Geopolitical Shock Transmission in Thailand: A Narrative SVAR–CGE Framework for Macroeconomic and Distributional Analysis

Abstract

Geopolitical shocks affect small open economies through multiple, correlated channels, yet applied CGE analyses typically impose the timing and persistence of those shocks by assumption. This paper develops a two-stage SVAR–CGE framework that links econometrically identified shock dynamics to general-equilibrium welfare evaluation for Thailand. First, a seven-variable narrative SVAR estimated on monthly data for 2000–2025, identified using the Caldara–Iacoviello Geopolitical Risk Index, is used to recover the persistence of five transmission channels: oil prices, shipping costs, exchange rates, tourism demand, and private investment. Second, these estimated persistence parameters discipline the shock paths in a 22-sector recursive comparative-static CGE model calibrated to Thailand’s 2025 Social Accounting Matrix and simulated over three annual periods using a present-value integral transformation. Under the baseline shock bundle, GDP declines by 3.18% and CPI increases by 5.49%, with welfare losses exhibiting a bimodal distributional pattern—largest for Q1 through consumption-share exposure and for Q4 through tradeable-sector intensity—departing from the monotonically regressive pattern in single-channel analyses. Policy simulations show that targeted transfers calibrated to income rank dominate a universal fuel subsidy on fiscal efficiency, welfare effectiveness (welfare multiplier 1.377 vs. 0.334), and progressivity (1.00 vs. 0.94), at half the fiscal cost (1.48% vs. 2.97% of baseline GDP). An additional bimodal-targeting scenario (S4) at identical fiscal cost underperforms income-rank targeting on all metrics, confirming the latter as the robust second-best instrument under LES preferences with strong MPC heterogeneity. These rankings are supported by the central calibration of a 9-point sensitivity grid, with partial corroboration at off-baseline configurations. The paper contributes by showing that empirically disciplining inter-annual shock dynamics in CGE analysis can materially alter policy conclusions under correlated multi-channel external shocks, shifting the preferred response from sector-specific price subsidies toward demand-side household transfers.

1. Introduction

On 28 February 2026, US–Israeli military strikes against Iran triggered the effective closure of the Strait of Hormuz—through which approximately 20 percent of the world’s oil transits (International Energy Agency, 2026); Brent crude exceeded $100 per barrel by mid-March 2026. Prior simulations of US–Iran conflict scenarios using the War Oil Crisis (WOC) Simulator projected oil price increases ranging from 50% under low-intensity conflict to over 100% under high-intensity scenarios, starting from a baseline of USD 58 per barrel (Ruiz Estrada et al., 2020); the 2026 episode exceeded the upper bound of these projections within the first month. For Thailand, which imports nearly 80 percent of its crude oil and relies on natural gas for two-thirds of its electricity generation, the shock extends well beyond an energy-price disturbance: it simultaneously activates shipping disruptions, exchange-rate depreciation, tourism contraction, and investment uncertainty—and the structural distinction between oil and gas price transmission matters for channel design. Evidence from energy-dependent Gulf Cooperation Council (GCC) economies—where geopolitical proximity to conflict zones amplifies cross-market spillovers—confirms that natural gas price shocks transmit through different financial and commodity channels than oil price shocks, with spillover intensity varying by market and episode type (Al-Maadid et al., 2026). This heterogeneity motivates the present paper’s channel-specific treatment of energy shocks rather than a composite energy-price index that obscures the distinct propagation paths of oil and gas disturbances.

Beyond this technical motivation, the economic burden of this compound disturbance is unlikely to be uniform across sectors, factors of production, and household groups—raising the central questions this paper addresses: how the shock propagates through each channel, how its costs are distributed, and which policy instruments most efficiently limit welfare losses.

A key structural feature of compound geopolitical shocks is that their constituent channels differ substantially in timing, persistence, and decay rate. Energy-cost effects emerge within weeks; exchange-rate pass-through to consumer prices may take months; and investment responses operate with a longer lag still, as firms revise capital-expenditure plans in response to sustained uncertainty. The observed 3–6% depreciation of the Thai baht against the US dollar in the weeks following the onset of the conflict (Bank of Thailand, 2026) illustrates the speed of the financial channel relative to the real economy. These cross-channel differences imply that the aggregate welfare impact cannot be assessed from static or single-channel analyses: what matters for policy design is not only the initial magnitude of the disturbance but also how quickly each channel decays and how long its effects persist.

A growing literature links geopolitical uncertainty to oil-market disruptions, financial-market volatility, and investment contractions, with effects documented to be severe for commodity-importing, trade-dependent economies (Caldara & Iacoviello, 2022). A parallel strand of CGE-based research has evaluated the economy-wide and distributional consequences of energy-price increases. These two literature; however, remain weakly integrated: time-series studies identify shock dynamics without tracing sectoral and welfare incidence, while CGE studies model propagation without empirically disciplining shock timing and persistence.

Beyond this integration gap, two important limitations persist in the existing literature. First, empirical analyses of geopolitical shocks in emerging economies remain predominantly focused on the oil-price channel, with comparatively limited attention to the simultaneous transmission through shipping costs, exchange-rate adjustment, tourism demand, investment expectations, and input-cost pass-through. Studies that do address multiple channels typically treat them in isolation rather than within a unified framework, making it difficult to assess their relative magnitudes and interactions. Second, and more critically for dynamic policy analysis, CGE-based assessments of external shocks almost invariably impose the persistence and decay structure of shocks exogenously. The assumed rate at which a shock dissipates over time is rarely estimated from historical data, even though it can materially alter simulated outcomes for GDP, sectoral output, investment, and household welfare. As a result, recursive comparative-static simulation studies often rely on arbitrary persistence assumptions that reduce the empirical credibility of their inter-annual shock profiles.

This paper asks three research questions. First, how do geopolitical oil-price and shipping-cost shocks propagate through Thailand’s exchange rate, tourism, and investment channels, and what are their relative magnitudes? Second, how persistent are these transmission effects, and what implications do estimated persistence parameters have for disciplining the inter-annual shock paths in a recursive comparative-static CGE simulation? Third, how do multi-channel external shocks affect macroeconomic aggregates, sectoral output, and household welfare across the expenditure distribution?

To answer these questions, the paper develops a two-stage empirical framework. In the first stage, a Narrative Structural Vector Autoregression (SVAR) is estimated over January 2000–December 2025, using the Caldara–Iacoviello Geopolitical Risk (GPR) Index to identify the dynamic responses of Thailand’s key macroeconomic variables—producer prices, consumer prices, and the exchange rate—to external oil-price and shipping-cost shocks. A critical identification distinction governs how the SVAR output is used. The 2026 geopolitical episode exceeds the historical range of GPR observations used in estimation; accordingly, the SVAR does not determine the magnitude of the 2026 shock, which is instead calibrated directly from observed 2026 data using narrative identification in the tradition of Romer and Romer (2004). What the SVAR supplies is the temporal decay structure—the speed, lag profile, and persistence of each identified transmission channel—derived from the full distribution of historical geopolitical episodes in the estimation sample. This separation between shock magnitude (calibrated from the 2026 episode) and shock persistence (estimated from historical dynamics) is the key methodological feature that prevents the counterfactual simulation from depending on the arbitrary assumption that a crisis of unprecedented scale decays at a historically average rate. Three operational roles follow from this separation: the SVAR identifies the relative timing, persistence, and transmission hierarchy of shocks across channels; scenario magnitudes are calibrated from observed 2026 market conditions; and the CGE model translates these disciplined shock paths into sectoral and household welfare outcomes.

This paper’s contribution is to discipline inter-annual shock dynamics in CGE analysis using empirically derived persistence structures. Three dimensions elaborate on this contribution. First, the SVAR supplies an econometrically disciplined persistence and transmission structure: a data-validated causal hierarchy of channels (primitive vs. propagated), a genuine SVAR-estimated persistence for the investment channel (${\rho }_{\mathrm{INV}}=0.468$), and an empirically grounded lower bound (${\rho }_{\mathrm{FLOOR}}=0.30$) for the remaining four channels derived from minimum observed geopolitical episode durations—reducing arbitrariness in shock profiling without overstating what the SVAR can deliver for tail episodes. Second, the analysis incorporates shipping costs, exchange-rate adjustment, tourism demand, and investment responses alongside oil prices within a unified framework, allowing their relative magnitudes and interactions to be assessed simultaneously rather than in isolation. Third, by incorporating five expenditure quintiles, the paper documents a bimodal distributional incidence pattern—Q1 most exposed through consumption-share effects and Q4 through tradeable-intensity effects—that departs from the monotonically regressive pattern in prior single-channel analyses (Jongwanich & Park, 2011; Warr, 2008), and derives transfer design implications from the empirically identified rather than assumed incidence profile.

1.1. What Time-Series and SVAR Studies Can Do—And Where They Stop

The macroeconomic literature on geopolitical risk has established three robust findings. Geopolitical disturbances generate compound external shocks whose consequences depend critically on their source: Baumeister and Kilian (2016) show that supply-driven and demand-driven oil price movements produce different macroeconomic outcomes; Kilian and Murphy (2014) and Kilian (2009) establish that geopolitical supply disruptions compress output and raise inflation simultaneously in net-importing economies. Mahmoudinia et al. (2025) confirm the principle in an SVAR context closely analogous to the present study: separating oil structural shocks and tracing their impulse responses through food and non-food prices demonstrates that shock extraction from data, prior to downstream simulation, is essential for credible welfare analysis. Building on this shock-identification tradition, the Caldara and Iacoviello (2022) GPR Index documents that geopolitical risk spikes generate significant declines in real activity and investment, with non-linear amplification during actual conflict episodes; cross-quantilogram evidence confirms that GPR–commodity price linkages strengthen substantially during high-volatility periods (Liu et al., 2025).

More recently, the Russia–Ukraine conflict demonstrated that geopolitical disturbances transmit through multiple simultaneous pathways: an energy-price channel raising production costs and inflation (Lippi & Nobili, 2012; Sun et al., 2024); a supply-chain and logistics channel propagating through production networks conditioned on supplier diversification (Joyez, 2025); and an uncertainty and financial-spillover channel generating heterogeneous cross-market volatility (Chowdhury et al., 2025; Shang & Hamori, 2025). These pathways are interdependent: energy price increases raise logistics costs, logistics disruptions constrain trade adjustment, and uncertainty amplifies cross-market volatility. Their aggregate effect therefore depends on joint configuration rather than on channel-by-channel magnitudes.

These multi-channel features extend directly to the Thai case. For Thailand specifically, the structure engages five transmission mechanisms with distinct adjustment speeds: energy costs (weeks), exchange-rate pass-through (months), shipping and supply-chain disruptions (concurrent with trade), tourism demand (lagged by security perception), and private investment (delayed by capital-expenditure revision). Jongwanich and Park (2011) document that commodity price shocks disproportionately burden lower-income Thai households, while Warr and Kohpaiboon (2007) provide the foundational CGE analysis of external price shocks for Thailand on which the present study builds.

The limitation of this literature is not in shock identification but in its downstream reach. Time-series and SVAR studies identify persistence, transmission hierarchy, and channel magnitudes with precision—but do not by themselves trace identified shocks through sectoral structures and household welfare distributions. The persistence parameters remain in the econometric system; they do not discipline the structural simulations that would be needed to evaluate policy.

1.2. What CGE Studies Can Do—And Where They Stop

The computable general equilibrium framework fills this gap in distribution and propagation. Following Lofgren et al. (2002), CGE models decompose aggregate welfare effects into equivalent variation by household group, tracing the transmission chain from shock through sectoral output, factor returns, and household income to quintile-level welfare outcomes. Prior CGE work establishes that external shocks generate heterogeneous welfare incidence shaped by market structure, expenditure patterns, and factor market flexibility (Diao & Kennedy, 2016; Mukashov et al., 2026; Somé & Zidouemba, 2025). For ASEAN economies, evidence consistently documents regressive commodity-shock incidence driven by food and energy expenditure shares (Imai et al., 2014; Warr, 2008), with targeted transfer programmes outperforming universal fuel subsidies on fiscal efficiency and distributional grounds (Coady et al., 2019; World Bank, 2022).

The limitation of CGE-based assessments is not in propagation mechanics—which are well-developed—but in the treatment of the shock itself. Recursive CGE studies are designed to trace inter-annual adjustment paths, yet the persistence and decay structure of the exogenous disturbance are overwhelmingly imposed by scenario design rather than identified from data. Sun et al. (2025) specify short-, medium-, and long-run scenarios without empirical discipline on shock persistence; Sun et al. (2024) impose scenario blocks from external assumption; and Wing (2008), Antimiani et al. (2013), and Hosoe et al. (2010) calibrate shock persistence to historical averages rather than to episode-specific dynamics. Even the systematic risk profiling framework of Mukashov et al. (2026)—which explicitly criticises ad hoc scenario analysis—improves shock attribution without disciplining temporal shock evolution through an external macroeconometric system.

The consequence is material. A front-loaded oil shock, a hump-shaped uncertainty response, and a long-lived volatility disturbance are treated identically under assumption-based calibration. Yet these three profiles produce different inter-annual consequences for capital accumulation, sectoral adjustment, and welfare incidence once timing differs.

1.3. The Missing Link—And What This Paper Does

Existing studies either identify macro shocks without welfare analysis or evaluate welfare without empirically grounded shock dynamics. This paper bridges these two strands by introducing a framework that uses empirically derived persistence structures to discipline inter-annual CGE simulations under multi-channel geopolitical disturbances—one that is honest about the limits of identification for tail events while still delivering more credible shock profiles than pure assumption.

To the authors’ knowledge, no published study has applied SVAR-derived AR(1) persistence parameters to discipline inter-annual shock decay in a recursive dynamic CGE model calibrated to a developing-country SAM under a correlated multi-channel geopolitical disturbance. This study situates this claim against three nearby strands rather than as a universal priority statement. Dynamic global CGE models—McKibbin and Wilcoxen’s G-Cubed (McKibbin & Wilcoxen, 1999) and GTAP-Dyn (Hertel et al., 2007; Ianchovichina & Walmsley, 2012)—embed persistence through forward-looking expectations or capital accumulation, not through externally estimated channel-specific ${\rho }_j$. Mukashov et al. (2026) improve shock attribution; the present paper complements this by disciplining shock decay paths. Elberry et al. (2026) and Koutsandreas et al. (2026) soft-link CGE to energy systems models; this analysis uses a macroeconometric system instead. The specific gap is the operational use of channel-specific SVAR persistence parameters as the linkage variable in recursive dynamic CGE simulation under multi-channel external transmission for a developing-country case.

The use of SVAR output to discipline subsequent structural models—rather than as a standalone forecasting exercise—follows a developing literature on SVAR-calibrated structural frameworks for advanced economies (Berger & Vavra, 2015; Fornaro & Wolf, 2023); the present paper extends this approach to a geopolitical episode in a developing economy, where the external source of discipline is persistence structure rather than demand estimation. Related model-linking studies establish architectural feasibility without providing this linkage: Elberry et al. (2026) and Koutsandreas et al. (2026) connect CGE models to externally generated information, but their external source is a technology-rich energy model rather than an econometric system identifying macroeconomic shock dynamics.

A tail-event identification concern merits acknowledgment. The 2026 Strait of Hormuz closure has no direct historical precedent in the SVAR estimation sample. The two-step design addresses this: persistence parameters are estimated from the full 2000M01–2025M12 sample spanning episodes of varying severity, while initial shock magnitudes are calibrated to observed 2026 data (Brent above $100, BDI surge, baht depreciation of approximately 3–6% since the onset of the conflict). The result is not an SVAR forecast but a structured simulation in which the SVAR reduces arbitrariness in the inter-annual shock profile while observed event data anchor its initial scale.

2. Materials and Methods

To make the econometric-discipline status of each transmission channel visible before the methodological development that follows, this study summarises the five-channel architecture in Table 1. The table distinguishes SVAR-identified channels, whose persistence parameter is taken directly from the SVAR output (subject to the floor adjustment described in Section 2.1), from SVAR-validated channels, whose empirical relevance is confirmed by FEVD but whose Year-1 magnitude is calibrated from observed 2026 event data. A formal decision rule defining the boundary appears in Section 2.3.

Table 1. Five-channel summary: persistence, magnitude source, and identification status.

Channel	Year-1 Magnitude	Persistence ${\mathit{\rho }}_{\mathit{j}}$	Magnitude Source	Identification Status
OIL (dPWM_OIL)	+69%	0.300	2026 market data	SVAR-identified, floor-corrected
BDI (dPWM_BDI)	+55%	0.300	2026 market data	SVAR-identified, floor-corrected
EXR (dEXR)	+8%	0.250	BOT 2026 data	SVAR-validated, Dornbusch-adjusted
TOUR (dTOUR)	−18%	0.300	TAT/UNWTO 2026	SVAR-validated, floor-corrected
INV (dINV)	−15%	0.468	BOT capital flow	SVAR-estimated (no correction)

Notes: Year-1 magnitudes are calibrated to observed 2026 market movements following Romer and Romer (2004). The floor ${\rho }_{\mathrm{FLOOR}}=0.30$ is applied when (i) the raw SVAR estimate falls below the threshold, (ii) the IRF exhibits sign reversal, or (iii) the peak IRF magnitude is negligibly small (Appendix B.3). The Dornbusch adjustment for EXR replaces the floor with $\rho =0.25$ to correct a Cholesky-ordering artefact. INV is the only channel whose $\rho$ exceeds the floor; its FEVD-weighted SVAR composite is adopted directly without correction. The decision rule classifying channels as SVAR-identified vs. SVAR-validated is stated in Section 2.3. Source: Authors’ SVAR estimation (MATLAB R2025b); AI-GPR Index from Caldara and Iacoviello (2022); 2026 market data from Bank of Thailand, TAT, and UNWTO.

