Introduction

Kiribati's fiscal position is unusual even by the standards of this series. Fiscal deficits of 22% of GDP in 2024 and 14% in 2025 would signal a crisis in almost any other country in this project; in Kiribati they are close to business as usual, financed by design through withdrawals from the RERF, a fund established in 1956 from phosphate-mining royalties on Banaba and preserved through independence specifically to give a resource-poor, land-scarce nation a permanent income stream once the phosphate itself ran out. The fund has operated under a return-based withdrawal rule intended to stabilise its real per-capita value, and the IMF's own language in its most recent mission statement is a direct warning about that rule's fragility: weak 2025 financial-market returns could jeopardise the withdrawal already budgeted for 2026, precisely the scenario a well-designed rule should be tested against before, not after, it happens.

Two further pressures sit alongside the RERF question. The copra subsidy, which supports outer-island households producing the country's traditional coconut-based export, rose from 6% to 9% of GDP in the 2025 budget — an enormous figure for a single subsidy line — and the IMF's assessment was that this increase was balanced by improved excise tax collection. That is a specific, testable fiscal claim, not an established fact, and this paper's Model B is built to test it rather than repeat it. Separately, fishing licence revenue, built on Kiribati's exceptionally large exclusive economic zone relative to its tiny land area, is described in the IMF's own recent reporting as in trend decline and subject to considerable volatility tied to global financial and commodity conditions, at the same time as the Fund recommends gradually reducing SOE and other tax exemptions and raising excise rates further.

The research question this paper takes up has three parts: how fragile the RERF's return-based withdrawal rule actually is under a realistic sequence of poor investment years, and what adjustment mechanism should trigger before the fund's real value erodes; whether the copra-subsidy/excise pairing the IMF described as balanced genuinely holds when modelled explicitly, and what it means for outer-island households if it does not; and how fishing licence revenue's trend decline should be distinguished from its year-to-year volatility in Kiribati's own budget planning, so a bad year is not mistaken for a permanent one, or a good year for a recovery. As throughout this series, the organising identity for the tax-side question splits the gap between potential and actual excise and SOE-related revenue into a policy component and a compliance component:

R^{pot}_{excise+SOE,t} - R_{excise+SOE,t} = \underbrace{\left(R^{pot}_{excise+SOE,t} - R^{law}_{excise+SOE,t}\right)}_{\text{policy gap: SOE exemptions, excise rate design}} + \underbrace{\left(R^{law}_{excise+SOE,t} - R_{excise+SOE,t}\right)}_{\text{compliance gap}}

Current tax issues

2.1 A fund financing deficits most countries could not sustain

RERF withdrawals financed a fiscal deficit of 22% of GDP in 2024 and 14% in 2025. These are financing decisions, not fiscal failures in the ordinary sense — the fund exists precisely to be drawn on — but a return-based withdrawal rule only works if it actually constrains withdrawals when returns disappoint, and the IMF's own concern, that weak 2025 returns could jeopardise the 2026 budgeted withdrawal, is a direct statement that the rule's binding constraint has not yet been tested by a real shortfall.

2.2 A subsidy that grew faster than most tax bases could

The copra subsidy rising from 6% to 9% of GDP in one budget cycle is, in absolute terms, one of the largest single fiscal-policy shifts documented anywhere in this series. The IMF's characterisation, that improved excise tax collection balanced the increase, is worth testing on its own terms: excise collection improving is not the same as excise collection improving by enough, and the two are rarely announced with the matching precision a genuine offset would require.

2.3 Fishing revenue: trend decline layered on volatility

Kiribati's fishing licence revenue depends on its exclusive economic zone, one of the largest in the world relative to its land area, and on global tuna markets and the terms Kiribati can negotiate, bilaterally and through the Parties to the Nauru Agreement Vessel Day Scheme, with distant-water fishing nations. The IMF's language distinguishes two separate phenomena worth separating analytically: a trend decline, and volatility around whatever the trend is in a given year. Budgeting against the wrong one, treating a volatile bad year as a trend, or a trend decline as merely a bad year, produces the wrong policy response either way.

