with audit experience, supported for about 12 months by external help for the build and training. Regional pooling with Tonga and Vanuatu on customs reference prices, audit methods and the disaster loss model would lower cost for all three, since the same small-island problems recur. The software cost is negligible; the data agreements and the bank-access legislation are the work.

Sequencing: the forecasting workbench first, because the Central Bank and customs data exist and the structural figure can appear in the next budget; the VAGST gap estimate in parallel; the investment dashboard from the FMIS data; the risk engine on customs and provident-fund data while the bank legislation is prepared; the matching service when it passes; the buffer calculator as soon as the loss distribution is fitted; the property register last.

Limitations

• Elasticities and loss parameters are assumed; the Central Bank of Samoa and the Ministry of Finance hold the series needed to estimate them.
• The execution model treats capacity as a scalar. In practice procurement rules and donor conditions bind more than staff numbers on some projects.
• Bank-data matching depends on legislation and on the banks' willingness; correspondent-banking sensitivities in Samoa argue for narrow, well-governed access.
• The FY2026 dip may include effects I have not identified; the Article IV staff report should be read for the authorities' own explanation.

7.1 The limitations discussed

Assumed parameters. Every figure in the paper rests on parameters that Samoa's own data can estimate and this paper has not. The remittance elasticities, the border share, the congestion function, the marginal product of public capital and the disaster loss distribution are all specified with the data that identify them, and the paper's illustrative numbers are placeholders for those estimates. The order of magnitude of each result is likely to survive estimation; the levels are not to be quoted.

Capacity as a scalar. The execution model treats project-management capacity as a single number. In practice, procurement rules, donor conditions, land access and contractor availability each bind on different projects, and a project can stall for reasons that no pipeline cap addresses. The survival analysis proposed in section 8.4 would show which constraints dominate; the cap is a first-order instrument, not a complete one.

Bank data and confidentiality. The audit model's strongest source is also its most sensitive. Samoa's banks operate under correspondent-banking relationships that are alert to any expansion of state access to customer data, and a badly designed provision could cost more in de-risking than it yields in VAGST. The paper's design is narrow for that reason, and the Central Bank's role as regulator is central to making it acceptable.

The FY2026 dip. The paper's original account attributed the dip to remittances; the IMF's July 2026 statement attributes it to domestic demand, agriculture and fuel prices with remittances holding up. Section 8 revises the reading. The lesson for the model is that the decomposition must include energy-price and domestic-demand components, and that a single-cause explanation of any revenue year is likely to be wrong.

Small samples. Four disaster events, sixteen years of quarterly revenue and a project database since 2015 are small samples for the models proposed. Pooling with Tonga, Vanuatu and Fiji on the disaster and customs models is the practical remedy, and the regional cooperation the paper proposes is a statistical necessity as much as an administrative economy.

Extended modelling: improvements and further research

8.1 Research update, September 2026: what caused the slowdown

The IMF's July 2026 concluding statement and the September Board release explain the FY2026 dip more precisely than section 2 did. Growth of 0.4% in FY2025/26 reflects weak household consumption, subdued private investment, lower agricultural output, the fading effect of the 2024 CHOGM, and higher fuel and shipping costs; Brent reached about US$120 a barrel in April 2026. The Central Bank reports real GDP contracting 4.5% in the December 2025 quarter with commerce down sharply and tax revenue falling by about 12 million tala in the quarter. Tourism and remittances held up. So the revenue shock was a domestic-demand and energy-price shock, not a remittance shock, and the IMF now recommends a temporary fiscal expansion for vulnerable households and productive investment. This changes the reading of Model A: the elasticity of revenue to fuel prices and to domestic commerce belongs in the nowcast alongside remittances, and the structural-revenue rule must separate an energy-price shock (temporary, run down buffers) from a remittance shift (persistent, adjust the base).

8.2 Model A, nowcast and structural revenue: improvements

• Add energy and shipping cost drivers. Extend the elasticity regression with the landed fuel price and a freight index. Import duty and VAGST on fuel rise with price, while VAGST on everything else falls as households cut back; the net sign is an empirical question the FY2026 data can answer.
• Separate temporary and persistent components. Fit an unobserved-components model that decomposes revenue into trend, a remittance cycle, an energy-price component and noise. Only the trend enters the structural balance; the components determine whether the response is buffer drawdown or base adjustment.
• Quarterly nowcasting with mixed frequencies. Customs entries are daily, remittances monthly, GDP quarterly. A MIDAS or Bayesian structural time-series nowcast gives Treasury a four-week early warning of the kind that would have flagged the December 2025 quarter.
• Grant calendar as a state variable. Grants swing from 8.3% to 2.5% to 6.9% of GDP. Model committed and pipeline grants by donor and project with disbursement probabilities, and report the grant forecast as a distribution; the headline balance should be shown with and without the grant tail.