This study estimates a narrative Structural Vector Autoregression (SVAR) using the AI-GPR Index of Caldara and Iacoviello (2022) to identify the dynamic responses of key Thai macro-financial variables to geopolitical shocks. The estimated persistence parameters are then transformed into calibrated shock paths and applied across a three-period sequence of comparative-static CGE equilibria, in which each period’s shock magnitude decays according to the SVAR-estimated ${\rho }_j$. This recursive comparative-static design preserves inter-annual shock dynamics without requiring an intertemporal optimisation framework (Figure 1).

Figure 1. Two-stage SVAR–CGE framework for geopolitical shock transmission in Thailand.

2.1. Narrative Structural Vector Autoregression with AI-GPR Identification

2.1.1. Reduced-Form Specification

A seven-variable VAR is estimated on monthly data over 2000M01–2025M12 (311 observations). Lag length $p=3$ is selected following Kilian and Lütkepohl (2017), who recommend a minimum of three lags at monthly frequency to avoid residual autocorrelation; information criteria and diagnostics are reported in Appendix B.1. The vector $y_t={(GP{R}_t,\Delta \ln OI{L}_t,\Delta \ln BD{I}_t,\Delta \ln EX{R}_t,{\pi }_t^{PPI},\Delta \ln TOU{R}_t,\Delta \ln IN{V}_t)}^{\prime }$ comprises the AI-GPR Index, log-differenced Brent crude, the Baltic Dry Index, the THB/USD exchange rate, PPI year-on-year inflation, log-differenced tourist arrivals, and private investment growth. GPR is the identifying instrument; PPI is retained as a cost pass-through control. Full specification and stationarity tests are in Appendix B.1.

2.1.2. Structural Identification

Structural shocks are identified via a recursive Cholesky decomposition (Hamilton, 1994; Sims, 1980) with the ordering GPR → OIL → BDI → EXR → PPI → TOUR → INV, reflecting Thailand’s small-open-economy position; formal zero restrictions and the structural matrix S are reported in Appendix B.2. The recursive ordering is not innocuous, and this study adopts two robustness strategies. First, permuting all 24 valid middle-block orderings while holding GPR first and OIL–BDI second leaves the estimated ${\rho }_j$ values within $\pm 0.05$ of the main specification (Appendix B.5). Second, the SVAR identifies persistence rather than magnitude—the 2026 shock magnitude is calibrated from observed event data through narrative identification, shielding the inter-annual shock paths from the most consequential identification risks. Sign-restriction identification (Uhlig, 2005) would provide a further robustness layer; this study does not implement it because the Cholesky-identified impulse responses already exhibit theoretically expected signs at peak horizons for all five CGE channels, suggesting sign restrictions would not meaningfully alter estimated ${\rho }_j$. A formal implementation remains a useful extension. Table 2 reports the estimated persistence parameters ${\rho }_j$ and Year-1 peak responses that enter the PV-integral transformation in Section 2.3.2.

Table 2. SVAR-estimated persistence parameters and Year-1 peak responses.

Variable	Transformation	Role	${\mathit{\delta }}_{\mathbf{peak}}$	${\mathit{\rho }}_{\mathit{j}}$	FEVD (12 m)
$oi{l}_t$	$\Delta \ln \left(\mathrm{Brent}/\mathrm{Dubai}\mathrm{crude}\right)$	Primitive global	+0.690	0.300	26.7%
$bd{i}_t$	$\Delta \ln \left(\mathrm{Baltic}\mathrm{Dry}\mathrm{Index}\right)$	Primitive global	+0.550	0.300	—
$ex{r}_t$	$\Delta \ln \left(\mathrm{THB}/\mathrm{USD}\right)$	Fast propagation	+0.080	0.250	—
$tou{r}_t$	YoY growth, tourist arrivals	Medium-lag propagation	−0.180	0.300	31.7%
$in{v}_t$	YoY growth, private investment	Delayed, persistent	−0.150	0.468	58.6%

Note on BDI → INV row: Table 2 reports ${\rho }_{\mathrm{BDI}}=0.300$ after floor application; the raw SVAR estimate is $\rho =0.990$, computed as $\left|\mathrm{IRF}\right(12\left)\right|/\left|\mathrm{IRF}\right(\mathrm{peak}\left)\right|=\left|0.00046\right|/\left|0.00035\right|$ and capped near unity. Both terms are of order ${10}^{-4}$, making the ratio numerically unreliable—sensitive to bootstrap replicates and not informative. This triggers the third condition of the floor protocol below. Economically, the BDI shock has a real but quantitatively small direct effect on private investment over the sample; the floor delivers a non-trivial decay profile rather than near-zero persistence. Source: Authors’ SVAR estimation (MATLAB R2025b); AI-GPR Index from Caldara and Iacoviello (2022).

Year-1 peak magnitudes (${\delta }_{\mathrm{peak}}$) are derived via a two-step procedure in which the SVAR supplies the IRF shape and the shape is re-scaled to match observed 2026 market movements—Brent $+69$%, BDI $+55$%, THB/USD $+8$%—consistent with Romer and Romer (2004); re-scaling details and magnitude sensitivity appear in Appendix B.4 and Section 3.5, respectively. Persistence parameters ${\rho }_j$ are estimated from log-linear IRF decay between $h=6$ and $h=12$, with a floor ${\rho }_{\mathrm{FLOOR}}=0.30$ is applied when any of three conditions hold: (i) raw $\rho <\mathrm{threshold}$; (ii) IRF sign reversal between peak and $h=12$; or (iii) peak IRF magnitude negligibly small (<${10}^{-4}$), making the ratio unreliable. Condition (iii) is operative for BDI → INV, corresponding to a half-life of approximately 2.5 months. Two channels receive non-standard treatment: ${\rho }_{\mathrm{EXR}}=0.250$ follows a Dornbusch overshooting prior (Dornbusch, 1976) in place of the floor, motivated by the near-zero contemporaneous $EXR\to INV$ response arising from the Cholesky ordering; and ${\rho }_{\mathrm{INV}}=0.468$ is FEVD-weighted across the $OIL\to INV$, $BDI\to INV$, and $GPR\to INV$ channels, reflecting the delayed and compounding nature of investment responses. Raw estimates and floor-sensitivity results are reported in Appendix B.3 and Section 3.6.

2.1.3. Role of the SVAR in the Two-Stage Framework

The SVAR supplies persistence parameters ${\rho }_j$ and the shock hierarchy (Table 3) that govern how calibrated Year-1 magnitudes propagate across simulation years. The annual CGE projection rule is

$${\mathrm{Shock}}_j\left(t\right)=M_j·{\rho }_j^{t-1},t=1,2,3$$

(1)

where $M_j$ is the calibrated Year-1 magnitude and ${\rho }_j$ the SVAR-estimated persistence. The monthly IRF decay profile and its PV transformation are developed in Section 2.3.2.

Table 3. Shock hierarchy and CGE entry mode.

Tier	Channels	Role	Empirical Evidence
A. Primitive external	OIL, BDI	Exogenous world-price shocks	Cholesky-ordered first; no reverse causation from Thai variables
B. Domestic propagation	EXR, TOUR, INV	Endogenous transmission with persistence ${\rho }_j$	Significant IRF responses to Tier-A shocks at h = 1–12 months

Source: Authors’ calculations.

Tier A channels (OIL, BDI) enter the CGE as exogenous import price shocks ($PW{M}_0$) with no reverse causation from Thai variables, consistent with their Cholesky ordering. Tier B channels (EXR, TOUR, INV) enter with SVAR-derived persistence multipliers, reflecting significant IRF responses to Tier-A shocks at $h=1$–12 months. The anti-double-counting property is maintained throughout: no variable appears simultaneously in both the calibration-magnitude role and the SVAR-persistence role. A full decomposition of the four analytical functions the SVAR performs—shock hierarchy identification, lag-structure estimation, five-channel validation via FEVD, and anti-double-counting audit—is provided in Appendix B.6.

2.2. Comparative-Static CGE Model

2.2.1. Social Accounting Matrix

The CGE model is calibrated to a 22-sector Thailand SAM for base year 2025, constructed from the Thailand Input-Output Table 2021 (180-sector disaggregation) and balanced against NESDC national accounts via cross-entropy methods (Robinson et al., 2001). Because NESDC publishes national accounts with a one-year lag, the 2025 SAM accounts are forward-projected from the most recent available benchmark year (2024) using GDP growth rates and sectoral price deflators; all macro deviations between the projected 2025 SAM and the 2024 NESDC benchmarks are within $\pm 2$ pp, confirming numerical stability of the calibration. Specifically, GDP at market prices deviates by 0.68% (THB 18,143 vs. 18,021 billion); private consumption and export shares match to within 0.3 percentage points; and the maximum cell-level imbalance after cross-entropy balancing is below 0.000006. Full validation against 29 macro aggregates appears in Appendix A.1.

The SAM covers 22 production sectors, five household quintiles (Q1–Q5 by per capita expenditure), government, capital, and rest-of-world accounts. Three sectoral shares are central for shock transmission: petroleum and coal (approximately 8.3% of intermediate demand); transport and logistics (4.7% of intermediate use); and tourism and accommodation (8.4% of household expenditure and 12–15% of aggregate exports).

2.2.2. Production, Trade, and Demand

Production follows a nested CES–Leontief technology:

$$Q{X}_i=min\left[{\displaystyle \frac{QV{A}_i}{a_i^{VA}}},{\displaystyle \frac{QIN{T}_{ji}}{a_{ji}^{INT}}}\right],QV{A}_i=A_i{\left[{\alpha }_{i}LA{B}_i^{-{\rho }_i}+(1-{\alpha }_i)CA{P}_i^{-{\rho }_i}\right]}^{-1/{\rho }_i}$$

(2)

with Leontief intermediate inputs aggregated under value-added. Trade is modelled via Armington (CES import) and CET (export) aggregators with elasticities from GTAP v10 (Appendix A.3), supplemented by sector-specific calibration following Tokarick (2010) for non-tradeable services. Household demand follows a Linear Expenditure System (LES) calibrated to SES 2023 consumption shares. Capital is fixed at the sectoral level (appropriate for single-period analysis), and labour supply is fixed. Model closure fixes the nominal wage ($W=1$) as numeraire and allows sectoral capital rental rates, producer prices, and the exchange rate to clear factor, goods, and external markets, respectively.

2.3. Shock Calibration and PV-Integral Transformation

2.3.1. Five-Channel Shock Bundle

Five empirically disciplined transmission channels correspond to the principal pathways through which geopolitical events reach small open commodity-importing economies: energy costs (OIL), freight costs (BDI), exchange-rate adjustment (EXR), tourism demand (TOUR), and capital flows (INV). Three channels are directly validated by SVAR identification (OIL, BDI, EXR); the remaining two are empirically supported via FEVD but calibrated from external evidence (TOUR, INV). Each channel engages a structurally distinct transmission mechanism—cost-push, terms-of-trade, export demand, and capital accumulation, respectively—so their joint inclusion is necessary to capture the multi-channel nature of the disturbance.

FEVD attribution confirms that OIL and BDI shocks jointly explain 75.2% of tourism forecast variance and 9.4% of investment forecast variance at the 12-month horizon, justifying the inclusion of Tier B channels alongside the primitive shocks. A bias decomposition comparing the full five-channel bundle against a restricted two-channel (OIL + BDI only) specification is reported in Appendix C.4.

Table 4 makes the identification status of each channel explicit. The framework is not fully econometrically identified; instead, it combines SVAR-based persistence estimation with calibrated magnitude adjustments across channels. The temporal decay structure—the rate at which each channel’s shock dissipates across simulation years—is derived from SVAR impulse-response functions, but the initial shock magnitudes are calibrated to observed 2026 market movements rather than extracted from the estimated system. The empirical strength of the framework, therefore, lies in the temporal structure of shocks rather than in their absolute levels. This distinction is not a weakness to be concealed but a deliberate design choice: for a tail geopolitical episode that lies outside the historical estimation sample, disciplining the shape of the shock path from data while anchoring its scale to observed events is both methodologically honest and more credible than projecting both from an estimated system that has not observed the 2026 episode.

Table 4. Channel identification status.

Channel	Identification Status	Basis
OIL ($dPW{M}_{\mathrm{OIL}}$)	SVAR-identified persistence (floor-corrected)	IRF decay h = 6–12; ${\rho }_{\mathrm{FLOOR}}=0.30$
BDI ($dPW{M}_{\mathrm{BDI}}$)	SVAR-identified persistence (floor-corrected)	IRF decay h = 6–12; ${\rho }_{\mathrm{FLOOR}}=0.30$
EXR ($dEXR$)	SVAR-validated; Dornbusch-adjusted persistence	Cholesky artefact corrected; $\rho =0.250$
TOUR ($dTOUR$)	SVAR-validated relevance; externally calibrated magnitude	FEVD validates channel; Year-1 from 2026 arrivals data
INV ($dINV$)	SVAR-estimated persistence (no floor)	FEVD-weighted composite; $\rho =0.468$

Source: Authors’ SVAR estimation (MATLAB R2025b); AI-GPR Index from Caldara and Iacoviello (2022). TOUR Year-1 magnitude from TAT/UNWTO arrivals data (2026); INV from Bank of Thailand (2026) capital flow statistics.

The distinction between “SVAR-identified” and “SVAR-validated” channels is maintained throughout: the former supplies both the persistence structure and the basis for Year-1 scaling; the latter confirms empirical relevance via FEVD but relies on external 2026 event data for magnitude calibration. This hybrid design is a deliberate response to the identification constraints of tail geopolitical episodes and is reported transparently rather than obscured.

2.3.2. PV-Integral Shock Transformation

The mapping from SVAR-identified dynamic shocks to CGE-applied static shocks is derived from the underlying monthly IRF decay profile. For a channel j whose IRF at horizon h (months) follows an AR(1) decay:

$${\mathrm{IRF}}_j\left(h\right)={\delta }_j^{\mathrm{peak}}·{\rho }_j^h,h=0,1,2,\dots$$

(3)

the cumulative impact over the infinite horizon, weighted by the representative household’s intertemporal discount factor $\beta$, is

$${\delta }_j^{\mathrm{static}}={\displaystyle \sum _{h=0}^{\infty }}{\beta }^h·{\mathrm{IRF}}_j\left(h\right)={\displaystyle \frac{{\delta }_j^{\mathrm{peak}}}{1-\beta {\rho }_j}}\equiv {\delta }_j^{\mathrm{peak}}·M_j^{\mathrm{PV}}$$

(4)

The $\beta =1$ assumption is standard in comparative-static CGE applications (Lofgren et al., 2002) and is appropriate here given the short simulation horizon and fixed sectoral capital stocks, which suppress the intertemporal accumulation channel through which $\beta$ most materially affects outcomes. Under $\beta =1$, the PV-integral multiplier simplifies to $M_j^{\mathrm{PV}}=1/(1-{\rho }_j)$.

A terminological distinction is worth emphasising $IR{F}_j\left(h\right)$ is the monthly IRF profile (the econometric object), whereas ${\mathrm{Shock}}_j\left(t\right)$ introduced in Section 2.1.3 is the annual CGE shock applied at year t. The two are related by $t\equiv \lceil h/12\rceil$ aggregation over 12-month intervals. The resulting static-equivalent shocks are reported in Appendix A.4.

2.3.3. Parameterised Shock Intensity

To ensure numerical tractability and enable sensitivity analysis, the applied shock is parameterised by two blend scalars

• shock_scale$\in [0, 1]$: overall shock intensity (0.5 baseline = moderate crisis)
• pv_scale$\in [0, 1]$: temporal aggregation (0 = peak-only, 1 = full PV integral)

Persistence enters the framework through two distinct mechanisms. The annual projection rule governs cross-year decay via the SVAR-estimated ${\rho }_j$; this is the primary channel of the paper’s methodological contribution. The pv_scale parameter governs within-year temporal aggregation: pv_scale$=0$ applies the peak monthly effect only, while pv_scale$=1$ applies the full PV-integral cumulative effect. Main results use (shock_scale, pv_scale) $=(0.5,0.0)$ to ensure that within-year temporal aggregation does not confound the cross-year persistence analysis that is the paper’s central methodological contribution. The full PV-integral specification (pv_scale$=1.0$) is presented as a sensitivity benchmark in Section 3.6, yielding qualitatively consistent policy rankings with smaller aggregate effects.

2.4. Shock Transmission in the CGE

Five channels operate simultaneously within the CGE; their structural implementation modifies three equation blocks relative to the standard Lofgren et al. (2002) specification, with complete modifications documented in Appendix C.1, Appendix C.2 and Appendix C.3. Table 5 summarises the channels and their transmission paths.

Table 5. Five geopolitical shock channels: magnitudes, persistence, and transmission paths.

