Introduction

Few countries have had so little need to tax. Since 2005 the Petroleum Fund has financed more than 80% of each budget, and the Fund now stands near nine times non-oil GDP. Domestic revenue has been an afterthought, raised through low flat income taxes, a narrow sales tax on imports, a services tax on a few sectors, and excise.

That period has ended. The question for this paper is what domestic revenue path, combined with what spending path, keeps the Fund from running out, and how the January 2027 VAT should be designed so that it contributes as much as a new tax in a low-capacity administration realistically can.

The organising identity splits the revenue gap for tax k into policy and compliance components:

R^{pot}_{k} - R_{k} = \underbrace{\left(R^{pot}_{k} - R^{law}_{k}\right)}_{\text{policy gap: no VAT, low rates}} + \underbrace{\left(R^{law}_{k} - R_{k}\right)}_{\text{compliance gap}}

Timor-Leste adds a third feature that the other Pacific cases lack: the tax base is itself a function of Fund withdrawals, because public spending drives non-oil GDP.

Current tax issues

2.1 Withdrawals far above sustainable income

The Petroleum Fund Law defines Estimated Sustainable Income (ESI) as 3% of petroleum wealth. On a Fund of about US$18 billion and no remaining field receipts, ESI is roughly US$0.54 billion. The 2025 budget was about US$2.6 billion. The IMF's 2025 mission proposed a 2026 central government envelope of about US$1.85 billion. Excess withdrawals are legal with parliamentary justification and have become routine.

2.2 No broad consumption tax

The indirect system taxes imports at low rates and domestic services in a few sectors. Domestic value added in retail, construction and most services is untaxed. The government has committed to a VAT by January 2027, a date both the IMF and World Bank cite. At the date of this paper that is under four months away, with enabling law, taxpayer registration, IT, refund procedures and public education all on the critical path. A slipped or under-prepared launch is the main near-term risk.

2.3 Thin direct tax base that recycles public money

Income tax rates are low and flat (about 10%; confirm against the Taxes and Duties Act). The formal private sector is small, and the largest taxpayers are contractors to government and donors. Domestic revenue is therefore partly Fund money returning through the tax system. When spending is cut, as it must be, tax falls with it. Any projection that holds the base fixed while cutting spending overstates revenue.

2.4 Upstream uncertainty

Bayu-Undan has ceased production. Greater Sunrise remains undeveloped, with first gas not expected before 2032 to 2035. Whatever fiscal terms apply, the state's take will depend on auditing multi-billion-dollar development costs. That capacity takes years to build and must exist before first gas.

2.5 Wage bill and recurrent pressure

The IMF flags a sharply rising public wage bill and recommends containing recurrent expenditure while protecting health, education and capital spending. Revenue that is absorbed by recurrent pay does not move the depletion date.

2.6 Administration

The Tax Authority has limited experience with credit-invoice taxes, refunds or risk-based audit. A large share of commerce is cash-based, and the currency is the US dollar, so there is no inflation tax or exchange-rate adjustment to cushion a fiscal cliff. Adjustment would fall directly on spending.

Models

3.1 Model A: stochastic Petroleum Fund dynamics

Model A gives the depletion date as a probability distribution, not a single year. With balance F, stochastic return r, petroleum receipts P (zero until Greater Sunrise), spending G and domestic revenue D:

F_{t+1} = F_t\,(1 + r_t) + P_t - \left(G_t - D_t\right), \qquad r_t \sim N\!\left(\bar r,\, \sigma_r^2\right), \qquad T^{*} = \min\{t : F_t \le 0\}

Greater Sunrise enters as a risky option: with probability p-S the project reaches first gas in year t-S, after which the state receives its fiscal take on production q at price p less costs c:

P_t = \mathbb{1}\!\left[\text{Sunrise}\right] \cdot \mathbb{1}\!\left[t \ge t_S\right] \cdot \gamma \left(p_t q_t - c_t\right), \qquad \Pr(\text{Sunrise}) = p_S

The statutory benchmark is ESI, and the economically meaningful condition is that the non-oil primary deficit stays within the real return net of population growth n:

