Introduction

Fiji is not short of tax administration capacity by Pacific standards. The Fiji Revenue and Customs Service (FRCS) runs a self-service online filing platform (TPOS, live since December 2019), an electronic VAT Monitoring System with accredited fiscal devices at the point of sale, and posted a record $3.51 billion in net revenue for the year to July 2026, 4.0% above forecast. The problem is not collection machinery. It is that fiscal policy gave back a large share of the base just as the buffer to absorb the next shock was most needed, and that public debt, at 80% of GDP, is still working down from a pandemic-era peak near 90%.

The VAT rate cut from 15% to 12.5% in the FY2025-26 budget reduced tax revenue by an estimated $250 million and helped push the budgeted overall deficit to around 6.5% of GDP before the 2026 oil-price shock arrived. The IMF's May 2026 Article IV mission was explicit: restore the 15% rate, hold the size of government steady, and rebuild buffers, because Fiji remains exposed to cyclones and other shocks with little fiscal room to respond. Fiji's own history makes the point concrete: Cyclone Winston in February 2016 was the most intense storm ever recorded in the Southern Hemisphere at landfall, causing damage estimated near 20% of GDP; Cyclones Yasa (2020) and Ana (2021) followed within five years.

The research question is therefore not whether Fiji can collect tax, but how much of the VAT cut should be reversed, how fast, with what compensation for the households who feel it most, and how the proceeds should be split between paying down debt and building a reserve against the next cyclone. A second, quieter question follows from FRCS's own infrastructure: with VMS already recording point-of-sale transactions, what is the most direct way to detect the specific form of evasion that infrastructure creates an opportunity to catch, which is suppression or manipulation of the device record itself, as distinct from the under-declaration risk that dominates tax administrations without electronic point-of-sale coverage.

As in the Vanuatu, Tonga and Timor-Leste papers in this series, the organising identity splits the gap between potential and actual VAT revenue for year t into a policy component and a compliance component:

R^{pot}_{VAT,t} - R_{VAT,t} = \underbrace{\left(R^{pot}_{VAT,t} - R^{law}_{VAT,t}\right)}_{\text{policy gap: the rate cut and remaining exemptions}} + \underbrace{\left(R^{law}_{VAT,t} - R_{VAT,t}\right)}_{\text{compliance gap: suppression at the point of sale}}

Current tax issues

2.1 A buffer given away before it was rebuilt

Fiji's FY2025-26 budget cut the VAT rate from 15% to 12.5%, at an estimated cost of $250 million in revenue — a large figure against a total FRCS collection of $3.51 billion for the year. The cut was popular and disinflationary in the short run; the IMF credited VAT and duty reductions with a sharp fall in headline inflation through 2025. But it arrived just as debt, at 80% of GDP, remained elevated after the pandemic-era peak near 90%, and just before a 2026 oil-price shock widened the current account deficit and put growth on a lower path. The 2026 Article IV mission's assessment was direct: the FY2025-26 budget "leaves the government with very little room to handle unexpected needs for extra spending," and rebuilding buffers should come through growth-friendly deficit reduction, including restoring the VAT rate, rather than through further spending cuts alone.

2.2 Debt still elevated, reserves still thin

Public debt has fallen from its FY2020-21 peak of 89.8% of GDP but remains at 80%, and external reserve coverage has been falling even as the current account position nominally improved. This matters for how a VAT reversal should be used: revenue recovered has at least two competing claims on it, debt reduction and a disaster buffer, and a model that ignores the debt claim will overstate how much can go into a reserve.