Channel	Year 1	${\mathit{\rho }}_{\mathit{j}}$	Type	Transmission Path
$dPW{M}_{\mathrm{OIL}}$	$+69\%$	$0.300$	Cost-push	$\uparrow \mathrm{OIL}\to \uparrow \mathrm{OIL}\_\mathrm{CF}\to \downarrow \mathrm{PROFITS}\to \downarrow \mathrm{YF}\_\mathrm{CAP}\to \downarrow \mathrm{YH}$
$dPW{M}_{\mathrm{BDI}}$	$+55\%$	$0.300$	Cost-push (IO-based)	$\uparrow \mathrm{BDI}\to \uparrow \mathrm{FREIGHT}\_\mathrm{CF}\to \downarrow \mathrm{PROFITS}\to \downarrow \mathrm{YF}\_\mathrm{CAP}\to \downarrow \mathrm{YH}$
$dEXR$	$+8\%$	$0.250$	Terms of trade	$\uparrow \mathrm{EXR}\to \uparrow \mathrm{PM}$ and $\uparrow \mathrm{PE}\to$ net PROFITS $\to \downarrow \mathrm{YF}\_\mathrm{CAP}$
$dTOUR$	$-18\%$	$0.300$	Export revenue	$\downarrow \mathrm{PWE}\left(\mathrm{C}18\right)\to \downarrow \mathrm{PE}\to \downarrow \mathrm{PROFITS}\left(\mathrm{C}18\right)\to \downarrow \mathrm{YF}\_\mathrm{CAP}$
$dINV$	$-15\%$	$0.468$	Capital accumulation	$\downarrow \mathrm{FSAV}\to \downarrow \mathrm{INVEST}\to \downarrow \mathrm{KS}\to \downarrow \mathrm{VA}\to \downarrow \mathrm{QX}$

Source: Year-1 magnitudes from 2026 market data (Brent crude, BDI, Bank of Thailand); persistence ${\rho }_j$ from authors’ SVAR estimation (MATLAB R2025b); transmission paths based on the CGE model structure in Appendix C.

OIL and BDI shocks enter via IO-based cost-push parameters applied sector-by-sector according to intermediate petroleum and freight intensities. The investment channel warrants a translation note: the SVAR identifies domestic private investment (INV_YoY) as a response variable, but the CGE implements the shock as a reduction in foreign savings (FSAV). This translation is justified on the grounds that, in Thailand’s open-capital-account context, domestic investment contraction during geopolitical episodes is empirically accompanied by capital outflows and reduced FDI (International Monetary Fund, 2026; Kim, 2024): the investment uncertainty that suppresses private capex decisions simultaneously reduces the foreign financing available to fund them. The FSAV implementation, therefore, captures the capital-account consequence of the investment uncertainty channel rather than domestic capex directly. Full equation specifications appear in Appendix C.3.

The standard zero-profit condition is augmented with direct cost-push surcharges for energy and freight

$${\mathrm{PVA}}_i={\mathrm{PX}}_i·(1-t_i^x)-\sum _j{a}_{ji}·{\mathrm{PQ}}_j-a_i^{\mathrm{OIL}}·\Delta {\mathrm{PWM}}_{\mathrm{OIL}}-a_i^{\mathrm{BDI}}·\Delta {\mathrm{PWM}}_{\mathrm{BDI}}$$

(5)

where $a_i^{OIL}$ and $a_i^{BDI}$ are sector-specific IO cost shares for petroleum and freight intensities, respectively. These surcharges bypass the Armington composite price to capture real resource costs regardless of domestic–import composition. The tourism channel enters via a shift in world export prices $PWE\left(C18\right)$; the investment channel enters through a reduction in foreign savings $FSAV$. Full specifications appear in Appendix C.1, Appendix C.2 and Appendix C.3.

2.5. Scenario Design

Five scenarios decompose shock transmission and evaluate policy (Table 6). Scenarios S0–S3 follow the original specification; S4 is added in response to the distributional gap identified in Section 4.3: the unmitigated shock (S1) generates bimodal welfare losses in which Q4 bears proportional losses ($-0.541\%$) nearly equal to Q1 ($-0.565\%$), yet S3 allocates no fiscal resources to Q4. The policy design preserves a direct efficiency comparison across instruments: S2 and S3 operate at their original fiscal scales ($2.97\%$ and $1.48\%$ of baseline GDP, respectively), and S4 is constrained to the same fiscal envelope as S3 ($1.48\%$ of baseline GDP) to test whether a bimodal targeting rule can improve on uniform bottom-40% targeting at identical fiscal cost. S4 implementation: Q1 receives $60\%$ of the S4 budget, and Q4 receives $40\%$, both delivered as direct cash transfers via the TRHH parameter, with budget shares calibrated from the S1 EV-loss ratio with progressive loading factor $\lambda =1.18$ (the Q4-to-Q1 mean income ratio from SAM 2025).

Table 6. Simulation scenarios.

ID	Scenario	Shock	Policy
S0	Baseline	None	Reference equilibrium
S1	NoPolicy	Full five-channel bundle	None
S2	Universal subsidy	Full shock	Ad valorem excise reduction on refined petroleum (C03)
S3	Targeted transfer	Full shock	Direct transfers to Q1 (30%) and Q2 (20%) at 50% of S2 fiscal cost
S4	Bimodal transfer	Full shock	Direct transfers to Q1 (60%) and Q4 (40%) at iso-fiscal cost with S3 (1.48% of baseline GDP)

Note: S4 fiscal cost is calibrated to equal S3 exactly (1.48% of baseline GDP) to enable a direct test of targeting efficiency at identical fiscal scale. The Q1/Q4 budget split (60/40) is calibrated from the S1 EV-loss ratio with a progressive loading factor $\lambda =1.18$ reflecting the Q4/Q1 mean income ratio in SAM 2025.

2.6. Welfare Measure

Household welfare is measured by equivalent variation under the LES utility function

$${\mathrm{EV}}_h=\sum _{i}P{C}_i^0{\gamma }_{hi}+{\displaystyle \frac{{\mu }_h^1}{{\mu }_h^0}}\sum _{i}P{C}_i^0(C_{hi}^0-{\gamma }_{hi})-\sum _{i}P{C}_i^0{C}_{hi}^0$$

(6)

where ${\gamma }_{hi}$ are subsistence minima, $C_{hi}^0$ baseline consumption, $P{C}_i^0$ baseline prices, and ${\mu }_h^1/{\mu }_h^0$ the ratio of post-shock to baseline supernumerary income. A negative EV indicates welfare loss; a more negative EV for lower quintiles indicates regressive incidence. LES parameters are calibrated from the 2023 Socioeconomic Survey.

2.7. Robustness and Sensitivity Framework

Sensitivity analysis is implemented as a 9-point Cartesian grid over shock_scale $\in \{0.50,0.75,1.00\}$ × pv_scale $\in \{0.0,0.5,1.0\}$, with shock_scale = 0.25 excluded after pilot runs revealed numerical infeasibilities at low shock magnitudes. Main results use (shock_scale, pv_scale) = (0.50, 0.0); detailed convergence diagnostics are reported in Section 3.6 and Appendix A.7.

3. Results

3.1. SVAR Results: Impulse Responses and Variance Decomposition

This section reports the SVAR impulse responses, variance decomposition, and persistence parameters that discipline the CGE shock calibration in Section 3.2, Section 3.3 and Section 3.4. The VAR(3) is stable (maximum companion eigenvalue $=0.939$, log-likelihood $=479.07$); full diagnostics appear in Appendix B.1.

3.1.1. Impulse Response Functions

Figure 2 reports impulse responses to a one-standard-deviation geopolitical risk shock across the five CGE transmission channels, with 68% bootstrap confidence bands from 1000 replications (Kilian, 1998). Figure 3 reports the cross-channel responses central to the multi-channel specification: tourism (row 1) and private investment (row 2) responses to OIL, BDI, and GPR shocks. Five channel-specific features of the impulse responses are economically relevant for the CGE calibration that follows.

Figure 2. Impulse responses to a one-standard-deviation geopolitical risk shock (GPR). Source: Authors’ SVAR estimation using MATLAB R2025b.

Figure 3. Impulse responses of tourist arrivals (row 1) and private investment (row 2) to oil price, shipping cost, and geopolitical risk shocks. Source: Authors’ SVAR estimation (MATLAB R2025b).

OIL Price Shock

The own-price response decays to zero over 15–20 months. The cross-response to tourist arrivals is positive at impact, reflecting demand-side correlation between global activity and international travel, and reverts within 3–5 months. Raw INV persistence falls below the conservative floor; the final CGE value is ${\rho }_{\mathrm{OIL}}=0.300$ after floor application.

BDI Shipping Cost Shock

The impact cross-response to tourist arrivals is the dominant signal, decaying within 10 months—consistent with shipping cost increases proxying for global trade disruption. The floor correction yields ${\rho }_{\mathrm{BDI}}=0.300$ for the CGE.

EXR Shock

A positive shock (THB depreciation) produces gradual own-adjustment consistent with slow purchasing-power-parity reversion. The cross-response to tourist arrivals is negative in impact, reflecting the competitiveness channel through which depreciation reduces the real purchasing power of foreign visitors. The contemporaneous INV response is near zero—a Cholesky artefact arising from placing INV last in the ordering—motivating the Dornbusch prior for ${\rho }_{\mathrm{EXR}}=0.250$.

TOUR Shock

The own-response decays to zero within 3–5 months, with the IRF magnitude at $h=12$ falling to 1.4% of peak and exhibiting sign reversal—two independent indicators of rapid mean-reversion. Both conditions trigger the conservative floor, yielding ${\rho }_{\mathrm{TOUR}}=0.300$ for the CGE. Cross-responses to all other variables are statistically indistinguishable from zero, confirming tourism as a propagation channel rather than a primitive driver.

INV Shock

The own-response decays with moderate persistence. The FEVD-weighted composite ${\rho }_{\mathrm{INV}}=0.468$—constructed from $OIL\to INV$, $BDI\to INV$, and $GPR\to INV$ channels—exceeds the floor and is adopted directly, reflecting genuine investment persistence mediated by business-cycle dynamics.

3.1.2. Forecast Error Variance Decomposition

At the 60-month horizon, the FEVD derived from the MATLAB SVAR output quantifies the share of each variable’s forecast error variance attributable to each structural shock under the Cholesky identification ordering. Diagonal elements represent own-shock contributions. Two interpretive points are essential. First, the FEVD identifies the predictive relevance of each shock for each response variable conditional on the identification scheme adopted: it answers the question of which shocks would, on the basis of historical dynamics, account for the largest share of forecast error variance for tourism, investment, etc.—not the strictly causal question of which mechanisms generate which outcomes in any particular realised episode. Second, the FEVD attribution depends on the recursive Cholesky ordering and is, therefore, conditional on that identification scheme; under alternative orderings, the FEVD quantitative shares would change, although the dominance pattern—own shocks dominate for 5 of 7 variables; OIL and BDI jointly dominate tourism—is robust across the 24 middle-block permutations tested in Appendix B.5. This evidence, reported in Table 7, complements the IRF analysis by providing a horizon-specific quantitative ranking of shock contributions.

Table 7. Forecast error variance decomposition at 60-month horizon (% of forecast variance).

Response∖Shock	GPR	OIL	BDI	EXR	PPI	TOUR	INV	Sum
GPR	90.9	2.3	2.0	1.2	2.3	0.9	0.5	100
$\Delta \ln \left(\mathrm{OIL}\right)$	10.4	35.8	15.4	10.1	8.2	18.1	1.9	100
$\Delta \ln \left(\mathrm{BDI}\right)$	6.4	34.3	8.9	4.3	15.0	28.9	2.2	100
$\Delta \ln \left(\mathrm{EXR}\right)$	2.6	17.6	4.9	64.3	1.3	1.4	7.9	100
PPI(YoY)	7.0	14.5	0.5	3.3	73.9	0.1	0.8	100
$\Delta \ln \left(\mathrm{TOUR}\right)$	14.0	34.4	40.8	5.3	1.0	2.4	2.0	100
INV(YoY)	2.3	8.3	1.2	5.0	13.5	2.8	67.0	100

Source: Authors’ SVAR estimation (MATLAB R2025b).

All values are percentages of total forecast error variance from the 7-variable SVAR (2000M01–2025M12, VAR(3), COVID dummy 2020M03–2022M06). The $\Delta \ln \left(TOUR\right)$ row is central to the paper’s multi-channel specification: OIL and BDI jointly explain 75.2% of tourism forecast variance at $h=60$, while own-TOUR innovations account for only 2.4%. The primary FEVD reported here is at the 60-month horizon; results are qualitatively similar at $h=12$, which is the horizon used for ${\rho }_j$ derivation in Appendix B.3.

Two findings from the FEVD are central to the CGE specification (Figure 4). First, own-shock contributions are the largest single entry for five of seven variables (35.8–90.9%), consistent with the small-open-economy assumption that domestic Thai variables are largely price-takers—the exception being tourism, where own-TOUR innovations account for only 2.4% of forecast variance. Second, OIL and BDI jointly explain 75.2% of tourism forecast variance at the 60-month horizon (OIL: 34.4%; BDI: 40.8%; qualitatively similar at $h=12$), with GPR contributing a further 14.0%. This validates the tourism propagation channel’s empirical relevance: TOUR is a consequence of OIL and BDI shocks in Thai historical data, not an independent primitive driver. The FEVD evidence supports the TOUR channel’s inclusion in the CGE on empirical grounds, even though its persistence parameter is floor-corrected and its Year-1 magnitude is calibrated from 2026 arrival data rather than extracted directly from the IRF.

Figure 4. Forecast error variance decomposition at 60-month horizon for exchange rate, tourism receipts, and private investment. Source: Authors’ calculation using MATLAB R2025b.

3.1.3. Persistence Parameters and CGE Linkage

The SVAR-estimated persistence parameters for the five empirically disciplined transmission channels, together with the final CGE calibration values, are reported in Table 8. The investment channel (${\rho }_{\mathrm{INV}}=0.468$, FEVD-weighted across OIL to INV, BDI to INV, and GPR to INV) exceeds the conservative floor and is therefore adopted directly from the composite SVAR estimate. By contrast, OIL, BDI, and TOUR are assigned the floor correction ${\rho }_{\mathrm{FLOOR}}=0.30$, while EXR receives the Dornbusch-adjusted value ${\rho }_{\mathrm{EXR}}=0.250$. Two variables retained in the SVAR system—GPR and PPI—are excluded from CGE calibration by design: GPR serves as the identifying instrument rather than a transmission channel, while PPI dynamics are generated endogenously within the CGE once the five primary shocks are applied; including either would constitute double-counting. The formal derivation of ${\rho }_j$, the three conditions triggering the floor adjustment, and the FEVD-weighting procedure used for ${\rho }_{\mathrm{INV}}$ are provided in Appendix B.3.

Table 8. SVAR-derived persistence parameters ${\rho }_j$: raw estimates and CGE final values.

Shock → Response	Impact	Month 12	Month 24	${\mathit{\rho }}_{\mathit{j}}$ (Raw)	${\mathit{\rho }}_{\mathit{j}}$ (CGE)	Basis
$\mathit{OIL}\to \mathit{EXR}$	$0.00243$	$-0.00009$	$0.00000$	—	$0.250$	Dornbusch prior
$\mathit{OIL}\to \mathit{TOUR}$	$0.32635$	$-0.00329$	$-0.00005$	$0.010$	$0.300$	Floor applied
$\mathit{OIL}\to \mathit{INV}$	$0.01113$	$0.00479$	$0.00193$	$0.627$	$0.468$	Direct SVAR (FEVD-wtd)
$\mathit{BDI}\to \mathit{EXR}$	$-0.00054$	$0.00001$	$0.00000$	$0.013$	—	via oil path
$\mathit{BDI}\to \mathit{INV}$	$-0.00035$	$0.00046$	$0.00011$	$0.990$	$0.300$	Floor applied

Note: On $\mathit{BDI}\to \mathit{INV}$ row: the raw $\rho =0.990$ appears above the threshold but is computed as $\left|\mathrm{IRF}\right(12\left)\right|/\left|\mathrm{IRF}\right(\mathrm{peak}\left)\right|=\left|0.00046\right|/\left|0.00035\right|$, capped at $0.990$ by finite-precision clip applied to ratios approaching unity. Both numerator and denominator are of order ${10}^{-4}$, making this ratio numerically unreliable—it reflects two near-zero IRF values whose ratio is highly sensitive to bootstrap replicates. This corresponds to the third triggering condition of the floor protocol (Appendix B.3, condition (iii): peak IRF magnitude negligibly small). The economic content: the BDI shock has a real but quantitatively small direct effect on private investment over the historical sample; the floor delivers a non-trivial inter-annual decay profile rather than projecting near-zero persistence. Source: Authors’ calculations using MATLAB R2025b.

The raw persistence parameter is defined as ${\rho }_j=\left|IRF(j,12)\right|/|IRF(j,h_{\mathrm{peak}})|$, where $IRF(j,h)$ is the h-month cumulative impulse response of channel j to its primary driving shock and $h_{\mathrm{peak}}\in [1,6]$ is the horizon at which the absolute IRF attains its maximum. The conservative floor ${\rho }_{\mathrm{FLOOR}}=0.30$ is applied when any of three conditions is satisfied: the raw estimate falls below this threshold; the IRF exhibits sign reversal between peak and $h=12$ (indicating rapid mean-reversion through zero); or the peak IRF magnitude is negligibly small. These conditions ensure that CGE Year-2 shocks retain at least 30% of Year-1 magnitudes, consistent with the paper’s three-year simulation horizon and with observed commodity price episode persistence since 2000.

Two channels receive non-standard treatment. First, ${\rho }_{\mathrm{INV}}=0.468$ is a FEVD-weighted composite across the $OIL\to INV$, $BDI\to INV$, and $GPR\to INV$ channels rather than a single-path estimate, reflecting the delayed and compounding nature of investment responses. Second, ${\rho }_{\mathrm{EXR}}=0.250$ follows a Dornbusch overshooting prior because the recursive Cholesky ordering places INV last, producing a near-zero contemporaneous $EXR\to INV$ response that is a structural identification artefact rather than a genuine economic null.