ESI_t = 0.03 \left( F_t + PV(P) \right), \qquad G_t - D_t \;\le\; \left(\bar r - \pi - n\right)\left(F_t + \mathbb{E}[PV(P)]\right)

3.2 Model B: VAT revenue, compliance ramp and threshold

Steady-state VAT revenue is the rate times final consumption times C-efficiency, where C-efficiency is the product of policy efficiency (how much consumption the law reaches) and compliance:

V_t = \tau \cdot C_t \cdot CE_t, \qquad CE_t = \left(1 - g^{pol}\right) \cdot c_t

A new VAT does not reach steady-state compliance at launch. Compliance follows a logistic ramp from launch date t-0 with half-life t-half:

c_t = \frac{c_{\infty}}{1 + e^{-\kappa\,(t - t_0 - t_{1/2})}}

Net revenue must subtract the taxes VAT replaces (import sales tax, services tax) and provide for refunds to exporters and investors:

\Delta R_t = V_t - R^{sales}_t - R^{services}_t - \text{Refunds}_t

The registration threshold is the most consequential design choice for a low-capacity administration. Following Keen and Mintz, the optimal turnover threshold z-star balances revenue from the marginal firm against administration cost A and compliance cost Gamma, with delta the marginal value of public funds and v the ratio of value added to turnover:

z^{*} = \frac{\delta A + \Gamma}{\left(\delta - 1\right) \tau\, v}

High administration cost and a low rate both push the threshold up. With firm-size distribution f(z), the share of potential revenue retained above a threshold is:

\Omega(z^{*}) = \frac{\int_{z^{*}}^{\infty} z\, v\, f(z)\, dz}{\int_{0}^{\infty} z\, v\, f(z)\, dz}

In economies where a few hundred importers, contractors, telecoms, banks and hotels produce most formal value added, omega stays above 85% even at a high threshold. A high threshold costs little revenue and makes the launch manageable.

3.3 Model C: tax base with spending recycling

Domestic revenue across taxes k depends on non-oil GDP, which depends on public spending through a multiplier m:

D_t = \sum_k \tau_{k,t}\, b_{k}\, Y^{N}_t\, c_{k,t}, \qquad Y^{N}_t = Y^{N}_{t-1}\left(1+g\right) + m\left(G_t - G_{t-1}\right)

The implication is a revenue leak from consolidation. Each dollar of spending cut lowers the deficit by less than a dollar:

\frac{\partial \left(G_t - D_t\right)}{\partial G_t} = 1 - m \sum_k \tau_k b_k c_k

With a multiplier of 0.6 and a domestic tax ratio of 12% of non-oil GDP, the leak is about 7 cents per dollar. The leak grows as the tax ratio rises, so it matters more after VAT than before.

3.4 Model D: the required revenue and spending path

Model D works backwards from the goal. Choose paths for the domestic revenue ratio d and spending ratio g that minimise adjustment cost, subject to a survival probability for the Fund and a cap on how fast the tax ratio can rise given administrative capacity:

\min_{\{d_t, g_t\}} \sum_t \beta^{t} \left[ a\,(\Delta d_t)^2 + b\,(\Delta g_t)^2 \right] \quad \text{s.t.} \quad \Pr\!\left(T^{*} > 2060\right) \ge 0.9, \quad 0 \le \Delta d_t \le 0.75\%\; Y^{N}, \quad g_t \ge g^{min}

The floor g-min protects health, education and capital spending. Because survival is a probability from Model A, the problem is solved by simulation: propose a path, run the Monte Carlo, check the constraint, adjust.

3.5 Model E: audit selection for a new VAT and for upstream costs

Timor-Leste has almost no audit history, so selection must be label-free. For VAT, each registrant's declared turnover is tested against customs imports with a sector mark-up, government payment records (the state is the largest customer and knows what it paid each contractor), and bank credits:

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

The government payment record is the strongest source available to any tax office in the region: a contractor cannot declare less turnover than the Treasury paid it. For Greater Sunrise, the same logic applies to development costs: benchmark declared capital cost per unit of capacity against independent engineering cost indices and partner-country filings before costs enter the recovery pool.