2.3 A VAT Monitoring System that most Pacific peers do not have

Unlike Tonga, Vanuatu or Timor-Leste, Fiji already operates a VAT Monitoring System: accredited electronic fiscal devices at the point of sale, with an FRCS accreditation table for compliant equipment. This changes the shape of Fiji's compliance problem. Where Tonga's tax is collected mostly at the wharf and the evasion risk is under-valuation of imports, Fiji's VAT is collected mostly at the till, and the risk shifts to the integrity of the device record itself: sales suppression through "zapper" software that deletes or rewrites transactions after the fact, phantomware that runs a second, unrecorded register, or simple failure to issue a device receipt for cash sales. FRCS's own CIT data shows over 35,000 taxpayers file corporate returns; VMS coverage of the retail and hospitality population that generates most consumption-tax risk is the more relevant base for this paper's compliance model.

2.4 Disaster exposure with a short return period

Fiji sits inside one of the most cyclone-exposed belts in the Pacific. Cyclone Winston (February 2016, Category 5 at landfall) caused damage estimated near 20% of GDP and is among the most intense storms ever recorded in the Southern Hemisphere. Cyclone Yasa followed in December 2020 and Cyclone Ana in January 2021, both within the same five-year window. A return period this short means a buffer sized to a 1-in-20-year loss, the convention used in Tonga's paper in this series, may still be tested more than once in a working career, and any fiscal rule needs to say explicitly how draws are replenished before the next event.

2.5 Governance and the record itself

Where Tonga's governance risk sits in officer discretion over exemption certificates, Fiji's analogous risk sits in the device record: a compliant-looking VMS receipt is only as trustworthy as the software and hardware chain that produced it. FRCS's own accreditation process is the first control; a monitoring model that can detect a device that has quietly stopped behaving like its accredited peers is the second, and is what Model B below is built to provide.

Models

3.1 Model A: VAT reversal costing and incidence

Model A answers the IMF's instruction in reverse of Tonga's Model A: instead of pricing what removing an exemption would raise, it prices what restoring a rate would raise, and who would feel it. The revenue effect of moving the rate from tau-0 (12.5%) toward tau-1 (up to 15%) on base B, allowing consumption to respond through the compensated own-price elasticity epsilon, is:

\Delta R(\tau_1) = (\tau_1 - \tau_0) \cdot B \cdot \left(1 - \varepsilon \cdot \frac{\tau_1 - \tau_0}{1+\tau_0}\right)

B is built from FRCS's own domestic and import VAT collection data disaggregated by commodity group, so the base does not need to be re-estimated from survey data the way Tonga's did. Because Fiji's 2023 VAT increase and 2025 VAT decrease are both recent, real, and administratively documented events, epsilon itself is estimable directly from Fiji's own collection time series around each rate change, rather than assumed from elsewhere in the region — a natural experiment Tonga and Vanuatu do not have.

Incidence follows the same logic as Tonga's targeting ratio: the share of the extra tax borne by the bottom two income quintiles Q40, from household expenditure survey weights w and spending x on VAT-bearing goods:

\eta = \frac{\sum_{h \in Q40} w_h\, x_h}{\sum_h w_h\, x_h}

A partial-restoration rule sets tau-1 below 15% and funds a compensating transfer sized to the burden on Q40, so the net decision is a triple: the rate chosen, the split of proceeds between debt reduction and the buffer (Model C), and the size of any offsetting transfer.

\Gamma = \Delta R(\tau_1) \cdot \eta, \qquad \Delta R^{net} = \Delta R(\tau_1) \cdot (1-\eta) - \Gamma^{admin}

where Gamma-admin is the administrative cost of running the transfer, which for a rate change (unlike an exemption removal) is usually small if an existing cash-transfer channel, such as the Social Pension Scheme or Poverty Benefit Scheme, is used rather than a new one built for this purpose.

3.2 Model B: VMS sales-suppression and device-integrity detection

Model B scores each accredited device, and the taxpayer it belongs to, for the statistical signature of suppression rather than for a single flagged transaction, since a device that deletes records leaves a gap the device's own log cannot show directly. Three independent tests are combined.