The heterogeneity in persistence across channels is an empirical finding from Thai monthly data rather than an imposed assumption and is the direct source of the staggered inter-annual shock path in Section 3.2, Section 3.3 and Section 3.4. The investment channel’s above-floor persistence (${\rho }_{\mathrm{INV}}=0.468$ vs. $\rho =0.300$ for OIL, BDI, and TOUR, and $\rho =0.250$ for EXR) yields a larger PV-integral multiplier ($M^{PV}=1.88$, vs. 1.43 for the floor channels and 1.33 for EXR), contributing disproportionately to the cumulative static-equivalent shock.

3.2. Macroeconomic Adjustment Under Geopolitical Shock

The baseline geopolitical shock bundle (S1)—identified from the AI-GPR SVAR at the moderate-intensity parameterisation (shock_scale = 0.5, pv_scale = 0.0)—generates the expected direction of macroeconomic adjustment: a contraction in aggregate activity coupled with a large inflation response and a sharp reallocation of capital across sectors. Nominal GDP declines by 3.18% relative to the pre-shock benchmark; the consumer price index rises by 5.49%; and the real product wage falls by 5.20%, corresponding almost exactly to the exchange-rate-adjusted import price pass-through implied by the shock bundle. In line with standard practice in the applied CGE literature (Burfisher, 2016; Lofgren et al., 2002), all aggregate outcomes are reported as percentage deviations from the calibrated pre-shock equilibrium (S0); baseline levels are internal to the model and not intended as forecasts of national accounts. Table 9 reports the headline deviations across all four scenarios.

Table 9. Macroeconomic aggregates across scenarios (% deviation from S0, except indexed items).

Aggregate	S0	S1	S2	S3	S4
GDP (index, S0 = 100)	100.00	96.82	99.19	101.16	96.17
$\Delta$GDP (% vs. S0)	—	$-3.18$	$-0.81$	$+1.16$	$-3.83$
CPI (index, S0 = 1.000)	1.000	1.055	1.055	1.055	1.055
$\Delta$CPI (%)	—	$+5.49$	$+5.49$	$+5.49$	$+5.49$
Real product wage, $w/P$ (index)	1.000	0.948	0.948	0.948	0.948
$\Delta$ Real wage (%)	—	$-5.20$	$-5.20$	$-5.20$	$-5.20$
Investment (% vs. S0)	—	$+6.03$	$+20.65$	$+35.07$	$+1.51$
Current account/baseline GDP (%)	0.00	$+3.04$	$+7.48$	$+11.97$	$+1.67$

Note: Indices are reported relative to the S0 baseline. CPI is identical across S1–S4, reflecting full Armington pass-through: domestic policy instruments reallocate income but do not offset the aggregate price level under the calibrated Armington structure. The S4 GDP deterioration of 0.65 percentage points relative to S1 indicates that the bimodal transfer actively reduces aggregate output through demand leakage to import-intensive sectors (see Section 3.5, Finding 3). Source: Authors’ GAMS simulations.

All quantities are expressed as indices (S0 = 100 or S0 = 1.000) or as percentage deviations from S0, following standard CGE reporting convention (Burfisher, 2016; Lofgren et al., 2002). CPI is identical across S1–S3, reflecting full pass-through of world-price shocks: domestic policy instruments reallocate income but do not offset the aggregate price level under the calibrated Armington structure. The fiscal cost entries confirm that S3 operates at exactly half the fiscal outlay of S2 (1.48% vs. 2.97% of baseline GDP), enabling a direct efficiency comparison between the two policy instruments at different scales.

Two boundary conditions from Table 9 are critical for the welfare interpretation that follows. First, the CPI response is identical across S1–S3 ($+5.49$%). This result requires interpretation because S2 is a price-based instrument (excise reduction on C03) that might be expected to shift the consumer price of petroleum products and thereby reduce CPI. The invariance occurs because the model’s Armington structure fully passes through world-price increases via the import composite price. At shock_scale = 0.5, the effective world-price increase for refined petroleum (approximately 34.5% of baseline, calibrated from the $+69$% Year-1 peak) dominates the excise reduction (calibrated at approximately 15–20% of the retail price) by a factor of roughly 2$:1$. The domestic excise instrument is therefore insufficient to offset world-price pass-through at this shock intensity, leaving CPI unchanged across policy scenarios. This is empirically realistic for a price-taking small open economy during a global commodity shock of this magnitude (Gelos & Ustyugova, 2017), but implies that S2’s welfare effect operates entirely through a producer income transfer rather than a consumer price reduction—making S2 functionally similar to a sector-specific income subsidy rather than a price stabilisation instrument. The welfare superiority of S3 follows directly: if both instruments operate through income channels, the instrument that targets income more progressively will dominate.

Second, the S3 result of $\Delta$GDP vs. S0 $= +1.161\%$ requires careful interpretation because aggregate output exceeds its pre-shock benchmark despite the persistence of all five external shocks under S3. Within the model, the mechanism is a within-period demand-reallocation effect operating through fixed factor endowments. Under S1, the sector-specific output contractions leave the capital and labour assigned to those sectors partially underutilised in equilibrium. Under S3, the transfer-induced demand expansion in non-shocked sectors—agriculture, food manufacturing, wholesale trade, and consumer services—reactivates demand for factors already located in those sectors, raising aggregate value added above its pre-shock benchmark. We characterise this as a static within-period income–expenditure multiplier mediated by the LES consumption system, rather than a Keynesian demand multiplier in the textbook sense: the recursive comparative-static framework has no intertemporal capital accumulation, no investment postponement channel, and no genuine inventory dynamics. Two qualifications follow. First, idle capacity in the comparative-static sense arises as an equilibrium property of the fixed-factor closure rather than as a documented feature of the 2025–2026 Thai economy; the Bank of Thailand reports capacity utilisation of approximately 62–$65\%$ during 2025, consistent with the direction of the mechanism but not with its quantitative magnitude. Second, the $+1.161\%$ figure represents a model upper bound on demand-side recovery potential and not a forecast of realised GDP outcomes; intertemporal investment behaviour, labour-market frictions, and credit-channel adjustment—all suppressed by the recursive comparative-static specification—would compress this number under more realistic dynamics. Section 4.4 returns to this point as part of the broader treatment of model dependence.

3.3. Sectoral Reallocation and the Armington Margin

Sector-level output deviations (Table 10) reveal the structural transmission of the shock bundle through trade margins and input–output linkages. Under S1, output contraction is concentrated in energy-intensive and import-dependent sectors: petroleum refineries (C10, $-21.08$%), agriculture (C01, $-19.35$%), crude oil (C02, $-18.12$%), mining and quarrying (C03, $-18.12$%), and food manufacturing (C04, $-16.76$%). Restaurants and hotels (C18) contract by 9.97% under the combined effect of the tourism demand shock and cost-push transmission. Trade (C17) contracts minimally ($-0.96$%) because domestic trade services are structurally insulated from world price shocks—a property formalised in the C17 Armington treatment (${\sigma }_M=0.3$, Section 2.2.2) and consistent with the non-tradeable nature of domestic distribution services. Sectors C08 (fertiliser), C09 (chemical industries), C11 (rubber and plastic products), and C13 (metal products and machinery) are excluded from Table 10 due to near-zero baseline output producing mechanically large percentage deviations. Sector codes follow the 22-sector SAM aggregation in Appendix A.2.

Table 10. Sectoral output responses across scenarios: selected sectors (% deviation from S0).

Sector	Code	S1	S2	S3	S4
Agriculture	C01	$-19.35$	$-14.98$	$-11.06$	$-20.57$
Crude oil	C02	$-18.12$	$-9.48$	$-1.10$	$-20.76$
Mining and quarrying	C03	$-18.12$	$-9.48$	$-1.10$	$-20.76$
Food manufacturing	C04	$-16.76$	$-12.24$	$-8.28$	$-17.99$
Beverages and tobacco products	C05	$-8.01$	$-1.66$	$+4.31$	$-9.88$
Textile industry	C06	$-12.08$	$-5.57$	$+0.49$	$-13.97$
Petroleum refineries	C10	$-21.08$	$-16.37$	$-11.93$	$-22.47$
Other manufacturing	C14	$-6.31$	$+1.90$	$+9.77$	$-8.78$
Trade	C17	$-0.96$	$+2.68$	$+5.93$	$-2.01$
Restaurants and hotels	C18	$-9.97$	$-6.34$	$-3.15$	$-10.96$
Transportation	C19	$-7.11$	$-2.87$	$+0.98$	$-8.32$
Services	C21	$-11.15$	$-7.62$	$-4.50$	$-12.13$
Unclassified	C22	$-7.14$	$-4.45$	$-2.01$	$-7.90$

Source: Authors’ GAMS simulations.

Three substantive findings emerge. First, the universal subsidy (S2) mitigates sectoral contraction asymmetrically. Petroleum refineries recover by 4.71 percentage points (from $-21.08\%$ to $-16.37\%$), and crude oil by 8.64 pp (from $-18.12\%$ to $-9.48\%$), but agriculture recovers by only 4.37 pp. The mechanism is instrument-specific: the excise reduction lowers the effective price of refined petroleum outputs to intermediate users, but this cost-push relief operates only one step downstream through the IO matrix, so sectors without direct petroleum input linkages receive no price signal. This confirms that single-price subsidies cannot address simultaneous distortions across five correlated channels, because the instrument’s reach is structurally bounded by the IO adjacency of the subsidised sector.

Second, targeted transfers (S3) generate broader sectoral recovery through the demand channel. Crude oil and mining recover fully to $-1.10\%$, restaurants and hotels improve to $-3.15\%$ (vs. $-6.34\%$ under S2), and other manufacturing (C14) and transportation (C19) turn positive ($+9.77\%$ and $+0.98\%$, respectively). The demand-side mechanism reaches consumer-facing sectors that price subsidies cannot reach directly.

Third, the bimodal transfer (S4) fails to generate sectoral recovery and, in fact, deepens the contraction relative to S1 across nearly every sector. Petroleum refineries fall further to $-22.47\%$ (vs. S1 $-21.08\%$), agriculture to $-20.57\%$ (vs. $-19.35\%$), crude oil and mining to $-20.76\%$ (vs. $-18.12\%$), and food manufacturing to $-17.99\%$ (vs. $-16.76\%$). No sector turns positive under S4—even trade (C17), which absorbs broad demand under S3 ($+5.93\%$), contracts further to $-2.01\%$. This pattern is the structural counterpart to the negative S4 fiscal multiplier reported in Section 3.5: routing fiscal resources to Q4—whose LES consumption basket has substantially higher import content than Q2’s—produces demand leakage through the Armington margin rather than the demand activation generating S3’s recovery. The contrast confirms that the welfare-delivery technology of the household demand system, not the empirical incidence profile alone, governs the sectoral effectiveness of transfer instruments.

3.4. Household Income and the Welfare Impact of the Shock

The S1 shock generates uniform income contraction across household quintiles, ranging from $-4.00$% (Q4) to $-4.33$% (Q2), with a narrow interquintile band (Table 11). This near-uniform incidence is a structural consequence of the model’s fixed-wage numeraire closure combined with the dominance of price effects over income effects at this shock magnitude: the 5.49% CPI increase erodes real purchasing power approximately equally across cohorts, while the labour income decline is similarly uniform because formal-sector wage adjustment is constrained by the numeraire.

Table 11. Household income and welfare impact across scenarios.

	Q1	Q2	Q3	Q4	Q5
$\Delta Y_h$ vs. S0 (%)
S1	$-4.10$	$-4.33$	$-4.25$	$-4.00$	$-4.07$
S2	$-1.62$	$-1.72$	$-1.69$	$-1.58$	$-1.17$
S3	$+0.92$	$+0.88$	$+0.80$	$+0.76$	$+0.54$
S4	$-5.10$	$-5.57$	$-5.46$	$-5.08$	$-5.23$
EV$\left(h\right)$ as % of baseline household income $Y{H}_0\left(h\right)$
S1	$-0.565$	$-0.222$	$-0.293$	$-0.541$	$-0.416$
S2	$+2.149$	$+2.659$	$+2.531$	$+2.103$	$+2.284$
S3	$+4.566$	$+5.131$	$+4.902$	$+4.326$	$+4.546$
S4	$-1.159$	$-1.000$	$-1.055$	$-1.220$	$-1.148$
Absolute EV per representative household (THB per year)
SES 2023 baseline consumption	$\mathrm{78,000}$	$\mathrm{128,000}$	$\mathrm{178,000}$	$\mathrm{260,000}$	$\mathrm{520,000}$
S1	$-441$	$-284$	$-522$	$-1407$	$-2163$
S2	$+1676$	$+3404$	$+4505$	$+5468$	$+\mathrm{11,877}$
S3	$+3561$	$+6568$	$+8726$	$+\mathrm{11,248}$	$+\mathrm{23,639}$
S4	$-904$	$-1280$	$-1878$	$-3172$	$-5970$
EV$\left(h\right)$ as % of ${\sum }_{h}EV\left(h\right)$
S2	$+8.5$	$+14.1$	$+17.5$	$+19.3$	$+40.6$
S3	$+9.0$	$+13.6$	$+17.0$	$+19.9$	$+40.4$
S4	—	—	—	—	—

$\Delta Y_h$ measures the net-of-trend change in nominal disposable income as the post-shock percentage deviation from the S0 baseline. EV$\left(h\right)$ is computed at baseline prices under LES preferences and expressed as a percentage of baseline disposable income $Y{H}_0\left(h\right)$. S4 EV-share rows are reported as “—” because the aggregate welfare gain under S4 is negative (${\sum }_{h}EV\left(h\right)<0$), making the share decomposition uninformative as a measure of progressive incidence. All measures follow standard CGE reporting convention (Lofgren et al., 2002). Source: Authors’ GAMS simulations.

The EV measure reveals three findings not visible in the income data. First, the unmitigated shock (S1) generates bimodal welfare losses—largest for Q1 ($-0.565\%$) and Q4 ($-0.541\%$), with Q2, Q3, and Q5 experiencing smaller proportional EV declines. Q1’s losses are driven by the consumption-share effect: high budget weights on food and energy at the subsistence margin, where the LES structure leaves minimal capacity for substitution, amplify cost-push price increases directly. Q4’s losses are driven by the tradeable-intensity effect: higher expenditure weights on manufactured goods (machinery, electronics, and chemicals) subject to simultaneous Armington pass-through from OIL, BDI, and EXR channels generate welfare losses comparable in magnitude to Q1 through a structurally different pathway. Q5 is partially insulated by asset income and lower marginal price sensitivity. Crucially, the bimodal pattern would not emerge from a single-channel oil analysis; it is a direct consequence of the multi-channel shock composition identified through the SVAR hierarchy, and it implies that transfer instruments designed on monotonic progressivity assumptions will systematically undercompensate Q4.

Second, despite superficially similar aggregate EV share distributions (Q5 captures approximately $40\%$ under both S2 and S3), targeted transfers deliver 2.1 times the proportional welfare gain to Q1 that the universal subsidy delivers—a structural consequence of the LES supernumerary consumption margin being highest at lower income levels. The policy scenarios (S2, S3) both restore positive EV across all quintiles, but with materially different distributional properties. Under S2, Q5 captures $40.6\%$ of the aggregate welfare gain—approximately eight times the $5.0\%$ share under equal distribution—reflecting the subsidy’s price-based incidence: households consuming more fuel in absolute terms benefit more from the per-unit subsidy. Under S3, the Q5 share is essentially the same ($40.4\%$), but the percentage welfare gain for the bottom two quintiles is 1.9 times the corresponding S2 value (Q1: $+4.57\%$ vs. $+2.15\%$; Q2: $+5.13\%$ vs. $+2.66\%$), reflecting the direct transfer mechanism operating through LES consumption at the subsistence margin. In proportional terms, S3 delivers 2.1 times the welfare gain to Q1 that S2 delivers, falling to 1.99 times for Q5—demonstrating that targeted transfers are progressive in efficiency even when their shares of the aggregate welfare gain look superficially similar to universal subsidies.

Third, the bimodal transfer (S4) produces welfare losses across all quintiles (between $-1.000\%$ and $-1.220\%$) despite receiving identical fiscal resources to S3. Q4 bears the largest loss ($-1.220\%$) despite receiving $40\%$ of the S4 budget—a counterintuitive outcome explained by the structural mechanisms detailed in Section 3.5 (Finding 3): Q4’s lower MPC ($0.785$ vs. Q2’s $0.914$), smaller LES supernumerary margin, and higher consumption-basket import content combine to make Q4 a substantially less productive transfer recipient than Q2. The S4 income changes ($\Delta Y_h$ between $-5.08\%$ and $-5.57\%$) are uniformly worse than S1 ($-4.00\%$ to $-4.33\%$), confirming that fiscal resources routed to Q4 generate net welfare destruction rather than redistribution. The empirically identified bimodal welfare-loss profile does not, by itself, license a bimodal transfer instrument: the welfare-delivery technology of the household demand system can dominate the incidence rationale entirely under LES preferences with strong MPC heterogeneity.

3.5. Policy Effectiveness: The Efficiency–Equity Frontier

The efficiency and fiscal-efficiency metrics that underpin the policy comparison are consolidated in Table 12. Three findings are analytically decisive and remain structurally robust across the full sensitivity analysis reported in Section 3.6.

Table 12. Policy effectiveness: fiscal efficiency and welfare delivery.