8.3 Model B, VAGST gap and audit: improvements

• Estimate the gap with two independent methods and reconcile. Top-down from national accounts and HIES, and bottom-up from a stratified random audit of 150 registrants. Report both and the reconciliation, as RA-GAP practice requires; a single method is not credible for a first estimate.
• Bank credits as the primary custodian source. With excess liquidity, deposit records are unusually complete. Model the ratio of bank credits to declared VAGST turnover by sector and firm size, with a robust benchmark; the residual is the shortfall score. This needs statutory access and access logging, but no new technology.
• Threshold bunching. Test for bunching of declared turnover just below the VAGST registration threshold; the excess mass gives the elasticity of declared turnover to registration and the size of the informal tail that is near-formal.
• Selection with a random share and propensity weighting, as in the companion papers, so that the semi-supervised layer learns the evasion pattern rather than the selector's habits.

8.4 Model C, public-investment execution: improvements

• Estimate the execution function on the project portfolio. The Ministry of Finance holds appropriation and disbursement by project since 2015. Fit execution rate as a function of concurrent project count, donor, procurement type and implementing agency; the coefficient on concurrent projects is the congestion parameter kappa assumed earlier.
• Survival analysis of project delay. Time from appropriation to first disbursement, with hazard depending on project size and procurement route, predicts which projects in the pipeline will stall and by how long.
• Link executed investment to the tax base. Estimate the marginal product of public capital theta with a growth regression on executed capital spending, instrumented by donor disbursement timing. This closes the loop between the revenue models and the investment case.

8.5 Model D, disaster buffer: improvements

• Calibrate on the 2009 tsunami and Cyclone Evan (2012). Monthly revenue around both events gives severity and recovery-curve estimates specific to Samoa.
• Joint modelling with the energy-price shock. FY2026 shows that non-disaster shocks also deplete revenue. A general fiscal buffer sized on the joint distribution of disaster and terms-of-trade losses is more useful than a disaster-only buffer, and the IMF's fiscal-space language supports it.
• Layered financing across cash buffer, contingent credit and parametric cover, optimised on PCRIC premium quotes, as in the Tonga paper.

8.6 New models proposed

• Fuel-tax incidence and the vulnerable-household transfer. With fuel prices high and a temporary expansion recommended, model the incidence of fuel excise and VAGST by household decile from the HIES, and design a transfer that offsets the burden on the bottom 40% at least cost.
• Property and rental income register, with a hedonic model on lease and listing data, to bring the assets remittances buy into the tax base as the PAYE base migrates.
• Labour-mobility fiscal model, as proposed for Tonga: PAYE lost per departing worker against VAGST gained from remittances spent, by sector and household type.

8.7 Further research

• A first published VAGST gap estimate with a reconciled top-down and bottom-up method.
• An unobserved-components decomposition of Samoa's revenue, 2010 to 2026, distinguishing remittance, energy and domestic-demand cycles.
• An execution-function estimate on the Ministry of Finance project database.
• Regional pooling with Tonga, Vanuatu and Fiji on customs reference prices and audit methods.

Sources for this section: IMF Staff Concluding Statement, 2026 Article IV, 29 July 2026; IMF Press Release 26/291, 16 September 2026; PINA, IMF warns Samoa's recovery has lost momentum, 3 August 2026; Central Bank of Samoa, Selected Economic Indicators, April 2026.

A. Worked example 1: the remittance channel and structural revenue

Assume tax revenue of about WST 726 million (22% of a GDP near WST 3.3 billion), of which 55% is collected at or near the border and moves with remittance-financed imports with an elasticity of 0.9.

Structural revenue: with remittances 8% above trend and arrivals 5% above, and a tourism-sensitive share of 15% of revenue with elasticity 0.6, structural revenue is about WST 696 million and the windfall about WST 30 million. The budget should be set on 696 and the 30 should go to the buffer.

B. Worked example 2: an illustrative VAGST gap

Final consumption WST 2,600 million; VAGST 15%; exemptions remove 18% of the base; collections WST 280 million.

A compliance gap of 12% is about WST 40 million a year. Closing a third of it through the audit model would fund the analytics cell many times over.

C. Worked example 3: a trader scored

A registered trader declares turnover of WST 1.2 million. Customs imports were WST 600,000 and the sector mark-up is 1.6; provident-fund contributions imply wages of WST 180,000 in a sector where wages are 14% of turnover; bank credits to the business account were WST 1.5 million. Each estimator's noise is taken as 0.12 in log terms.