2.4 SOE and other exemptions still on the table

The IMF's 2025 consolidation recommendations explicitly include gradually reducing SOE and other tax exemptions and raising excise rates further, alongside further rationalising copra subsidies and streamlining SOE subsidies. These sit on the policy side of the revenue-gap identity above: not evasion, but statutory relief the government has chosen to grant and could choose to unwind.

2.5 Debt low, but distress risk high

Public debt fell to 8% of GDP in 2025, low in absolute terms, but the IMF still assesses the risk of debt distress as high, because of long-term climate-related vulnerabilities rather than current borrowing levels. This is a materially different debt story from Fiji's or Solomon Islands' papers in this series: the constraint is not today's debt service but the fiscal space climate adaptation investment will eventually require, layered on top of the RERF and subsidy pressures already described.

2.6 A currency account gap that compounds the picture

The current account deficit reached 18.8% of GDP in 2025, assessed by the IMF as substantially weaker than fundamentals would imply, with government spending itself contributing to high import demand. This is downstream of, not separate from, the fiscal choices in sections 2.1-2.4: RERF-financed spending and subsidy transfers both flow substantially into imported consumption in an economy with very limited domestic production capacity.

Models

3.1 Model A: RERF return-based withdrawal rule under stress

The fund evolves under its own return-based rule, where the permitted withdrawal in year t is tied to a moving average of recent real returns rather than a fixed share of the fund's value, the standard design for a rule meant to smooth withdrawals across good and bad years:

RERF_{t+1} = RERF_t(1+\rho_t) - W_t, \qquad W_t = w \cdot \overline{\rho}_{t-k:t-1} \cdot RERF_t

where rho_t is the realised real return in year t, w is a policy withdrawal multiplier, and the k-year trailing average return governs the permitted draw. This paper stress-tests the rule against a sequence of below-median returns rather than the single-year shortfall the IMF's language implies, since fund-depleting risk compounds across consecutive poor years:

\rho_t \sim \mathcal{N}(\mu_\rho, \sigma_\rho^2), \qquad \text{stress path: } \rho_t = \mu_\rho - \sigma_\rho \text{ for } t \in \{t_0, t_0+1\}

The rule's adequacy is judged by whether the fund's real per-capita value, RERF_t / P_t (P the population), is preserved or eroded over the stress path relative to a baseline of average returns, and by how large an adjustment — either a temporarily higher contribution from fiscal surplus years or a temporarily lower w — would be needed to restore it:

\Delta w^{restore} : \quad \frac{RERF_{t_0+5}}{P_{t_0+5}}\bigg|_{w-\Delta w} \geq \frac{RERF_{t_0}}{P_{t_0}}

3.2 Model B: copra subsidy and excise fiscal-neutrality and incidence

Model B tests the IMF's balanced characterisation directly rather than assuming it:

\text{NetFiscalShift}_t = \big(\text{ExciseRevenue}_t - \text{ExciseRevenue}_{t-1}\big) - \big(\text{CopraSubsidy}_t - \text{CopraSubsidy}_{t-1}\big)

A NetFiscalShift near zero supports the balanced characterisation; a materially negative figure means the subsidy increase outpaced the excise offset regardless of how the two were described together. Incidence is tracked separately, since even a genuinely balanced pair can still redistribute: excise falls predominantly on imported consumption goods, which weighs more heavily, in relative terms, on households without significant copra income, while the subsidy itself flows to outer-island copra-producing households specifically:

\eta^{excise}_{Q40} = \frac{\sum_{h \in Q40} w_h\, x^{excise}_h}{\sum_h w_h\, x^{excise}_h}, \qquad \eta^{copra}_{outer} = \frac{\text{Copra subsidy received by outer-island households}}{\text{Total copra subsidy}}

A fiscally neutral pairing can still be a redistribution from urban, import-dependent households to outer-island copra producers, or the reverse; Model B reports both the net fiscal position and this distributional split rather than treating aggregate neutrality as the end of the question.