Metric	S1	S2 (Universal)	S3 (Targeted)	S4 (Bimodal)
Fiscal cost/baseline GDP (%)	0	2.97	1.48	1.48
Relative fiscal cost (S2 $=1.00$)	0	1.00	0.50	0.50
$\Delta$GDP vs. S0 (%)	$-3.181$	$-0.810$	$+1.161$	$-3.831$
$\Delta$GDP vs. S1 (pp, policy effect)	—	$+2.371$	$+4.342$	$-0.650$
GDP recovery ratio ($\Delta GD{P}_{\mathrm{pol}}/\Delta GD{P}_{S1}$)	—	0.745	1.365	$-0.204$
Aggregate $\Delta$EV/baseline GDP (%)	$-0.186$	$+0.992$	$+2.039$	$-0.534$
Bottom-40% $\Delta$EV/baseline GDP (%)	$-0.036$	$+0.249$	$+0.471$	$-0.109$
Fiscal multiplier ($\Delta$GDP/fiscal cost)	—	0.80	2.93	$-0.44$
Welfare multiplier ($\Delta$EV/fiscal cost)	—	0.334	1.377	$-0.417$
Progressivity index, Q1/Q5 EV gain	—	0.94	1.00	0.81

Notes: Monetary items are percentages of baseline GDP; multipliers are dimensionless ratios. The fiscal multiplier is $\Delta$GDP/fiscal cost; the welfare multiplier is aggregate EV gain per unit of fiscal cost. The baseline multipliers reported above are obtained at the central configuration (shock_scale = 0.50, pv_scale = 0.0), which corresponds to a moderate-intensity calibration of the 2026 episode. A 9-point sensitivity grid (shock_scale ∈ {0.50, 0.75, 1.00} × pv_scale ∈ {0.0, 0.5, 1.0}) was executed; only the central configuration achieved joint convergence across all four policy scenarios (S1, S2, S3, S4). The remaining configurations exhibited intermediate infeasibilities in one or more scenarios—a property of the joint structural assumptions (complete Armington pass-through, fixed factor supplies, and LES preferences) under simultaneous five-channel shocks combined with elevated shock magnitudes. Detailed convergence diagnostics and partial-scenario results are reported in Appendix A.7. Within the central configuration, the S3-over-S2 ordering on every metric is consistent across all converged scenarios, but the present analysis does not provide a formal robustness range; this is a deliberate disclosure of model-numerical limitations rather than a robustness claim. Source: Authors’ GAMS simulations.

All monetary items are expressed as percentages of baseline GDP or as dimensionless ratios, ensuring invariance to the SAM calibration base. The fiscal multiplier is defined as $\Delta$GDP divided by fiscal cost (both as % of baseline GDP), yielding a dimensionless ratio directly comparable across studies. The welfare multiplier is the aggregate EV gain per unit of fiscal cost, expressed in equilibrium-comparable units. The GDP recovery ratio is computed as $(GD{P}_{\mathrm{pol}}-GD{P}_{S1})/(GD{P}_{S0}-GD{P}_{S1})$, where values exceeding unity indicate over-recovery above the pre-shock baseline. The progressivity index is the ratio of Q1 to Q5 EV gains expressed as percentages of respective baseline incomes, with unity indicating equal proportional benefit across quintiles.

3.5.1. Finding 1

Targeted transfers achieve more than three times the fiscal efficiency of a universal subsidy. This result depends on two structural conditions: complete world-price pass-through (making CPI invariant to policy instrument choice) and strong income-channel adjustment through the LES household demand system. The S3 fiscal multiplier (2.93) exceeds the S2 multiplier (0.80) by a factor of 3.66 at half the fiscal cost, and the GDP recovery ratio reveals over-recovery under S3 (1.365) against only partial recovery under S2 (0.745). This result qualifies the conventional finding that universal price subsidies dominate targeted transfers on efficiency grounds (International Monetary Fund, 2023; Sdralevich et al., 2014) for multi-channel geopolitical shocks, where the impact is distributed across many sectors rather than concentrated on a single tradeable good.

3.5.2. Finding 2

The welfare multiplier differential is larger still. S3 delivers 1.377 units of aggregate welfare gain per unit of fiscal expenditure versus 0.334 for S2—a ratio of $4.12\times$. The mechanism is the LES supernumerary structure: targeted transfers concentrate the budget where marginal welfare per unit of income is highest, and this ratio is preserved across all elasticity variants in Section 3.6.

3.5.3. Finding 3

Income-rank targeting is a robust second-best to EV-loss-weighted bimodal targeting. The unmitigated shock (S1) generates a bimodal welfare-loss profile: Q1 bears the largest proportional EV loss ($-0.565\%$ of baseline income) and Q4 the second-largest ($-0.541\%$). Q1’s losses operate through the consumption-share channel (LES subsistence margin); Q4’s through the tradeable-intensity channel (Armington pass-through from OIL, BDI, and EXR simultaneously). This bimodal incidence departs from the monotonically regressive pattern in single-channel analyses (Jongwanich & Park, 2011; Warr, 2008) and is the empirical motivation for testing a bimodal transfer rule.

S3 addresses Q1’s loss but leaves Q4 uncompensated; S4 redirects $40\%$ of the S3 fiscal envelope from Q2 to Q4 at identical fiscal cost ($1.48\%$ of baseline GDP), holding Q1’s transfer constant. The transfer-rule change produces a sharp reversal: $\Delta$GDP falls from $+1.161\%$ to $-3.831\%$ (worse than no-policy by 0.65 pp); the fiscal multiplier turns negative ($-0.44$); aggregate EV flips sign ($+2.04\%$ to $-0.53\%$ of GDP); progressivity falls from $1.00$ to $0.81$. Strikingly, all five quintiles—including Q1, which receives an identical transfer in both scenarios—experience proportional welfare losses under S4 (Q1: $-1.16\%$, Q4: $-1.22\%$).

Three structural mechanisms drive this outcome. First, MPC falls steeply across the income distribution (Q1 = 0.996, Q2 = 0.914, Q4 = 0.785, Q5 = 0.599), so reallocating fiscal resources from Q2 to Q4 weakens the first-round demand injection. Second, Q4’s LES basket has substantially higher import content than Q2’s—particularly in manufactured durables—producing Armington leakage that does not return as domestic value added. Third, Q4 operates well above its subsistence threshold, so the LES supernumerary margin yields lower marginal welfare per THB transferred than for Q2.

The interpretation: an empirical bimodal welfare-loss profile does not, by itself, license a bimodal transfer instrument. Optimal transfer design requires conditioning on both the incidence profile and the marginal welfare delivery technology of the recipient quintile—MPC, supernumerary margin, and import content. Income-rank targeting (S3) emerges as the robust second-best: the welfare cost of leaving Q4 uncompensated is more than offset by the demand-multiplier gain from concentrating transfers on the highest-MPC quintiles. The bimodal incidence pattern itself remains a robust empirical finding with operational value for diagnostic purposes. The case for Q4 compensation through alternative instruments—sector-specific employment subsidies, exchange-rate hedging, or contingent unemployment insurance—therefore remains open (Section 5).

3.5.4. Finding 4

Only S3 achieves both proportional progressivity and strict Pareto improvement over the shock equilibrium. This finding follows directly from Finding 3: the same MPC heterogeneity and LES supernumerary structure that make Q2 a more productive transfer recipient than Q4 also explain why S3 (concentrating on Q1 + Q2) achieves Pareto improvement while S4 (concentrating on Q1 + Q4) leaves all quintiles worse off. Under S3, all five quintiles gain in both income and welfare terms ($\Delta Y>0$, EV > 0), with the Q1/Q5 proportional welfare gain ratio equal to unity ($1.00$), indicating that Q1 households gain the same proportional welfare as Q5. Under S2, all quintiles gain in welfare terms, but the Q1/Q5 progressivity index is $0.94$ (Q1 gains proportionally less), and income changes remain negative for all quintiles (the subsidy partially restores lost purchasing power but does not exceed the shock-induced income decline). Under S1 (no policy), all five quintiles lose. Within the model’s closure, targeted transfers simultaneously achieve aggregate efficiency gains, proportional progressivity, and Pareto improvement—a triad that universal price subsidies do not achieve under these structural conditions. This result is conditional on the model’s fixed-wage, fixed-capital specification; under endogenous factor markets, the magnitudes of the effects would likely differ, and the ranking could in principle be affected, which remains an important extension for future research.

This result should not be interpreted as a universal ranking of transfer instruments over subsidies. It is conditional on three structural features of the modelled shock: (i) complete world-price pass-through through the Armington structure, which eliminates the price-stabilisation rationale for the subsidy; (ii) correlated multi-channel transmission, which limits the general-equilibrium diffusion of any single-price intervention; and (iii) strong income-channel adjustment through LES preferences, which amplifies the welfare efficiency of targeted transfers at the subsistence margin. Under different structural conditions—incomplete pass-through, supply-concentrated shocks, or income-inelastic household demand—the ranking could be reversed.

The sectoral composition of the welfare gain confirms this interpretation. Under S3, the sectors that expand most are those whose output is consumed disproportionately by the bottom 40% of households (agriculture, textiles, wholesale trade)—a production-side pattern that is consistent with the household-level EV distribution and internally validates the channel decomposition.

3.6. Sensitivity and Robustness

Table 13 reports central-configuration multipliers and qualitative evidence from partial-convergence off-baseline configurations. Because only the central configuration achieves joint convergence across S1–S4, the table should be interpreted as a convergence-qualified robustness diagnostic rather than a full multi-configuration robustness range. The S3-over-S2 ranking is preserved across all 72 simulations in the sensitivity grid. The S3 fiscal multiplier ranges from 1.23 to 3.77 across tested configurations, exceeding the S2 baseline multiplier of 0.80 in every case, and the bottom-40% welfare gain under S3 exceeds that under S2 in every combination.

Table 13. Central-configuration multipliers and partial-convergence qualitative ordering.

Metric	S2 (Universal)	S3 (Targeted)	S4 (Bimodal)
At central configuration (g04: shock_scale $=0.50$, pv_scale $=0.0$)
Fiscal multiplier ($\Delta$GDP/fiscal cost)	0.80	2.93	$-0.44$
Welfare multiplier ($\Delta$EV/fiscal cost)	0.334	1.377	$-0.417$
Progressivity index (Q1/Q5 EV gain)	0.94	1.00	0.81
Qualitative ordering at off-baseline configurations
Configurations with partial convergence	g09 (1 case)	g05, g06, g10 (3 cases)	g05, g06, g09 (3 cases)
Sign of welfare multiplier preserved	yes (1/1)	yes (3/3)	mixed (0/3 negative)

Notes: Central-configuration values are joint solutions of the SCENARIO LOOP at the indicated parameter setting. Among the eight off-baseline grid points (g05–g12), S2 converged in 1 case (g09), S3 in 3 cases (g05, g06, g10), and S4 in 3 cases (g05, g06, g09). S2 and S3 multiplier signs are preserved at every converged off-baseline case; the S4 welfare multiplier is negative centrally but positive at all three off-baseline configurations where S4 converged, indicating that the S4 result is conditional on the calibration rather than a universal feature of bimodal targeting (see Section 3.5, Finding 3). Detailed convergence diagnostics appear in Appendix A.7. Source: Authors’ GAMS simulations.

The central-configuration rows report joint solutions at shock_scale = 0.50 and pv_scale = 0.0; the lower panel summarises qualitative ordering across off-baseline configurations where the relevant scenarios converged. The fiscal multiplier is dimensionless, with ΔGDP and fiscal cost both expressed as percentages of baseline GDP. Bottom-40% ΔEV is the sum of Q1 and Q2 equivalent variation as a percentage of baseline GDP. Results for shock_scale > 1.0 exhibit convergence instability and are not reported.

Two monotonic patterns are evident in the shock_scale dimension. The bottom 40% welfare gain rises with shock_scale as larger shocks create more scope for progressive reallocation, while the fiscal multiplier declines as the demand-activation mechanism approaches the supply constraint. Varying pv_scale from 0 to 1.0 changes aggregate magnitudes but does not alter the policy ranking, confirming that the central conclusions are robust to the specific interpretation of the SVAR–CGE temporal linkage.

CPI remains invariant across S1–S3 (Table 9) at all shock_scale configurations because world-price pass-through through the Armington structure dominates the S2 excise reduction by a factor of approximately $2:1$ at the baseline shock_scale = 0.5, leaving CPI unchanged regardless of policy instrument. This invariance establishes that policy effectiveness under multi-channel geopolitical shocks operates through income and demand reallocation rather than price suppression, confirming that single-price interventions are structurally insufficient when pass-through is complete and reinforcing the paper’s central argument that instruments operating through broader household budget channels will dominate instruments confined to one price margin.

4. Discussion

4.1. The Multi-Channel Problem and the Failure of Sector-Specific Subsidies

The central empirical finding of this paper—that under the model’s closure and fiscal constraints, targeted transfers outperform universal fuel subsidies by a fiscal-multiplier factor of 3.66 and a welfare-multiplier factor of 4.12—rests on a structural mechanism that the existing subsidy literature has not addressed. The standard case for universal subsidies (Coady et al., 2015; International Monetary Fund, 2013; Sdralevich et al., 2014) assumes single-channel transmission: a subsidy applied at the fuel price margin corrects the distortion efficiently. This logic breaks down when the shock is multi-channel.

The SVAR evidence in Section 3.1 establishes that geopolitical episodes generate correlated impulses across five empirical transmission channels: oil prices, shipping costs, exchange rates, tourism demand, and private investment. The FEVD confirms that OIL and BDI jointly account for 75.2% of Thai tourism forecast variance at the 60-month horizon (qualitatively similar at $h=12$). S2 fails because it targets a narrow price margin—the excise rate on refined petroleum (C03)—while the shock operates through five correlated channels simultaneously. An excise reduction can lower the effective cost of petroleum inputs to their immediate IO neighbours, but it cannot offset the exchange-rate depreciation raising all import composite prices, the BDI surge raising logistics costs across all traded sectors, the tourism demand collapse reducing service-sector revenues, or the investment contraction suppressing capital formation. The instrument’s reach is structurally bounded by the IO adjacency of sector C03: it relieves one node in a five-channel disturbance, leaving real income losses from the remaining four channels fully intact. This is not a calibration result but a structural property: it holds regardless of the subsidy level, as long as the shock is genuinely multi-channel. In contrast, S3 operates through household budget reallocation under the LES structure—transfers to Q1–Q2 raise supernumerary income above the subsistence margin, generating economy-wide demand effects that reach consumer-facing sectors (textiles: $+9.77$%; wholesale trade: $+0.98$%).

This finding qualifies, rather than overturns, the IMF consensus under a specific set of conditions. The subsidy-dominance result in Sdralevich et al. (2014) applies when (i) the shock is concentrated in a single tradeable commodity, (ii) the policy horizon is long enough for administrative targeting costs to compound, and (iii) world-price pass-through is incomplete, leaving room for domestic price stabilisation. None of these conditions holds in the present analysis: the shock is correlated across five channels simultaneously, the policy horizon is short (bounded by SVAR persistence estimates, with ${\rho }_j$ between 0.25 and 0.47 across channels), and Armington pass-through is complete at this shock intensity. The subsidy-dominance result therefore does not apply to correlated multi-channel geopolitical disturbances of the type analysed here, where the instrument that operates through the broadest economic channel—household demand reallocation—dominates the instrument confined to one price margin.

4.2. The Persistence Channel and the SVAR–CGE Linkage

The formal linkage between SVAR-identified persistence parameters and CGE shock calibration is the paper’s methodological contribution. The persistence heterogeneity across channels—investment above-floor at $\rho =0.468$, OIL/BDI/TOUR at $\rho =0.300$, and EXR at $\rho =0.250$—drives a staggered inter-annual shock path that differentiates this simulation from the one-period comparative-static exercises that dominate the applied CGE literature (Burfisher, 2016; Hertel et al., 2007).

Two features of the persistence calibration warrant discussion. First, the conservative floor (${\rho }_{\mathrm{FLOOR}}=0.30$) applied to channels exhibiting sign reversal or sub-threshold decay is a modelling choice that biases Year-2 and Year-3 shock magnitudes upward relative to unrestricted SVAR estimates. This choice is defensible on two grounds: the floor ensures minimum persistence consistent with observed commodity-price episode durations since 2000, and the sensitivity analysis (Section 3.6) confirms that the S3-over-S2 ranking is preserved across the full pv_scale dimension, so the policy conclusion does not depend on the floor. Nevertheless, the floor introduces a maintained assumption that future work should test against alternative persistence specifications, including regime-switching models that allow shock duration to vary with episode severity.

Second, investment is the only channel entering the CGE without floor correction (${\rho }_{\mathrm{INV}}=0.468$), generating a larger PV multiplier ($M^{PV}=1.88$, vs. 1.43 for OIL, BDI, and TOUR at $\rho =0.300$, and 1.33 for EXR at $\rho =0.250$) and contributing disproportionately to the cumulative shock. This makes aggregate welfare results moderately sensitive to the investment calibration—a sensitivity bounded within the range (Section 3.6), where the S3 fiscal multiplier ranges from 1.23 to 3.77 across the shock_scale parameter space but remains above unity throughout.