Digit-conformity test. Genuine transaction values, aggregated over enough sales, follow Benford's law in their leading digits; a device or business systematically rounding or fabricating entries deviates from the expected first-digit distribution P(d) = log10(1 + 1/d):

\chi^2_i = \sum_{d=1}^{9} \frac{\left(O_{i,d} - N_i P(d)\right)^2}{N_i P(d)}

where O is the observed count of leading digit d for device i over N transactions in the window.

Sequence-integrity test. Accredited devices issue sequentially numbered receipts; a gap in the sequence not matched by a corresponding void or refund record is a hard flag rather than a statistical one:

H_i = \mathbb{1}\!\left[\text{receipt sequence for } i \text{ has an unexplained gap in window } w\right]

Peer-ratio test. Sales-to-declared-rent, sales-to-covers (for restaurants), or sales-to-floor-area ratios are compared against the peer distribution for the same business type and location, using the same robust median/MAD construction as the customs model in Tonga's paper, since a device that is technically compliant but simply under-reports volume will not fail the digit or sequence tests but will sit far below its peers:

z^{peer}_i = \frac{\ln \tilde{s}_{g} - \ln s_i}{1.4826 \cdot MAD_g}

The combined device risk score, and the escalation rule:

A_i = \omega_1 \chi^2_i + \omega_2 \max(0, z^{peer}_i) + \delta H_i, \qquad \text{escalate if } A_i > A^{*} \text{ or } H_i = 1

Escalation means a physical device audit, not an automatic assessment: this model flags candidates for FRCS's existing VMS accreditation and compliance team to inspect, consistent with the governance principle in section 5.5.

3.3 Model C: disaster buffer under a binding debt constraint

Model C extends Tonga's compound-Poisson buffer model with a second state variable, the debt ratio, since Fiji cannot treat buffer-building as costless while debt remains at 80% of GDP and the IMF is asking for consolidation. Disaster losses follow the same structure as Tonga's model:

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)

but lambda is calibrated to Fiji's own three-events-in-five-years recent history rather than assumed, giving a materially higher event rate than Tonga's illustrative 0.3/year. Recovered VAT revenue Delta-R-net from Model A is split each period between debt amortisation and the buffer by a rule that tightens as debt approaches a target ceiling d* (say, 70% of GDP, consistent with pre-pandemic norms) and relaxes once debt is comfortably below it:

\phi_t = \min\!\left(1,\ \max\!\left(0,\ \frac{D_t - d^{low}}{d^{*} - d^{low}}\right)\right), \qquad \text{Debt paydown}_t = \phi_t \cdot \Delta R^{net}_t, \qquad \text{Buffer contribution}_t = (1-\phi_t)\cdot \Delta R^{net}_t

phi_t is the share of recovered revenue directed to debt paydown; it is 1 while debt is at or above d*, falls linearly as debt approaches d-low (say, 60% of GDP), and is 0 once debt is comfortably below target, at which point the buffer receives the full amount. The buffer itself evolves as in Tonga's model, drawing down only on a declared disaster:

B_{t+1} = B_t(1+r) + \text{Buffer contribution}_t - \mathbb{1}[\text{disaster}_t]\min(B_t, L_t R_t)

The debt state evolves alongside it, and the model's output is not a single buffer target but a joint path for (D_t, B_t) under alternative choices of tau_1 (Model A) and d* — the question Fiji's fiscal authorities actually face, which is how to sequence debt reduction and self-insurance rather than treating them as separable.