3.3 Model C: fishing licence revenue decomposition and smoothing

Model C decomposes the fishing-licence revenue series into a slow-moving trend and a residual volatility component, using a standard structural decomposition rather than a simple moving average, so trend and cycle are separated formally rather than by eye:

F_t = T_t + C_t + \varepsilon_t, \qquad T_t = T_{t-1} + \beta_t, \qquad \beta_t = \beta_{t-1} + \eta_t

a local linear trend model (T the level, beta the evolving trend slope, C a cyclical component, estimated via a standard state-space filter). Budget planning then responds to T_t, the trend, for baseline revenue projections, and sizes a smoothing buffer to absorb C_t, the cyclical deviation, using the same compound-loss buffer logic as the disaster models elsewhere in this series but calibrated to revenue volatility rather than physical loss severity:

\text{SmoothingBuffer}^{target} = z_{0.9} \cdot \sigma_C

where sigma_C is the estimated standard deviation of the cyclical component and z_0.9 a coverage multiplier chosen to smooth all but the most severe downside years, consistent with how much of the shortfall the budget can reasonably self-insure rather than needing donor or RERF support to cover.

Algorithms and illustrative results

4.1 Algorithm 1: RERF withdrawal rule stress test

INPUT  RERF_0, mu_rho, sigma_rho, k-year trailing window, withdrawal multiplier w
FOR each simulated path (baseline: iid returns; stress: k consecutive years at
  mu_rho - sigma_rho, otherwise iid):
  val <- RERF_0
  trailing_returns <- [mu_rho] * k   # seed with expected return
  FOR t in 1..horizon:
    rho_t <- draw or stress value for this path
    avg_return <- mean(trailing_returns[-k:])
    W_t <- w * avg_return * val
    val <- val * (1 + rho_t) - W_t
    trailing_returns.append(rho_t)
    RECORD val, W_t
COMPARE baseline vs stress paths at horizon: erosion = 1 - stressed_val/baseline_val
IF erosion > tolerance:
  SOLVE for adjustment (temporary Delta_w reduction, or a top-up contribution rule)
  that restores real per-capita value within a specified recovery horizon

4.2 Algorithm 2: fishing revenue trend/cycle decomposition and buffer sizing

INPUT  Historical fishing licence revenue series F_1..F_T
FIT local linear trend state-space model (level T_t, slope beta_t, cycle C_t)
  via Kalman filter / maximum likelihood
FOR each year t: DECOMPOSE F_t into T_t (trend), C_t (cycle), residual
SET baseline budget revenue for t+1 to the trend forecast T_hat_{t+1}, not the
  most recent observed F_t
ESTIMATE sigma_C from the fitted cyclical component's historical standard deviation
SET smoothing buffer target = z_0.9 * sigma_C
ANNUALLY: contribute to the buffer when C_t > 0 (above-trend years), draw down
  when C_t < 0 (below-trend years), so the general budget is insulated from the
  cyclical component while the trend is tracked and reported separately

4.3 Illustrative results: RERF under a two-year return shock

Using illustrative parameters (5% expected real return, 10% return volatility, a 3-year trailing-average withdrawal rule, and a withdrawal multiplier chosen so baseline withdrawals roughly track the fund's expected growth), a ten-year simulated path with two consecutive below-median return years inserted early erodes the fund's real per-capita value by roughly 26% relative to a path without the shock, by year ten. The mechanism is compounding rather than the two bad years alone: because the withdrawal rule uses a trailing average, the depressed returns keep suppressing permitted withdrawals for several years after the shock itself has passed, but the fund's value has already fallen further behind the no-shock path than the return shortfall alone would suggest, since withdrawals continued (at a reduced but positive rate) throughout the low-return years rather than pausing. This is the concrete version of the IMF's own concern: the rule moderates but does not eliminate the risk a bad run of years poses to the fund's long-run value, and this paper's Algorithm 1 gives Kiribati's Ministry of Finance a way to quantify, rather than only anticipate, how much of an adjustment a real shock of this kind would require.

4.4 Illustrative arithmetic: the copra-excise pairing

On Kiribati's 2025 nominal GDP of roughly AUD 498-525 million, a subsidy increase from 6% to 9% of GDP represents an additional AUD 15-16 million a year, illustratively. For the IMF's balanced characterisation to hold in the NetFiscalShift sense of Model B, excise revenue would need to have risen by a comparable amount over the same period — a figure this paper does not have access to independently and which Kiribati's own Ministry of Finance and Economic Development budget documents should supply before the balanced characterisation is treated as confirmed rather than asserted.