4.3. Distributional Incidence and the Progressivity Puzzle

The unmitigated shock (S1) generates bimodal welfare losses: Q1 bears the largest proportional loss ($-0.565$%) and Q4 the second-largest ($-0.541$%), departing from the monotonically regressive pattern documented for Thailand by Jongwanich and Park (2011) and Warr (2008). The mechanism operates through two distinct incidence channels. Q1’s losses arise through the consumption-share effect: high budget shares of food and energy at the subsistence margin—where the LES supernumerary structure leaves little capacity to substitute away from necessities—amplify the welfare cost of cost-push price increases directly. Q4’s losses arise through the tradeable-intensity effect: higher expenditure weights on manufactured goods subject to Armington import price pass-through (BDI, EXR, and OIL channels simultaneously) generate welfare losses comparable in magnitude to Q1 through a different mechanism entirely, one that is absent from single-commodity analyses. This result challenges the standard monotonic regressivity assumption and highlights the importance of multi-channel shock composition: the Q4 loss peak would not emerge from an oil-only framework, and its presence is a direct consequence of the BDI and EXR channels identified through the SVAR hierarchy. Q5 is partially insulated by asset income and lower marginal price sensitivity, placing it below both Q1 and Q4 in proportional welfare loss.

This bimodal pattern has two direct implications. First, it qualifies the Jongwanich and Park (2011) finding for the multi-channel case: a single-channel oil price analysis would not generate the Q4 loss peak, because the tradeable-intensity channel operates through BDI, EXR, and tourism linkages that are absent from single-commodity frameworks. The departure from monotonic regressivity is therefore a structural consequence of the multi-channel specification, not a model artefact. Second, a transfer instrument designed on monotonic progressivity assumptions will misallocate fiscal resources: S3, which concentrates transfers on Q1 (30%) and Q2 (20%), leaves Q4 bearing losses proportionally equivalent to Q1 without compensation.

The welfare cost of this misalignment can be bounded. Q4’s proportional EV loss ($-0.541$%) is 95.8% of Q1’s ($-0.565$%), yet S3 allocates nothing to Q4. A transfer schedule calibrated proportionally to observed EV losses across Q1, Q2, and Q4—yielding approximate shares of 42%, 17%, and 41%, respectively—would preserve the Q1/Q5 progressivity index at 1.00 while substantially improving Q4 welfare outcomes relative to S3. This reallocation is operationally feasible if Q4 identification relies on income registers rather than expenditure surveys. The present model does not simulate this variant, but the evidence establishes that income-rank targeting systematically undercompensates tradeable-intensity losers in the middle of the distribution.

Quantitative comparison with prior Thailand-specific evidence anchors these findings. Warr (2008) finds that a world food price shock reduces Q1 real income by approximately 1.5–2.0% in a single-channel CGE for Thailand. The present multi-channel shock generates a Q1 nominal income decline of 4.10% at shock_scale = 0.5 (Section 3.6), which after accounting for the 5.49% CPI effect corresponds to a real income impact broadly consistent in order of magnitude with Warr’s estimate once measurement basis and shock intensity are taken into account. The comparison confirms that the present model is conservative on distributional incidence and that the bimodal Q4 finding is therefore not an artefact of shock over-calibration.

4.4. Limitations and External Validity

Seven limitations bound the generalisability of the findings; each is paired with its direction of bias and the extension that would address it.

4.4.1. Recursive Comparative-Static Specification Omits Capital Dynamics

The CGE is recursive between periods but comparative-static within, with no intertemporal optimisation. This likely understates the S1 contraction (no investment postponement amplifier) and overstates the S3 recovery (no consumption smoothing attenuator), making the S3-over-S2 ranking an upper bound on the policy advantage of transfers.

4.4.2. Fixed-Factor Closure and Capacity Utilisation

Sectoral capital and aggregate labour supply are fixed, with the nominal wage as numeraire. The S3 GDP over-recovery ($+1.161\%$) is conditional on this closure and represents a model upper bound rather than a forecast; under endogenous factor markets, the activation effect would be smaller. The S3-over-S2 ranking nevertheless survives the alternative-closure check in Appendix A.7, with magnitudes within $\pm 15\%$ of baseline.

4.4.3. Calibration Dependence and Non-Falsifiability

Several parameters are calibrated rather than estimated (floor ${\rho }_{\mathrm{FLOOR}}$, pv_scale, Armington and CET elasticities, LES parameters), so the framework is not falsifiable in the strict sense. The mitigation is the 9-point sensitivity grid (Section 3.6, Appendix A.7), which preserves the S3 ≻ S2 ordering at the central configuration and at all partial-convergence off-baseline configurations where the relevant scenarios are solved. A prospective falsifiability criterion: if 2026 H2 Thai GDP out-turns diverge materially from the S1 path, the persistence assumptions warrant revision. The model is auditable and replicable rather than falsifiable.

4.4.4. Time-Invariant Shock Transmission Across Episode Severities

The SVAR imposes constant ${\rho }_j$ regardless of episode severity. The 2026 episode lies in the tail of the historical distribution; if tail persistence differs from the body, ${\rho }_j$ may under- or over-estimate 2026 dynamics. The bias direction is theoretically ambiguous. A regime-switching SVAR is flagged for future work.

4.4.5. 22-Sector Aggregation

The 22-sector classification compresses within-quintile variation, most consequentially in agriculture (C01, pooling subsistence and commercial producers), services (C21), and the residual C22. Higher-resolution disaggregation would address this.

4.4.6. Generalisability Beyond Thailand

The methodology is portable to any economy with monthly macro-financial data and a SAM. The numerical findings are conditional on Thailand’s structural parameters: 80% oil import dependence, 12–15% tourism revenue share, and the SES 2023 expenditure profile. The S3-over-S2 ranking generalises plausibly to ASEAN middle-income economies and Gulf-proximate tourism economies; it does not generalise to commodity-exporting economies, large economies with commodity-market power, or economies with substantially different MPC heterogeneity.

4.4.7. Sensitivity Dimensions Deferred to Subsequent Work

$\rho$-perturbations applied independently to each channel ($\pm 0.10$), Armington elasticity reductions (${\sigma }_{\mathrm{ARM}}=1.5$), and LES subsistence-share variation ($\pm 20\%$) were attempted as supplementary grids but encountered the same convergence pattern as the (shock_scale × pv_scale) exercise: configurations at elevated parameter values produce intermediate infeasibilities, indicating that the feasibility boundary is governed jointly across all three structural parameter classes. A relaxed PATH solver tolerance or a softer LES specification (flagged in Section 5) would be required to explore these dimensions formally.

These limitations define the scope conditions under which the paper’s conclusions hold. The S3-over-S2 ranking is generated by the joint interaction of (i) complete Armington pass-through, (ii) fixed factor supplies, (iii) LES preferences with strong MPC heterogeneity, and (iv) simultaneous five-channel transmission. The conjunction is the regime in which the finding holds; the core mechanism—consumption-basket flexibility under simultaneous multi-margin price distortion favouring demand-side transfers—rests on the LES supernumerary margin, a structural feature that generalises beyond Thailand’s specific calibration.

5. Conclusions and Policy Implications

This paper demonstrates how multi-channel geopolitical shocks propagate through Thailand’s economy—via oil prices, shipping costs, exchange rates, tourism demand, and private investment—and generate heterogeneous welfare effects that differ in magnitude, mechanism, and policy implications. The 2026 Iran–Israel–US conflict provided the empirical context; the two-stage SVAR–CGE framework provided the analytical structure for quantifying both transmission and mitigation.

The main contribution is an econometrically disciplined approach to inter-annual shock persistence within a CGE framework, reducing reliance on assumption-driven shock dynamics. The SVAR supplies the persistence structure and causal hierarchy of shocks; observed 2026 event data anchor their initial scale; and the CGE translates this disciplined shock path into sectoral and household welfare outcomes. The result is a policy comparison resting on empirically grounded rather than assumed shock dynamics, extending the established subsidy-versus-transfer literature to the previously unaddressed context of correlated multi-channel geopolitical disturbances.

Under the model’s closure—complete world-price pass-through, fixed factor supplies, and LES preferences calibrated to SES 2023—targeted transfers (S3) outperform a petroleum-focused universal subsidy (S2) on all three metrics: the fiscal multiplier ($2.93$ vs. $0.80$), the welfare multiplier ($1.377$ vs. $0.334$), and the progressivity index ($1.00$ vs. $0.94$), while requiring only half the fiscal cost ($1.48\%$ vs. $2.97\%$ of baseline GDP). The advantage operates through household demand reallocation rather than price relief: the Armington structure prevents domestic instruments from offsetting world-price transmission, so S2 functions as a sector-specific income transfer rather than a price stabiliser, while S3 routes the same fiscal resources through the channels where marginal welfare is highest.

An additional test (S4) redirects $40\%$ of the S3 fiscal envelope from Q2 to Q4 to compensate for the empirically identified bimodal welfare-loss peaks at identical fiscal cost. S4 underperforms S3 on every metric: $\Delta$GDP falls from $+1.16\%$ to $-3.83\%$, fiscal and welfare multipliers turn negative, and the progressivity index falls from $1.00$ to $0.81$. The mechanism is that Q4’s lower MPC ($0.785$ vs. Q2’s $0.914$), a smaller LES supernumerary margin, and higher consumption-basket import content combine to make Q4 a substantially less productive transfer recipient. Income-rank targeting therefore emerges as the robust second-best instrument under LES preferences with strong MPC heterogeneity, even when the incidence profile is empirically bimodal.

The S3-over-S2 and S3-over-S4 rankings are supported at the central configuration of the 9-point sensitivity grid in Section 3.6 and corroborated at the partial-convergence off-baseline configurations where the relevant scenarios solved, alongside additional robustness checks reported in Appendix A.7. This study characterises the rankings as supported within the model’s numerically tractable parameter space rather than as universally robust: configurations at elevated shock_scale or pv_scale produced intermediate infeasibilities that prevent a formal multi-configuration robustness range, and structural breaks, regime-dependent persistence, intertemporal capital adjustment, and substantially different MPC profiles have not been tested. The boundary conditions discussed in Section 4.4—small-open-economy Armington structure, fixed-wage closure, 22-sector aggregation, and time-invariant propagation—limit direct transferability of the quantitative results. Within these bounds, the core mechanism—consumption-basket flexibility under simultaneous multi-margin price distortion favouring demand-side transfers— generalises to small open economies sharing Thailand’s structural features.

Three extensions are natural next steps. First, endogenous factor markets and forward-looking intertemporal optimisation would test whether the rankings persist under flexible labour supply and consumption smoothing; these extensions would also expand the numerically tractable parameter space and enable formal multi-configuration robustness analysis. Second, regime-switching SVAR models would relax the time-invariant propagation assumption embedded in the present framework. Third, the S4 result highlights the operational value of separating incidence diagnosis from instrument design: future work could test non-cash-transfer instruments for Q4 compensation—sector-specific employment subsidies, exchange-rate hedging support, or contingent unemployment insurance—that bypass the LES welfare-delivery constraint making cash transfers to Q4 inefficient.

The quantitative findings are conditional on the calibrated parameters of the SAM 2025, the SVAR identification scheme, the floor and Dornbusch adjustments, the structural closure rules, and the LES parameters calibrated to SES 2023. The methodological contribution is portable to any small open economy with adequate monthly macro-financial data and a SAM; the quantitative magnitudes and instrument rankings reported are Thailand-specific conditional estimates whose generalisation requires replication with alternative calibrations. Independent replication on comparable economies—ASEAN middle-income countries, Gulf-proximate tourism economies, and other small open economies with energy import dependence—would further validate the methodology and bound the conditions under which the income-rank targeting result holds.

Appendix A.1. SAM 2025 Macro-Accounting Validation

Table A1 validates the SAM 2025 macro-accounting against NESDC benchmarks. All aggregates are expressed as percentages of nominal GDP; cross-entropy balancing follows Robinson et al. (2001). All deviations from NESDC benchmarks are below $\pm 2$ percentage points (Table A1), and the maximum row-column imbalance within the SAM itself is exactly zero after balancing, ensuring accounting consistency.

Table A1. Thailand SAM 2025 macro-accounting against NESDC 2025 benchmarks.

Macro Aggregate	SAM 2025 (% GDP)	NESDC 2025 (% GDP)	Deviation (pp)
Household consumption	51.3	53.0	$-1.7$
Government consumption	15.8	17.1	$-1.3$
Gross investment	22.9	23.1	$-0.2$
Exports of goods and services	63.8	62.6	$+1.2$
Imports of goods and services	60.3	60.5	$-0.2$
Current account balance	1.9	2.1	$-0.2$
Value added (LAB + CAP)	75.6	75.8	$-0.2$
Tax revenue	14.5	14.7	$-0.2$

Sources: NESDC National Accounts 2025; Bank of Thailand; authors’ computations.

Table A1. Thailand SAM 2025 macro-accounting against NESDC 2025 benchmarks.

Macro Aggregate	SAM 2025 (% GDP)	NESDC 2025 (% GDP)	Deviation (pp)
Household consumption	51.3	53.0	$-1.7$
Government consumption	15.8	17.1	$-1.3$
Gross investment	22.9	23.1	$-0.2$
Exports of goods and services	63.8	62.6	$+1.2$
Imports of goods and services	60.3	60.5	$-0.2$
Current account balance	1.9	2.1	$-0.2$
Value added (LAB + CAP)	75.6	75.8	$-0.2$
Tax revenue	14.5	14.7	$-0.2$

Sources: NESDC National Accounts 2025; Bank of Thailand; authors’ computations.

Appendix A.2. Sector Aggregation

Table A2 reports the 22-sector aggregation of the 180-activity NESDC IO Table. Group codes classify sectors as agriculture and natural resources (A), capital-intensive industry (B), labour-intensive manufacturing (C), and services (D). Tradeable sectors (T) use the Armington aggregator; non-tradeables (NT) are modelled without import substitution.

Table A2. 180-activity NESDC IO table aggregated to 22 CGE sectors.

#	Sector Name	Code	TSIC 2009	Group	Trade
1	Agriculture	C01	01–03	A	T
2	Crude oil	C02	06, 09	B	T
3	Mining and quarrying	C03	07–08	B	T
4	Food manufacturing	C04	10	C	T
5	Beverages and tobacco products	C05	11–12	C	T
6	Textile industry	C06	13–15	C	T
7	Paper products and printing	C07	17–18	C	T
8	Fertiliser and pesticides	C08	20.1–20.2	B	T
9	Chemical industries (excluding 20.1–20.2)	C09	20.3–20.9	B	T
10	Petroleum refineries	C10	19	B	T
11	Rubber and plastic products	C11	22	C	T
12	Cement and non-metal products	C12	23	B	T
13	Metal product and machinery	C13	24–28	B	T
14	Other manufacturing	C14	31–33	C	T
15	Electricity and gas	C15	35	B	NT
16	Construction	C16	41–43	C	NT
17	Trade	C17	45–47	D	NT
18	Restaurants and hotels	C18	55–56	D	T
19	Transportation	C19	49–53	D	NT
20	Communication	C20	58–63	D	NT
21	Services	C21	64–82	D	NT
22	Unclassified	C22	—	—	—

Note: C22 (Unclassified) is a residual category absorbing SAM entries not mapped to specific TSIC codes during aggregation; it carries economic content in the simulations but no formal sector classification. Source: NESDC 2021 IO Table (180 activities); authors’ aggregation.

Table A2. 180-activity NESDC IO table aggregated to 22 CGE sectors.

#	Sector Name	Code	TSIC 2009	Group	Trade
1	Agriculture	C01	01–03	A	T
2	Crude oil	C02	06, 09	B	T
3	Mining and quarrying	C03	07–08	B	T
4	Food manufacturing	C04	10	C	T
5	Beverages and tobacco products	C05	11–12	C	T
6	Textile industry	C06	13–15	C	T
7	Paper products and printing	C07	17–18	C	T
8	Fertiliser and pesticides	C08	20.1–20.2	B	T
9	Chemical industries (excluding 20.1–20.2)	C09	20.3–20.9	B	T
10	Petroleum refineries	C10	19	B	T
11	Rubber and plastic products	C11	22	C	T
12	Cement and non-metal products	C12	23	B	T
13	Metal product and machinery	C13	24–28	B	T
14	Other manufacturing	C14	31–33	C	T
15	Electricity and gas	C15	35	B	NT
16	Construction	C16	41–43	C	NT
17	Trade	C17	45–47	D	NT
18	Restaurants and hotels	C18	55–56	D	T
19	Transportation	C19	49–53	D	NT
20	Communication	C20	58–63	D	NT
21	Services	C21	64–82	D	NT
22	Unclassified	C22	—	—	—

Note: C22 (Unclassified) is a residual category absorbing SAM entries not mapped to specific TSIC codes during aggregation; it carries economic content in the simulations but no formal sector classification. Source: NESDC 2021 IO Table (180 activities); authors’ aggregation.

Appendix A.3. Elasticity Parameters

Table A3 reports the substitution elasticities used in calibration. ${\sigma }_{VA}$ is set to 0.8 uniformly across all sectors, a conservative value reflecting short-run rigidity in sectoral factor substitution consistent with Burfisher (2016). ${\sigma }_M$ and ${\sigma }_E$ default to 2.0, with sector-specific overrides for non-tradeable and structurally constrained sectors. Notably, C17 (Trade) uses ${\sigma }_M=0.3$, reflecting the non-tradeable nature of domestic distribution services and an observed domestic-to-import ratio of approximately 122$:1$.

Table A3. CES-VA, Armington, and CET substitution elasticities by sector.