Algorithms and illustrative results

4.1 Algorithm 1: Monte Carlo depletion

INPUT  F0, mean return rbar, volatility sigma, spending path G[t],
       revenue path D[t] from Model C, Sunrise probability pS, start year tS, take P[t]
FOR sim = 1 .. 20000:
  sunrise <- Bernoulli(pS)
  F <- F0
  FOR t = 2026 .. 2070:
    r <- Normal(rbar, sigma)
    P <- take[t] if sunrise and t >= tS else 0
    F <- F * (1 + r) + P - (G[t] - D[t])
    IF F <= 0: record T*[sim] = t; BREAK
RETURN median, 5th and 95th percentile of T*, and Pr(T* > 2060)

4.2 Algorithm 2: required-path solver

INITIALISE d[t] = current tax ratio, g[t] = current spending ratio
REPEAT:
  D[t], Y[t] <- Model C given d, g (recycling included)
  surv <- Pr(T* > 2060) from Algorithm 1
  IF surv >= 0.9: BREAK
  raise d[t] by up to 0.75% of non-oil GDP per year where VAT ramp allows
  lower g[t] toward g_min in proportion to cost weight b / a
UNTIL converged
RETURN paths d*, g*, and the implied annual withdrawal

4.3 Illustrative results

Inputs are mine, not official: opening balance US$18.0 billion, mean nominal return 4.5%, domestic revenue US$0.22 billion growing 5% a year, no Sunrise receipts. The baseline lands within a year of the government's 2037 estimate, which gives some comfort on calibration.

With return volatility of 8% a year, 20,000 simulations of the baseline give a median depletion year of 2036 and a 5th to 95th percentile range of 2033 to 2039. Good investment returns buy three years at most.

Reading. Restraint buys three years; VAT buys two more. A VAT raising about 4% of non-oil GDP net is a good outcome by regional standards and still covers only about 5% of the financing gap. Survival to 2060 without Sunrise requires withdrawals to fall toward US$0.6 to 0.7 billion, which means both a domestic revenue ratio rising toward 20% of non-oil GDP over 10 to 12 years and spending well below the current envelope. Model D exists to find the least painful version of that path.

4.4 VAT design parameters implied by Model B

Rate and targets are illustrative; calibrate on recent VAT launches in comparable economies.

Solutions, software, data science and AI models

Timor-Leste has an advantage the other two lack: it is building a VAT from a blank page, so digital controls can be designed in from launch instead of retrofitted. 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/>tax system, ASYCUDA, FMIS payments, e-invoices, Central Bank Fund data, payroll] --> B[Ingestion and data quality]
  B --> C[Warehouse<br/>bitemporal, single taxpayer view]
  C --> D[Fiscal lab<br/>Fund Monte Carlo, path solver, VAT microsimulation]
  D --> E[Budget documents and public simulator]
  C --> F[Risk engine<br/>Treasury matching, reconstruction, invoice graph]
  F --> G[Refund gate]
  F --> H[Audit allocator and case management]
  H --> I[Audit outcomes]
  I --> F
  C --> J[Upstream audit workbench]
  C --> K[Tax-law assistant<br/>three languages]

The fiscal lab serves the Ministry and Parliament; the risk engine serves the Tax Authority; both read the same governed warehouse.

5.4 Software stack

5.5 Governance of AI use

• Transparency of the fiscal model. Publish the simulator code and assumptions. A depletion date that anyone can reproduce is harder to argue away than one from a consultant's spreadsheet.
• Human decision. Risk scores select cases and hold refunds for review; officers assess. Refund holds carry a statutory time limit so that honest exporters are not starved of cash.
• Explainability. The primary audit score is a sum of record-based terms, led by the Treasury payment mismatch, which a contractor can verify against its own bank statements.
• Random audit share. A fixed share of audits and refund checks is random, to measure the compliance gap of the new VAT from year one.
• Data protection. Payroll and civil-registry linkage needs a legal basis and access logging; the language assistant never receives taxpayer-identifying data.