Solutions, software, data science and AI models

5.1 Issue-to-solution map

5.2 Model catalogue

5.3 Reference architecture

flowchart TD
  A[Sources: RERF financial statements, MFED budget records, Customs/excise collections, fishing licence agreements, SOE exemption register] --> B[Ingestion and data quality]
  B --> C[Warehouse: bitemporal, entity linked]
  C --> D[RERF policy-lab simulator]
  D --> E[MFED and RERF board dashboard]
  C --> F[Copra/excise net-shift tracker]
  C --> G[Fishing revenue trend/cycle model]
  G --> H[Revenue forecasting and buffer tool]
  C --> I[SOE exemption register]

The RERF simulator and fishing-revenue model both run on an annual budget-cycle cadence; the copra/excise tracker updates whenever new collection data is available.

5.4 Software stack

5.5 Governance of AI use

• The RERF stress test informs, it does not set, policy. The RERF Board and Ministry of Finance retain the withdrawal decision; Model A's role is to make the risk visible before it materialises.
• The copra/excise finding is reported regardless of outcome. If NetFiscalShift comes out materially negative, that result should be reported to Cabinet in the same terms as a neutral or positive result, not adjusted after the fact.
• The exemption register is public-facing where possible. Consistent with this series' general governance principle, a published exemption register supports the accountability the IMF's phase-out recommendation implies.
• No automated trigger. Even where Model A's stress test suggests an adjustment is needed, the adjustment itself — a lower withdrawal multiplier, a temporary surplus contribution rule — is a fiscal-policy decision for Cabinet and the RERF Board, not an automatic rule change.

5.6 Team and indicative effort

Given Kiribati's very small scale, a two-person function within the Ministry of Finance and Economic Development — one analyst covering the RERF and fiscal-modelling work, working with existing MFED budget and Customs data — is proportionate. All three models in this paper can be built and run with open-source tools at modest computational cost; the binding constraint is analyst time and access to RERF's own detailed return history, not infrastructure. Cost and staffing figures are estimates.

Limitations

• The RERF stress-test parameters in section 4.3 (expected return, volatility, withdrawal multiplier) are illustrative; RERF's actual asset allocation, return history and the precise mechanics of the current return-based rule are held by the RERF Board and Ministry of Finance and should replace these inputs before the model informs an actual withdrawal decision.
• The copra/excise NetFiscalShift calculation in section 4.4 uses approximate GDP figures to illustrate the arithmetic; it is not a substitute for the actual excise and subsidy line items in Kiribati's budget accounts, which this paper does not have access to.
• The fishing-revenue trend/cycle decomposition (Model C) needs a reasonably long historical series to estimate reliably; with a short series, the trend and cycle components will be poorly identified and the model's early estimates should be treated with corresponding caution.
• This paper does not model the fishing-licence revenue's negotiation dimension directly — the terms Kiribati secures through PNA and bilateral agreements are a policy and diplomatic question as much as a data-science one, and Model C takes the resulting revenue series as given rather than modelling the negotiation itself.
• Climate-related debt-distress risk (section 2.5) is named but not separately modelled in this paper; it interacts with all three models here (a climate shock could simultaneously depress RERF returns, strain the fishing-revenue base, and require new adaptation borrowing) but a fully joint model of that interaction is future work.

References

• International Monetary Fund. Kiribati: IMF Staff Completes 2026 Article IV Consultation. 4 March 2026.
• International Monetary Fund. IMF Executive Board Concludes 2025 Article IV Consultation with Kiribati. Press Release 25/247, 10 July 2025.
• International Monetary Fund. Kiribati: IMF Completes 2025 Article IV Consultation. Press Release 25/161, 27 May 2025.
• International Monetary Fund Country Report No. 25/[xxx]. Kiribati: 2025 Article IV Consultation — Staff Report.
• Ministry of Finance and Economic Development, Kiribati. 2026 Budget Speech.
• Parties to the Nauru Agreement (PNA). Vessel Day Scheme documentation.
• Methods: as elsewhere in this series — Hutton, E. (2017), The Revenue Administration Gap Analysis Program, IMF Technical Notes and Manuals 17/04; Harvey, A. C. (1989), Forecasting, Structural Time Series Models and the Kalman Filter, Cambridge University Press, for the trend/cycle decomposition in Model C. Method citations are from memory; verify before publication.