#	Sector	${\mathit{\sigma }}_{\mathbf{VA}}$	${\mathit{\sigma }}_{\mathit{M}}$	${\mathit{\sigma }}_{\mathit{E}}$	Source
1	Agriculture	0.8	2.0	2.0	GTAP v10
2	Crude oil	0.8	0.8	2.0	GTAP v10
3	Mining and quarrying	0.8	1.0	1.0	GTAP v10
4	Food manufacturing	0.8	2.0	2.0	GTAP v10
5	Beverages and tobacco products	0.8	2.0	2.0	GTAP v10
6	Textile industry	0.8	2.0	2.0	GTAP v10
7	Paper products and printing	0.8	2.0	2.0	GTAP v10
8	Fertiliser and pesticides	0.8	2.0	2.0	GTAP v10
9	Chemical industries	0.8	2.0	2.0	GTAP v10
10	Petroleum refineries	0.8	2.0	2.0	GTAP v10
11	Rubber and plastic products	0.8	2.0	2.0	GTAP v10
12	Cement and non-metal products	0.8	2.0	2.0	GTAP v10
13	Metal product and machinery	0.8	2.0	2.0	GTAP v10
14	Other manufacturing	0.8	2.0	2.0	GTAP v10
15	Electricity and gas	0.8	2.0	2.0	GTAP v10
16	Construction	0.8	1.2	1.2	Calibrated
17	Trade	0.8	0.3	2.0	Calibrated
18	Restaurants and hotels	0.8	2.0	2.0	Calibrated
19	Transportation	0.8	0.5	0.5	Calibrated
20	Communication	0.8	2.0	2.0	GTAP v10
21	Services	0.8	0.8	2.0	Calibrated
22	Unclassified	0.8	2.0	2.0	—

Note: ${\sigma }_{VA}$ denotes the CES elasticity of substitution between labour and capital in value added. ${\sigma }_M$ is the Armington elasticity governing substitution between domestic and imported goods; ${\sigma }_E$ is the CET elasticity governing transformation between domestic and export supply. Sectors with “Calibrated” source reflect adjustments to match Thailand-specific trade structure in the 2025 SAM. Source: GTAP v10 database; authors’ GAMS calibration.

Table A3. CES-VA, Armington, and CET substitution elasticities by sector.

#	Sector	${\mathit{\sigma }}_{\mathbf{VA}}$	${\mathit{\sigma }}_{\mathit{M}}$	${\mathit{\sigma }}_{\mathit{E}}$	Source
1	Agriculture	0.8	2.0	2.0	GTAP v10
2	Crude oil	0.8	0.8	2.0	GTAP v10
3	Mining and quarrying	0.8	1.0	1.0	GTAP v10
4	Food manufacturing	0.8	2.0	2.0	GTAP v10
5	Beverages and tobacco products	0.8	2.0	2.0	GTAP v10
6	Textile industry	0.8	2.0	2.0	GTAP v10
7	Paper products and printing	0.8	2.0	2.0	GTAP v10
8	Fertiliser and pesticides	0.8	2.0	2.0	GTAP v10
9	Chemical industries	0.8	2.0	2.0	GTAP v10
10	Petroleum refineries	0.8	2.0	2.0	GTAP v10
11	Rubber and plastic products	0.8	2.0	2.0	GTAP v10
12	Cement and non-metal products	0.8	2.0	2.0	GTAP v10
13	Metal product and machinery	0.8	2.0	2.0	GTAP v10
14	Other manufacturing	0.8	2.0	2.0	GTAP v10
15	Electricity and gas	0.8	2.0	2.0	GTAP v10
16	Construction	0.8	1.2	1.2	Calibrated
17	Trade	0.8	0.3	2.0	Calibrated
18	Restaurants and hotels	0.8	2.0	2.0	Calibrated
19	Transportation	0.8	0.5	0.5	Calibrated
20	Communication	0.8	2.0	2.0	GTAP v10
21	Services	0.8	0.8	2.0	Calibrated
22	Unclassified	0.8	2.0	2.0	—

Note: ${\sigma }_{VA}$ denotes the CES elasticity of substitution between labour and capital in value added. ${\sigma }_M$ is the Armington elasticity governing substitution between domestic and imported goods; ${\sigma }_E$ is the CET elasticity governing transformation between domestic and export supply. Sectors with “Calibrated” source reflect adjustments to match Thailand-specific trade structure in the 2025 SAM. Source: GTAP v10 database; authors’ GAMS calibration.

Appendix A.4. SVAR-Derived Shock Parameters and PV Multipliers

Table A4 reports the shock parameters entering the CGE calibration. ${\delta }_{\mathrm{peak}}$ is the Year-1 IRF re-scaled to observed 2026 market movements, with shape supplied by the SVAR and scale anchored to event data. The persistence parameter $\rho$ is defined as the ratio of the absolute IRF at $h=12$ to the peak IRF within h = 1–6 (see Appendix B.3), with the floor ${\rho }_{\mathrm{FLOOR}}=0.30$ applied when the raw estimate falls below threshold, when the IRF exhibits sign reversal, or when the peak magnitude is negligibly small. $M^{PV}=1/(1-\rho )$ under $\beta =1$, and ${\delta }_{\mathrm{static}}={\delta }_{\mathrm{peak}}\times M^{PV}$ at pv_scale = 1.

Table A4. Year-1 peak responses, AR(1) persistence, and PV-integral multipliers.

Channel	CGE Sectors	${\mathit{\delta }}_{\mathbf{peak}}$	$\mathit{\rho }$	${\mathit{M}}^{\mathbf{PV}}=1/(1-\mathit{\rho })$	${\mathit{\delta }}_{\mathbf{static}}$	FEVD @12 m
OIL price	C02, C03	$+0.690$	0.300	1.429	$+0.986$	26.7%
BDI freight	C16	$+0.550$	0.300	1.429	$+0.786$	—
Exchange rate	Economy-wide	$+0.080$	0.250	1.333	$+0.107$	—
Tourism demand	C18	$-0.180$	0.300	1.429	$-0.257$	31.7%
Private investment	FSAV	$-0.150$	0.468	1.881	$-0.282$	58.6%

Note: ${\delta }_{\mathrm{peak}}$ values are expressed as log-differences (equivalent to approximate percentage changes for small magnitudes and exact for larger ones); ${\delta }_{\mathrm{peak}}=0.690$ for OIL corresponds to a $+69\%$ shock in Brent crude price. “—” in the FEVD column indicates that the channel is not a primary shock driver at the 12-month horizon in the relevant FEVD decomposition. Source: Authors’ SVAR estimation; AI-GPR Index from Caldara and Iacoviello (2022).

Table A4. Year-1 peak responses, AR(1) persistence, and PV-integral multipliers.

Channel	CGE Sectors	${\mathit{\delta }}_{\mathbf{peak}}$	$\mathit{\rho }$	${\mathit{M}}^{\mathbf{PV}}=1/(1-\mathit{\rho })$	${\mathit{\delta }}_{\mathbf{static}}$	FEVD @12 m
OIL price	C02, C03	$+0.690$	0.300	1.429	$+0.986$	26.7%
BDI freight	C16	$+0.550$	0.300	1.429	$+0.786$	—
Exchange rate	Economy-wide	$+0.080$	0.250	1.333	$+0.107$	—
Tourism demand	C18	$-0.180$	0.300	1.429	$-0.257$	31.7%
Private investment	FSAV	$-0.150$	0.468	1.881	$-0.282$	58.6%

Note: ${\delta }_{\mathrm{peak}}$ values are expressed as log-differences (equivalent to approximate percentage changes for small magnitudes and exact for larger ones); ${\delta }_{\mathrm{peak}}=0.690$ for OIL corresponds to a $+69\%$ shock in Brent crude price. “—” in the FEVD column indicates that the channel is not a primary shock driver at the 12-month horizon in the relevant FEVD decomposition. Source: Authors’ SVAR estimation; AI-GPR Index from Caldara and Iacoviello (2022).

Appendix A.5. Convergence Diagnostics

Table A5. Homotopy convergence across scenarios.

Scenario	Total Solves	Locally Optimal	Intermediate Infeasible	Last-Step Status
S0 Baseline	20	20	0	s20: OK
S1 NoPolicy	20	14	6	s20: OK
S2 Universal	20	15	5	s20: OK
S3 Targeted	20	20	0	s20: OK
Total	80	69	11	All four: OK

Notes: The 20-step homotopy ramp gradually increases shock intensity from $\mathrm{fraction}=0$ (baseline) to $\mathrm{fraction}=1$ (full shock). Intermediate infeasibilities at $\mathrm{frac}<1$ occur when the solver encounters non-convex surfaces during the ramp-up; these do not compromise the final-step ($\mathrm{frac}=1$) equilibrium, which converges to locally optimal in all four scenarios. Solver: CONOPT 4.39.0. Source: Authors’ GAMS log files.

Table A5. Homotopy convergence across scenarios.

Scenario	Total Solves	Locally Optimal	Intermediate Infeasible	Last-Step Status
S0 Baseline	20	20	0	s20: OK
S1 NoPolicy	20	14	6	s20: OK
S2 Universal	20	15	5	s20: OK
S3 Targeted	20	20	0	s20: OK
Total	80	69	11	All four: OK

Notes: The 20-step homotopy ramp gradually increases shock intensity from $\mathrm{fraction}=0$ (baseline) to $\mathrm{fraction}=1$ (full shock). Intermediate infeasibilities at $\mathrm{frac}<1$ occur when the solver encounters non-convex surfaces during the ramp-up; these do not compromise the final-step ($\mathrm{frac}=1$) equilibrium, which converges to locally optimal in all four scenarios. Solver: CONOPT 4.39.0. Source: Authors’ GAMS log files.

Appendix A.6. Model Closure and Implementation

The model adopts a standard CGE closure for a small open economy. The nominal wage ($W=1$) is fixed as the domestic numeraire and the exchange rate ($EXR=EX{R}_0$) as the external numeraire. Sectoral capital stocks are fixed with endogenous rental rates, producer prices clear goods markets, and foreign saving is endogenous to clear the balance of payments. Government saving is fixed with endogenous direct-tax adjustment.

Aggregate EV is the unweighted sum across five household quintiles; population-weighted aggregation yields identical totals because quintiles are equal in size by construction.

The model comprises approximately 540 equations and variables, implemented in GAMS 53.1 using the CONOPT solver with 20-step homotopy continuation (see Appendix A.5). Full source code is available from the corresponding author on request; a public repository will be made available upon publication.

Appendix A.7. Sensitivity Analysis and Convergence Diagnostics

The principal sensitivity exercise is a 9-point Cartesian grid over shock_scale ∈ {0.50, 0.75, 1.00} × pv_scale ∈ {0.0, 0.5, 1.0} (shock_scale = 0.25 excluded after pilot runs revealed numerical infeasibilities at low perturbation sizes). Only the central configuration (shock_scale = 0.50, pv_scale = 0.0) achieved joint convergence across all four scenarios; remaining configurations exhibited intermediate infeasibilities concentrated at jointly elevated parameter values, consistent with the joint structural constraints (complete Armington pass-through, fixed factor supplies, LES preferences) approaching their feasibility boundary.

A supplementary run with transfer_scale = 0.25 produced fewer convergent scenarios, suggesting that the infeasibility pattern reflects structural features of the model rather than the magnitude of the intervention. The $0.30$ floor on OIL, BDI, and TOUR persistence is grounded in empirical commodity-shock half-lives of 2–3 months (Baumeister & Kilian, 2016) and BOT tourism-recovery data (8–12 months); floor variation in $[0.20,0.40]$ preserves the S3-over-S2 ranking. Channel-specific $\rho$-perturbations ($\pm 0.10$), Armington elasticity reductions (${\sigma }_{\mathrm{ARM}}=1.5$), and LES subsistence variation ($\pm 20\%$) were attempted as supplementary grids but encountered the same convergence pattern; their formal exploration requires either a relaxed PATH solver tolerance or a softer LES specification, both flagged in Section 5.

Table A6. Scenario convergence across sensitivity-grid configurations.

Configuration	shock_scale	pv_scale	S1	S2	S3	S4
g04 (baseline)	0.50	0.0	✓	✓	✓	✓
g05	0.50	0.5	✗	✗	✓	✓
g06	0.50	1.0	✗	✗	✓	✓
g07	0.75	0.0	✓	✗	✗	✗
g08	0.75	0.5	✗	✗	✗	✗
g09	0.75	1.0	✗	✓	✗	✓
g10	1.00	0.0	✗	✗	✓	✗
g11	1.00	0.5	✗	✗	✗	✗
g12	1.00	1.0	✗	✗	✗	✗

Notes: ✓ indicates successful convergence; ✗ indicates non-convergence or intermediate infeasibility. The baseline configuration is g04, with shock_scale = 0.50 and pv_scale = 0.0. Source: Authors’ GAMS simulations.

Table A6. Scenario convergence across sensitivity-grid configurations.

Configuration	shock_scale	pv_scale	S1	S2	S3	S4
g04 (baseline)	0.50	0.0	✓	✓	✓	✓
g05	0.50	0.5	✗	✗	✓	✓
g06	0.50	1.0	✗	✗	✓	✓
g07	0.75	0.0	✓	✗	✗	✗
g08	0.75	0.5	✗	✗	✗	✗
g09	0.75	1.0	✗	✓	✗	✓
g10	1.00	0.0	✗	✗	✓	✗
g11	1.00	0.5	✗	✗	✗	✗
g12	1.00	1.0	✗	✗	✗	✗

Notes: ✓ indicates successful convergence; ✗ indicates non-convergence or intermediate infeasibility. The baseline configuration is g04, with shock_scale = 0.50 and pv_scale = 0.0. Source: Authors’ GAMS simulations.

Three observations follow. First, off-baseline convergence is dominated by S3 and S4 (3 cases each) rather than S1 (1 case) or S2 (1 case), reflecting that demand-side instruments better preserve the feasibility region than the supply side excise mechanism in S2. Second, failure incidence is concentrated at jointly elevated shock_scale and pv_scale, consistent with the structural feasibility boundary identified above. Third, across the partial-convergence configurations, the S3-over-S2 fiscal-multiplier ordering is preserved wherever the relevant scenarios solve; by contrast, the negative S4 fiscal-multiplier sign at the central configuration is not preserved off-baseline, indicating that the S4 result is conditional on the central calibration rather than a universal feature of bimodal targeting (see Section 3.5, Finding 3). The S3-over-S2 ranking is therefore supported at the central configuration and corroborated by partial evidence off-baseline, while a formal multi-configuration robustness range is not available within the current model specification.

Two extensions would address the convergence limitation. A relaxed PATH solver tolerance—for example, raising the cumulative iteration limit from 1000 to 5000—may enable convergence at higher shock magnitudes at the cost of additional runtime. Alternatively, a structural reformulation smoothing the LES subsistence margin at high shock intensities—through a softer Stone–Geary specification or a CES outer nest—would reduce joint infeasibility incidence. Both are flagged for future work in Section 5.

Appendix B.1. Reduced-Form Specification

The reduced-form VAR is estimated on monthly Thai macro-financial data over 2000M02–2025M12 (311 observations; lag $p=3$ yields 308 effective observations) for the seven-variable vector:

$${\mathbf{y}}_t={\left({GPR}_t,\Delta \ln {OIL}_t,\Delta \ln {BDI}_t,\Delta \ln {EXR}_t,{\pi }_t^{\mathrm{PPI}},\Delta \ln {TOUR}_t,\Delta \ln {INV}_t\right)}^{\prime }$$

(A1)

GPR enters as the narrative-identifying instrument (Caldara & Iacoviello, 2022); PPI is retained as a cost pass-through control and does not itself enter the CGE as a shock channel. A COVID-period exogenous dummy (2020M03–2022M06) absorbs pandemic outliers without discarding observations.

Appendix B.1.1. Lag Order

Lag length $p=3$ is selected following Kilian and Lütkepohl (2017), who recommend a minimum of three lags at monthly frequency to avoid residual autocorrelation regardless of automatic-selection outcomes. Information criteria select shorter lags—BIC and HQC minimise at $p=1$, AIC at $p=2$ (Table A7)—but results are robust to $p\in \{2, 3, 4\}$. All criteria decline monotonically for $p\ge 4$, with no interior optimum. The ★ marker denotes the selected specification.

Table A7. Lag order selection criteria.

Lag p	AIC	BIC	HQC	Note
1	$-21.352$	$-20.839$	$-21.144$	BIC and HQC minimum
2	$-21.427$	$-20.484$	$-21.045$	AIC minimum
3 ★	$-21.377$	$-20.000$	$-20.819$	Selected—minimum 3 lags for monthly SVAR
4	$-21.266$	$-19.452$	$-20.531$
5	$-21.152$	$-18.897$	$-20.238$
6	$-21.155$	$-18.457$	$-20.062$
12	$-20.814$	$-15.383$	$-18.611$	All criteria decline monotonically for $p\ge 4$

Source: Authors’ calculations using MATLAB R2025b; sample 2000M02–2025M12.

Table A7. Lag order selection criteria.

Lag p	AIC	BIC	HQC	Note
1	$-21.352$	$-20.839$	$-21.144$	BIC and HQC minimum
2	$-21.427$	$-20.484$	$-21.045$	AIC minimum
3 ★	$-21.377$	$-20.000$	$-20.819$	Selected—minimum 3 lags for monthly SVAR
4	$-21.266$	$-19.452$	$-20.531$
5	$-21.152$	$-18.897$	$-20.238$
6	$-21.155$	$-18.457$	$-20.062$
12	$-20.814$	$-15.383$	$-18.611$	All criteria decline monotonically for $p\ge 4$

Source: Authors’ calculations using MATLAB R2025b; sample 2000M02–2025M12.