Algorithms and illustrative results

4.1 Algorithm 1: VAT reversal and split rule

INPUT  FRCS VAT base B by commodity group, elasticity eps estimated from the 2023
       increase and 2025 decrease, HIES weights w and spending x, debt ratio D_t,
       target d_star, floor d_low
FOR candidate rates tau_1 in {13.0%, 13.75%, 14.5%, 15.0%}:
  dR      <- (tau_1 - tau_0) * B * (1 - eps * (tau_1-tau_0)/(1+tau_0))
  eta     <- Q40 expenditure share on VAT-bearing goods from HIES
  Gamma   <- dR * eta
  dR_net  <- dR * (1 - eta) - admin_cost
  phi     <- clip((D_t - d_low) / (d_star - d_low), 0, 1)
  debt_paydown     <- phi * dR_net
  buffer_contrib   <- (1 - phi) * dR_net
  REPORT tau_1, dR, Gamma, debt_paydown, buffer_contrib
SELECT tau_1 balancing consolidation speed against near-term compensation cost

4.2 Algorithm 2: VMS device risk scoring

FOR each accredited device i, rolling window w:
  digits      <- leading digits of all transaction values for i in w
  chi2_i      <- Benford chi-square statistic on digits
  gaps        <- receipt sequence numbers for i in w
  H_i         <- 1 if an unexplained gap exists, else 0
  s_i         <- sales-to-[rent|covers|floor-area] ratio for i
  s_tilde,MAD <- robust centre/spread of s across peer group g (business type, location)
  z_peer_i    <- (ln(s_tilde) - ln(s_i)) / (1.4826 * MAD)
  A_i         <- w1*chi2_i + w2*max(0, z_peer_i) + delta*H_i
  IF A_i > A_star OR H_i = 1: escalate to physical device audit
  ELSE: retain in monitoring pool, recompute next window
MONTHLY: rank escalated devices by A_i for the compliance team's queue
AFTER each audit finding: recalibrate w1, w2, delta by logistic regression on outcomes

4.3 Illustrative results: disaster buffer under Fiji's event rate

Inputs are assumed from Fiji's own recent history, not estimated from a full loss series: event rate 0.6 a year (three named cyclones causing material fiscal impact in 2016-2021), median loss per event 8% of annual revenue, log standard deviation 0.85, annual loss capped at 90%. The run uses 200,000 simulated years.

A buffer sized to the 1-in-20-year loss is about 31% of annual revenue, roughly double Tonga's illustrative figure in this series, reflecting Fiji's shorter observed return period. At a $3.5 billion revenue base, that is on the order of $1.1 billion — larger than is plausible to hold in cash. The practical reading is not "build a $1.1 billion buffer" but that Fiji's self-insurance should cover only the more frequent, lower layers (up to perhaps the 90th percentile) in cash, with the tail transferred to the Pacific Resilience Facility, parametric cyclone insurance, and contingent credit lines, consistent with the layered-financing logic flagged as further research in Tonga's paper and formalised here as a design choice rather than an afterthought.

4.4 Illustrative arithmetic: VAT reversal

Assume, illustratively, an own-price elasticity of -0.3 for the VAT-bearing consumption basket and a Q40 expenditure share on VAT-bearing goods of 0.42 (VAT bears more heavily on lower-income households in relative terms because saving rates rise with income). Restoring the full 2.5-percentage-point cut:

\Delta R \approx 0.025 \times B \times \left(1 - 0.3 \times \frac{0.025}{1.125}\right) \approx 0.025 \times B \times 0.993

Against an approximate VAT base of $2.0 billion, this yields roughly $49-50 million per percentage point restored, or close to the $250 million originally given up if the full 2.5 points are reversed. With eta at 0.42, about $105 million of that would need to flow back to Q40 households as compensation to hold them harmless, leaving roughly $145 million net — before any split between debt paydown and the buffer under Model C. This is arithmetic illustrating the model's mechanics, not a costed government estimate; B, epsilon and eta all need estimation from FRCS and Fiji Bureau of Statistics microdata before any number here is used in a budget submission.

Solutions, software, data science and AI models

Fiji's solution centres on TPOS and VMS, the infrastructure FRCS already runs, rather than on building new collection channels from scratch.