Only positive z-scores count, so the score is 0.57 plus 1.86 weighted by source reliability. The pattern, low imports and high bank credits, is a service business whose revenue is not import-driven; the audit question is to reconcile WST 1.5 million of bank credits with WST 1.2 million of declared sales.

D. Worked example 4: execution and the future base

With a marginal product of public capital of 0.15 and a capital budget of WST 300 million, execution of 55% adds about WST 25 million of output a year and WST 5 million of tax; execution of 75% adds about WST 34 million and WST 7 million. The difference compounds across budgets.

E. Worked example 5: the disaster buffer

Event rate 0.25 a year, median loss per event 7% of annual revenue, log standard deviation 0.8, annual loss capped at 70%, 200,000 simulated years.

A buffer of 15% of revenue, filled at 2.5% of revenue a year plus half of periodic windfalls, reaches target in about five years.

F. Algorithms in full

F.1 Structural revenue and nowcast

ALGORITHM Structural_Revenue
QUARTERLY:
  fit ln R[k,t] = a[k] + bM[k] ln M[t] + bA[k] ln A[t] + bY[k] ln Y[t] + bF[k] ln Fuel[t] + g[k] D[t] + e  on rolling 10-year window
  trend M, A by HP filter or local linear trend
  R_struct[k] = R[k] * (M_trend/M)^bM * (A_trend/A)^bA
  windfall = max(0, sum R - sum R_struct)
MONTHLY:
  nowcast quarter revenue from weekly customs, monthly remittances and arrivals (MIDAS or Bayesian structural time series); report band
ANNUALLY:
  publish headline and structural balance; buoyancy by stream against structural GDP

F.2 VAGST gap and audit selection

ALGORITHM VAGST_Gap_And_Select
TOP-DOWN: VTTL from imports by HS code + HIES + government consumption; compliance and policy gaps
BOTTOM-UP: stratified random audit of 150 registrants; scale; reconcile with top-down
FOR each registrant i:
  Qhat[m,i] for m in {imports*markup, bank credits, wages/labour share, utilities/intensity}
  z[m,i] = (ln Qhat[m,i] - ln Qdec[i]) / sigma[m,s]
  H[i] = 1 if input credits claimed > suppliers' declared output tax to i, or sales < imports with no stock build
  A[i] = sum_m omega[m] max(0, z[m,i]) + gamma d_MCD(ratios_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 EY s.t. auditor hours by skill; random share >= 10%; selection rate by island within [0.5, 2] of base
ROUTE: identity breaches under cap -> letter; ranked -> officer; large -> senior review
AFTER audit: reweight omega by Bayesian update on observed turnover

F.3 Pipeline cap

ALGORITHM Pipeline_Cap
ESTIMATE e = C/(C + kappa n) on project data since 2015, with agency and donor effects
FIT survival model of appropriation-to-first-disbursement time
FOR target execution e*: n_max = C (1 - e*) / (kappa e*)
QUARTERLY: report pipeline n against n_max; predicted execution; projects at risk of delay from survival model

F.4 Buffer rule

ALGORITHM Buffer_Rule
FIT event rate lambda (trend allowed), severity distribution, recovery curve on 2009, 2012, 2018, 2019 events
SIMULATE annual loss L; B_star = quantile_0.95(L) * R_bar - expected insurance payout above that level
QUARTERLY: contribution = c R + psi windfall while B < B_star; drawdown only on declared disaster up to L R
EVERY 3 YEARS: refit; republish B_star
OPTIMISE layering: cash buffer to 95th percentile; parametric cover and PRF above, chosen on premium quotes

G. Data specification for Samoa

H. Sensitivity analysis

I. Validation protocol

• Back-test the structural revenue model on 2015 to 2025, checking that it would have flagged the FY2024 windfall and the FY2026 shortfall in advance.
• Reconcile the top-down and bottom-up VAGST gap estimates and publish both with the reconciliation.
• Hold out each compliant trader in turn to check the reconstruction bands.
• Run random audits at 10% and compare hit rates with model-selected cases.
• Back-test the execution function on projects since 2015.
• Refit the disaster distribution every three years and publish the buffer target.

References

• International Monetary Fund. Press Release 26/291: Executive Board Concludes 2026 Article IV Consultation with Samoa. 16 September 2026.
• International Monetary Fund. Press Release 25/023: Executive Board Concludes 2024 Article IV Consultation with Samoa. 31 January 2025.
• International Monetary Fund. Press Release 23/84: Executive Board Concludes 2023 Article IV Consultation with Samoa. 20 March 2023.
• Methods: Keen, M. (2013), The Anatomy of the VAT, IMF WP 13/111; Hutton, E. (2017), RA-GAP, IMF TNM 17/04; Cummins and Mahul (2009), Catastrophe Risk Financing in Developing Countries, World Bank. Method citations from memory; verify before publication.