5.6 Team and indicative effort

Two small units are needed: a fiscal modelling unit of three to four in the Ministry of Finance, and a risk and analytics unit of five to six in the Tax Authority. External support of 18 to 24 months is realistic given the VAT launch. Set against a financing gap above US$1 billion a year, the cost of both units is negligible; the scarce input is trained Timorese staff, so the programme should fund postgraduate data science training from the start. Staffing and duration are my estimates.

Limitations

• Scenario inputs are illustrative. The Ministry of Finance and the Central Bank hold the actual Fund return assumptions, domestic revenue series and spending plans.
• The fiscal multiplier for Timor-Leste is uncertain and import leakage is high, so m may be well below 0.6, which would shrink the recycling effect.
• Returns are modelled as independent normal draws. Real portfolios have fat tails and the Fund's equity share makes a bad early sequence more damaging than the average suggests.
• The model ignores borrowing. Concessional loans can smooth the path but do not change the intertemporal arithmetic.
• Greater Sunrise is treated as a single probability. Its fiscal terms, processing location and cost are all unsettled.

Extended modelling: improvements and further research

8.1 Research update, September 2026: the VAT date has moved

The January 2027 launch assumed in sections 2 and 3 is no longer credible. On 18 September 2026 Commission C of the National Parliament held its first public seminar on a draft VAT law that has not yet been published; the draft still has to pass the Council of Ministers before it reaches Parliament, and the Minister of Finance intends a public awareness campaign before Cabinet approval. The government still says 2027, and is creating a dedicated authority to administer the tax. In July 2026 the Tax Authority Commissioner led a 20-person delegation to Indonesia's Directorate General of Taxes to study VAT policy design, business processes, IT architecture and compliance risk management. Read together: the design work is serious and the law is late. A realistic launch is the second half of 2027 or 2028. For Model A this shifts the VAT ramp by one to two years and moves the illustrative depletion date in the restraint-plus-VAT scenario from about 2041 to about 2040. For Model B it means the design window is open: the threshold, rate and refund rules argued for below can still be written into the law.

8.2 Model A, Fund dynamics: improvements

• Replace normal returns with a block bootstrap of the Fund's own history. The Petroleum Fund has published quarterly returns since 2005, spanning 2008, 2020 and 2022. Resample blocks of eight quarters to preserve autocorrelation and fat tails; the 5th percentile depletion year will move earlier than the normal-return estimate.
• Model the equity share as a policy variable. The Fund's allocation between equities and bonds changes expected return and variance together. Run the depletion distribution across allocations and report the trade-off: a higher equity share raises the median depletion year and widens the lower tail.
• Endogenous withdrawals. Excess withdrawals have been approved every year. Model the withdrawal as a function of the previous year's shortfall and an election-cycle dummy, estimated on 2010 to 2025 data, rather than as an exogenous path. The projected depletion date under estimated behaviour is the honest baseline; the policy scenarios are deviations from it.
• Greater Sunrise as a decision tree with information arrival. First gas is uncertain in both timing and terms. Model the final investment decision as a node whose probability is updated on observable milestones (processing-location decision, financing, partner commitments), so that each budget can report a current probability rather than a fixed 2032 to 2035 range.

8.3 Model B, VAT design: improvements

• Calibrate the compliance ramp on documented launches. The half-life parameter was assumed. Fit the logistic ramp to C-efficiency paths from VAT introductions in comparable economies (Cambodia 1999, Lao PDR 2010, Fiji 1992, Papua New Guinea 1999, Bangladesh 1991, Vanuatu 1998) using IMF Government Finance Statistics and national data; report the fitted half-life distribution and use its median.
• Estimate the firm-size distribution before choosing the threshold. Business registration and income-tax turnover data give the empirical distribution of turnover. Fit a Pareto tail and compute the retained-revenue share at candidate thresholds directly, replacing the assertion that a few hundred firms hold most value added with a measured share.
• Refund risk as a queueing and fraud model. Refund claims arrive at a rate that depends on export and investment activity; fraudulent claims cluster in the first two years. Model the refund queue with a time limit and a risk gate, and simulate cash-flow exposure under alternative carry-forward periods, so that the six-month rule proposed earlier is chosen on numbers.
• Price incidence at launch. Estimate expected pass-through by product category from the Indonesian and Cambodian experience, and model the one-off CPI effect and its interaction with the dollarised economy, where there is no exchange-rate absorption.
• Interaction with the sales tax it replaces. Net revenue is gross VAT less the abolished import sales tax and services tax. Model the transition month by month, because there is a period in which the old taxes have stopped and the new one is not yet yielding.