5.1 Issue-to-solution map

5.2 Model catalogue

5.3 Reference architecture

flowchart TD
  A[Sources: VMS transaction logs, TPOS returns, HIES, FRCS collection series, cyclone loss records] --> B[Ingestion and data quality]
  B --> C[Warehouse: bitemporal, device and taxpayer linked]
  C --> D[Device scoring service: Benford test, sequence check, peer ratio]
  D --> E[VMS compliance queue]
  E --> F[Audit findings]
  F --> D
  C --> G[Policy lab: VAT reversal costing, debt/buffer split, disaster simulation]
  G --> H[Treasury and FRCS dashboards]

The device scoring service runs nightly against the prior day's VMS logs; the policy lab runs on demand for budget preparation.

5.4 Software stack

5.5 Governance of AI use

• Shadow mode first. Run device scoring alongside the existing VMS compliance selection for one full cyclone season before switching over, and publish the comparison of hit rates.
• No automated penalties. A high A_i score queues a device for physical audit; only the audit finding, not the score, supports any assessment.
• Explainable flags. Each escalated device shows its chi-square statistic, its peer ratio and reference group, and any sequence gap, so the compliance officer and the taxpayer can both see the basis for selection.
• Recalibration under review. Weight updates from audit outcomes are proposed by the model and approved by the compliance team before going live, not applied automatically.

5.6 Team and indicative effort

A small analytics function inside FRCS — one lead, one data engineer, one analyst working across VMS and TPOS data, and one Treasury liaison for the debt/buffer modelling — is realistic given FRCS's existing TPOS and VMS platforms already do the data capture. External build support of six to nine months is a reasonable estimate, shorter than in administrations building point-of-sale monitoring from nothing. Cost and staffing figures are estimates.

Limitations

• The $250 million VAT-cut cost and the elasticity and incidence figures in section 4.4 are illustrative arithmetic, not FRCS or Ministry of Finance estimates; the first deliverable is measurement, using FRCS's own collection series around the 2023 and 2025 rate changes.
• The disaster-loss calibration in section 4.3 uses three named events as an informal anchor, not a fitted historical loss series; a proper calibration needs Ministry of Finance revenue data for the months immediately following Winston, Yasa and Ana specifically.
• The debt-anchor thresholds (d* = 70%, d-low = 60% of GDP) are illustrative reference points based on pre-pandemic norms, not a Fiji-specific debt-sustainability analysis, which the IMF's own Article IV work is better placed to supply.
• The VMS device-scoring model assumes accredited devices report reliably enough to compute peer ratios; for very small or seasonal businesses, thin transaction volumes will make the Benford and peer-ratio tests unreliable, and the sequence-integrity check should carry more weight for that segment.
• Compensating transfers assume the Social Pension Scheme and Poverty Benefit Scheme reach the households a VAT restoration would burden most; coverage gaps, particularly for informal and rural households, need checking against FRCS and Ministry of Social Welfare data before the incidence estimate in section 4.4 is relied on.

References

• International Monetary Fund. IMF Executive Board Concludes 2026 Article IV Consultation with Fiji. Press Release 26/178, 29 May 2026.
• International Monetary Fund. IMF Executive Board Concludes 2025 Article IV Consultation with Fiji. Press Release 25/208, 20 June 2025.
• International Monetary Fund. IMF Staff Completes 2024 Article IV Mission to Fiji. Press Release 24/83, 11 March 2024.
• Fiji Revenue and Customs Service. Annual Report 2023-2024. Parliamentary Paper 161/25.
• Fiji Revenue and Customs Service. FRCS Records Highest Annual Revenue Collection to Date. August 2026.
• Fiji Revenue and Customs Service. VAT Monitoring System accreditation table, frcs.org.fj.
• Fiji Times. "IMF: Bring back 15% VAT." 2026.
• Methods: as in the Tonga paper in this series — 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; Nigrini, M. (2012), Benford's Law: Applications for Forensic Accounting, Auditing, and Fraud Detection, Wiley; Cummins, J. D. and Mahul, O. (2009), Catastrophe Risk Financing in Developing Countries, World Bank. Method citations are from memory; verify before publication.