8.4 Models C and D, base recycling and required path: improvements

• Estimate the multiplier. The fiscal multiplier in a dollarised, import-dependent economy is unknown. Use a local-projection estimate on quarterly public spending and non-oil GDP, instrumented by budget appropriations, to obtain it; a multiplier near 0.3 would halve the recycling leak assumed earlier.
• Solve the required path as a stochastic control problem. Replace the simulation search with dynamic programming over (Fund balance, tax ratio, spending ratio), with the survival constraint as a penalty. This gives a feedback rule, withdrawal as a function of the current balance, that the budget law can adopt directly, rather than a fixed path that is obsolete after one bad year.
• Add a welfare floor. Spending on health, education and cash transfers is fixed at a floor; the optimiser adjusts capital and wage spending. Report the depletion date under the floor and the cost of raising it.

8.5 Model E, audit selection: improvements

• Treasury payment matching as a hard constraint, not a score. A contractor's declared turnover cannot be below what the Treasury paid it. Encode this as an identity breach that generates an automatic query, with the difference assessed unless the contractor explains it. This is the highest-yield, lowest-skill control available to a new administration.
• Withholding at source on government payments. Model the revenue and compliance effect of a VAT and income-tax withholding on all government payments to vendors, with credit against final liability. In an economy where the state is the largest customer, this captures most of the formal base on day one.
• Upstream cost-audit benchmark. Before Sunrise, assemble a capex-per-tonne and opex-per-barrel-equivalent benchmark from public project data in Australia, Papua New Guinea and Southeast Asia, so that the cost-recovery pool can be tested at first gas.

8.6 New models proposed

• Readiness index. A checklist-scored readiness model (law, register, IT, refunds, education, staffing) with historical launch outcomes from other countries, giving a probability that a launch on a given date succeeds; use it for the go or no-go decision.
• Wage-bill dynamics. Model public employment and average pay as a function of the budget envelope and election cycles, and its effect on non-oil GDP and tax; the IMF's flagged wage-bill growth is the main competitor for the revenue VAT raises.
• Public fiscal simulator design. A single-page application with the Fund model compiled to JavaScript, letting Parliament and citizens move withdrawal, VAT and Sunrise assumptions and see the depletion distribution. Publishing the model is itself a fiscal-governance instrument.

8.7 Further research

• A published block-bootstrap depletion distribution using the Central Bank's actual return series.
• The empirical firm-size distribution and the retained-revenue curve for VAT thresholds.
• A fitted compliance ramp from comparable VAT launches.
• A local-projection estimate of the fiscal multiplier in Timor-Leste.
• Post-launch: a registered evaluation design comparing declared turnover of VAT registrants matched to Treasury payments before and after launch.

Sources for this section: VATCalc, Timor-Leste VAT implementation 2027 and the 18 September 2026 parliamentary seminar; Indonesia DGT, Timor-Leste Tax Authority study visit, July 2026.

References

• Ministry of Finance, Timor-Leste. General State Budget 2026 Report: Investing in National Transformation. October 2025.
• International Monetary Fund. IMF Staff Complete 2025 Article IV Mission to Timor-Leste. 2 July 2025.
• International Monetary Fund. IMF Concludes 2026 Article IV Mission to Timor-Leste, as republished by Mirage News. August 2026.
• World Bank. Timor-Leste Economic Report, December 2022.
• Bertelsmann Stiftung. BTI 2026 Timor-Leste Country Report.
• Timor-Leste's financial cliff draws closer in 2025. East Asia Forum, 27 January 2025.
• Methods: Keen, M. and Mintz, J. (2004), The optimal threshold for a value-added tax, Journal of Public Economics 88; Keen, M. (2013), The Anatomy of the VAT, IMF Working Paper 13/111; Ebrill, L. et al. (2001), The Modern VAT, IMF. Method citations are from memory; verify before publication.