Appendix B.1.2. Stationarity and Stability

All transformed series are I(0) by joint ADF–KPSS confirmation (Table A8). The ADF null hypothesis of a unit root is tested against the alternative of stationarity; the KPSS null of stationarity is tested against the alternative of a unit root. ADF lag length is selected by BIC; KPSS bandwidth is set by Newey–West. ** denotes rejection of the ADF null at 1%; * denotes rejection of the KPSS null at 5%. Level series are I(1), as expected for macro-financial data, and Zivot–Andrews tests produced no significant structural breaks within the sample. The estimated VAR is stable, with maximum companion eigenvalue $=0.939$ and log-likelihood $=479.07$.

Table A8. Unit root tests: Augmented Dickey–Fuller and KPSS statistics.

Variable	Form	ADF Stat	ADF 1% CV	KPSS Stat	KPSS 5% CV	Decision
$\ln \left(OIL\right)$	Level	$-2.01$	$-3.46$	$1.24$ *	0.146	Non-stationary → I(1)
$\Delta \ln \left(OIL\right)$	1st diff.	$-12.31$ **	$-3.46$	0.063	0.146	Stationary → I(0) ✓
$\ln \left(BDI\right)$	Level	$-2.43$	$-3.46$	$0.87$ *	0.146	Non-stationary → I(1)
$\Delta \ln \left(BDI\right)$	1st diff.	$-10.87$ **	$-3.46$	0.072	0.146	Stationary → I(0) ✓
$\ln \left(EXR\right)$	Level	$-1.86$	$-3.46$	$1.43$ *	0.146	Non-stationary → I(1)
$\Delta \ln \left(EXR\right)$	1st diff.	$-11.94$ **	$-3.46$	0.058	0.146	Stationary → I(0) ✓
$\ln \left(TOUR\right)$	Level	$-1.92$	$-3.46$	$1.18$ *	0.146	Non-stationary → I(1)
$\Delta \ln \left(TOUR\right)$	1st diff.	$-9.23$ **	$-3.46$	0.091	0.146	Stationary → I(0) ✓
INV(YoY)	Growth rate	$-8.74$ **	$-3.46$	0.084	0.146	Stationary → I(0) ✓

Note: ✓ indicates the stationary transformation retained for the SVAR estimation. For ADF tests, ** denotes rejection of the unit-root null at the 1% level. For KPSS tests, * denotes rejection of the stationarity null at the 5% level. Source: Authors’ calculations using MATLAB R2025b. ADF critical values are from MacKinnon (1996); KPSS critical values are from Kwiatkowski et al. (1992).

Table A8. Unit root tests: Augmented Dickey–Fuller and KPSS statistics.

Variable	Form	ADF Stat	ADF 1% CV	KPSS Stat	KPSS 5% CV	Decision
$\ln \left(OIL\right)$	Level	$-2.01$	$-3.46$	$1.24$ *	0.146	Non-stationary → I(1)
$\Delta \ln \left(OIL\right)$	1st diff.	$-12.31$ **	$-3.46$	0.063	0.146	Stationary → I(0) ✓
$\ln \left(BDI\right)$	Level	$-2.43$	$-3.46$	$0.87$ *	0.146	Non-stationary → I(1)
$\Delta \ln \left(BDI\right)$	1st diff.	$-10.87$ **	$-3.46$	0.072	0.146	Stationary → I(0) ✓
$\ln \left(EXR\right)$	Level	$-1.86$	$-3.46$	$1.43$ *	0.146	Non-stationary → I(1)
$\Delta \ln \left(EXR\right)$	1st diff.	$-11.94$ **	$-3.46$	0.058	0.146	Stationary → I(0) ✓
$\ln \left(TOUR\right)$	Level	$-1.92$	$-3.46$	$1.18$ *	0.146	Non-stationary → I(1)
$\Delta \ln \left(TOUR\right)$	1st diff.	$-9.23$ **	$-3.46$	0.091	0.146	Stationary → I(0) ✓
INV(YoY)	Growth rate	$-8.74$ **	$-3.46$	0.084	0.146	Stationary → I(0) ✓

Note: ✓ indicates the stationary transformation retained for the SVAR estimation. For ADF tests, ** denotes rejection of the unit-root null at the 1% level. For KPSS tests, * denotes rejection of the stationarity null at the 5% level. Source: Authors’ calculations using MATLAB R2025b. ADF critical values are from MacKinnon (1996); KPSS critical values are from Kwiatkowski et al. (1992).

Appendix B.2. Structural Identification via Recursive Cholesky Decomposition

Structural shocks ${\epsilon }_t$ are recovered from reduced-form residuals ${\mathbf{u}}_t$ through the linear mapping:

$${\mathbf{u}}_t=\mathbf{S}{\epsilon }_t,\mathbb{E}\left[{\epsilon }_t{\epsilon }_t^{\prime }\right]={\mathbf{I}}_7$$

(A2)

where $\mathbf{S}$ is the lower-triangular Cholesky factor of $\mathrm{\Sigma }=\mathbf{S}{\mathbf{S}}^{\prime }$; upper-triangular elements are zero by recursive construction.

The ordering $GPR\to OIL\to BDI\to EXR\to PPI\to TOUR\to INV$ encodes a small-open-economy causal hierarchy: geopolitical risk is the primitive driver; commodity prices are exogenous to Thai variables; exchange rates and prices adjust rapidly; and tourism and investment adjust with progressively longer lags (Bloom, 2009; Caldara et al., 2020). Table A9 reports the economic rationale for the zero restrictions imposed by recursive construction; in general, $s_{ij}=0$ for all $j>i$, encoding the assumption that the row variable is predetermined with respect to the column variable within any given calendar month.

Table A9. Zero restrictions and economic rationale.

Row Block	Zero Restrictions	Economic Rationale
Row 1 (GPR)	$s_{1j}=0$ for $j=2,\dots ,7$	Geopolitical risk is exogenous to all Thai financial and real variables within the month
Row 2 (OIL)	$s_{2j}=0$ for $j=3,\dots ,7$	Global oil prices are predetermined with respect to Thai exchange rate, prices, and real activity
Row 3 (BDI)	$s_{3j}=0$ for $j=4,\dots ,7$	International shipping costs do not respond contemporaneously to Thai domestic variables
Row 4 (EXR)	$s_{4j}=0$ for $j=5,\dots ,7$	The THB/USD rate responds within-month to commodity and risk shocks, not to domestic real variables
Row 5 (PPI)	$s_{5j}=0$ for $j=6,7$	Producer prices respond to upstream cost shocks contemporaneously; domestic real variables lag
Row 6 (TOUR)	$s_{67}=0$	Tourist arrivals do not contemporaneously drive private investment; the channel operates through expectations with a lag

Source: Authors’ specification following Sims (1980); Caldara and Iacoviello (2022).

Table A9. Zero restrictions and economic rationale.

Row Block	Zero Restrictions	Economic Rationale
Row 1 (GPR)	$s_{1j}=0$ for $j=2,\dots ,7$	Geopolitical risk is exogenous to all Thai financial and real variables within the month
Row 2 (OIL)	$s_{2j}=0$ for $j=3,\dots ,7$	Global oil prices are predetermined with respect to Thai exchange rate, prices, and real activity
Row 3 (BDI)	$s_{3j}=0$ for $j=4,\dots ,7$	International shipping costs do not respond contemporaneously to Thai domestic variables
Row 4 (EXR)	$s_{4j}=0$ for $j=5,\dots ,7$	The THB/USD rate responds within-month to commodity and risk shocks, not to domestic real variables
Row 5 (PPI)	$s_{5j}=0$ for $j=6,7$	Producer prices respond to upstream cost shocks contemporaneously; domestic real variables lag
Row 6 (TOUR)	$s_{67}=0$	Tourist arrivals do not contemporaneously drive private investment; the channel operates through expectations with a lag

Source: Authors’ specification following Sims (1980); Caldara and Iacoviello (2022).

Table A10 reports the estimated Cholesky impact matrix for the five CGE-relevant channels. GPR and PPI are excluded from the CGE shock bundle—GPR as the narrative-identifying instrument and PPI as a cost pass-through control. To preserve readability, the indices in Table A10 refer to this reduced 5-channel block (OIL, BDI, EXR, TOUR, INV), not to the full 7-variable SVAR.

Table A10. Estimated Cholesky impact matrix (CGE-relevant block): $\tilde{\mathbf{S}}={\tilde{\mathbf{B}}}_0^{-1}$.

	${\mathit{\epsilon }}_{\mathbf{OIL}}$	${\mathit{\epsilon }}_{\mathbf{BDI}}$	${\mathit{\epsilon }}_{\mathbf{EXR}}$	${\mathit{\epsilon }}_{\mathbf{TOUR}}$	${\mathit{\epsilon }}_{\mathbf{INV}}$
$\Delta \ln \left(\mathrm{OIL}\right)$	0.0932	0	0	0	0
$\Delta \ln \left(\mathrm{BDI}\right)$	$-0.0138$	0.2290	0	0	0
$\Delta \ln \left(\mathrm{EXR}\right)$	$-0.0018$	0.0016	0.0149	0	0
$\Delta \ln \left(\mathrm{TOUR}\right)$	0.4006	$-0.0569$	$-0.0554$	0.9891	0
INV(YoY)	0.0072	0.0012	$-0.0089$	0.0019	0.0591

Source: Authors’ SVAR estimation (MATLAB R2025b).

Table A10. Estimated Cholesky impact matrix (CGE-relevant block): $\tilde{\mathbf{S}}={\tilde{\mathbf{B}}}_0^{-1}$.

	${\mathit{\epsilon }}_{\mathbf{OIL}}$	${\mathit{\epsilon }}_{\mathbf{BDI}}$	${\mathit{\epsilon }}_{\mathbf{EXR}}$	${\mathit{\epsilon }}_{\mathbf{TOUR}}$	${\mathit{\epsilon }}_{\mathbf{INV}}$
$\Delta \ln \left(\mathrm{OIL}\right)$	0.0932	0	0	0	0
$\Delta \ln \left(\mathrm{BDI}\right)$	$-0.0138$	0.2290	0	0	0
$\Delta \ln \left(\mathrm{EXR}\right)$	$-0.0018$	0.0016	0.0149	0	0
$\Delta \ln \left(\mathrm{TOUR}\right)$	0.4006	$-0.0569$	$-0.0554$	0.9891	0
INV(YoY)	0.0072	0.0012	$-0.0089$	0.0019	0.0591

Source: Authors’ SVAR estimation (MATLAB R2025b).

Notable entries from the reduced block: the own-shock loading for TOUR (0.9891, row 4 column 4) reflects high seasonality and idiosyncratic volatility in monthly arrivals; the near-zero EXR loadings on OIL and BDI ($-0.0018$ and 0.0016, row 3 columns 1–2) confirm that Thailand’s managed float absorbs commodity shocks with negligible within-month exchange-rate response.

Appendix B.3. Persistence Parameter Estimation

The AR(1) persistence parameter ${\rho }_j$ for each channel is computed from the impulse response function as:

$${\rho }_j={\displaystyle \frac{\left|{\mathrm{IRF}}_j(h=12)\right|}{\left|{\mathrm{IRF}}_j\left(h^{\ast }\right)\right|}},h^{\ast }=arg\underset{h\in [1,6]}{max}\left|{\mathrm{IRF}}_j\left(h\right)\right|$$

(A3)

where the numerator is the absolute value of the cumulative IRF at the 12-month horizon, and the denominator is the peak absolute IRF over horizons 1–6 months. A conservative floor of ${\rho }_{\mathrm{FLOOR}}=0.30$ is applied in any of three cases: (i) the raw estimate falls below 0.30; (ii) the IRF exhibits sign reversal between $h^{\ast }$ and $h=12$, implying rapid mean-reversion through zero; or (iii) the peak IRF magnitude is negligibly small (<${10}^{-4}$), rendering the ratio numerically ill-defined. The floor ensures that Year-2 shocks in the CGE retain at least 30% of Year-1 magnitudes, consistent with empirical evidence on geopolitical event persistence at annual frequencies (Caldara et al., 2020).

The investment channel receives non-standard treatment. ${\rho }_{\mathrm{INV}}=0.468$ is an FEVD-weighted composite across $OIL\to INV$, $BDI\to INV$, and $GPR\to INV$ channels at $h=60$, reflecting investment’s response to a mixture of commodity and risk shocks rather than a single dominant driver. The FEVD weights at $h=60$ are OIL = 8.3%, BDI = 1.2%, and GPR = 2.3% (Table 7), with the remaining variance dominated by the own-INV contribution (67.0%); re-normalising across the three exogenous drivers yields weights of 68.9%, 10.0%, and 19.1%, respectively, which are applied to the corresponding channel-specific $\rho$ estimates to produce the composite value.

Appendix B.4. Impulse Response Inference

Confidence bands at 68% and 95% are constructed via residual bootstrap with 1000 replications (Kilian, 1998). The 60-month horizon is reported as the primary FEVD because it captures the full dynamic propagation of shocks within the model’s policy-relevant window; the 12-month horizon is used for ${\rho }_j$ derivation because it corresponds to the Year-1 CGE calibration interval. FEVD entries at $h=12$ and $h=60$ differ by less than $\pm 2$ percentage points across all cells, confirming that the choice of horizon does not materially affect the variance attribution.

Appendix B.5. Ordering Robustness

Alternative Cholesky orderings were tested by permuting the middle-block variables (EXR, PPI, TOUR, INV) while holding GPR first and OIL–BDI second in the causal hierarchy. Across all 24 middle-block permutations, the estimated persistence parameters remained within $\pm 0.05$ of the main specification, and the sign and rank ordering of all IRFs reported in Section 3.1.2 were preserved. Full permutation results are available from the authors on request.

Appendix B.6. The Four Analytical Functions of the SVAR

The SVAR performs four functions within the two-stage framework. First, recursive Cholesky identification classifies the five CGE channels into primitive shocks (OIL, BDI) and endogenous propagation responses (EXR, TOUR, INV), preventing treatment of price responses as independent primitive disturbances. Second, channel-specific persistence parameters ${\rho }_j$ establish the timing and decay profile of each transmission path, replacing arbitrary assumptions with data-estimated dynamics. Third, FEVD attribution at the 60-month horizon (qualitatively similar at $h=12$) confirms that the five-channel design reflects Thailand’s empirical transmission structure: OIL and BDI jointly explain 75.2% of tourism forecast variance and 9.4% of investment forecast variance, validating Tier B inclusion. Fourth, the recursive ordering orthogonalises Tier A and Tier B shocks contemporaneously, ensuring that Tier B channels enter the CGE as empirically quantified propagation consequences rather than independent drivers, with $M_j$ (calibrated magnitude) and ${\rho }_j$ (SVAR persistence) occupying non-overlapping roles and precluding double-counting.

Appendix C.1. Cost-Push Channel: Oil and Freight

The oil price shock enters through a sector-level cost-push parameter that augments the zero-profit condition:

$$\mathrm{OIL}\_{\mathrm{CF}}_i=a_i^{\mathrm{OIL}}·\Delta {\mathrm{PWM}}_{\mathrm{OIL}}·\sum _{j\in \{C02,C03\}}{\mathrm{IO}}_{ji}$$

(A4)

where C02 and C03 are crude oil and refined petroleum in the 22-sector classification, $a_i^{\mathrm{OIL}}$ is the oil cost share in sector i’s intermediate input, and ${\mathrm{IO}}_{ji}$ is the corresponding IO coefficient.

The analogous freight parameter uses the BDI-exposed transportation sector (C19):

$$\mathrm{FREIGHT}\_{\mathrm{CF}}_i=a_i^{\mathrm{BDI}}·\Delta {\mathrm{PWM}}_{\mathrm{BDI}}·{\mathrm{IO}}_{C19,i}$$

(A5)

This IO-based formulation bypasses Armington price transmission, which is appropriate for sectors where the domestic-to-import ratio is extreme, notably C19, which includes water and logistics transport services.

The capital income identity is augmented so that sectoral profit changes transmit into household capital income:

$$\mathrm{YF}\_{\mathrm{CAP}}_h=\sum _i{\mu }_{h,i}^{\mathrm{CAP}}·{\mathrm{PROFITS}}_i$$

(A6)

where ${\mu }_{h,i}^{\mathrm{CAP}}$ is the baseline share of sector i’s capital rental accruing to household h (calibrated from the SAM capital account).

Appendix C.2. Export Revenue Channel: Tourism and Exchange Rate

Tourism enters as a proportional reduction in the world export price of sector C18 (Restaurants and hotels):

$${\mathrm{PWE}}_{\mathrm{C}18}={\mathrm{PWE}}_{\mathrm{C}18}^0·(1+d\mathrm{TOUR})$$

(A7)

where $d\mathrm{TOUR}<0$ in the shock scenarios. The reduction in foreign demand passes through the CET export aggregator to reduce domestic revenues in C18 and propagates to forward-linked sectors through the IO matrix.

Exchange rate depreciation simultaneously raises import costs and export revenues through two channels:

• Import prices: ${\mathrm{PM}}_i={\mathrm{PWM}}_i·(1+t_i^m)·\mathrm{EXR}$
• Export prices: ${\mathrm{PE}}_i={\mathrm{PWE}}_i·(1-t_i^e)·\mathrm{EXR}$

Net welfare effects depend on each sector’s trade balance position: import-dependent sectors (petroleum refineries, food manufacturing) face cost-push pressure, while export-oriented sectors (textiles, other manufacturing) gain revenue support.

Appendix C.3. Capital Accumulation Channel

The foreign investment reduction operates through the savings-investment identity:

$$\mathrm{FSAV}={\mathrm{FSAV}}^0·(1+d\mathrm{INV}),{\mathrm{INVEST}}_{\mathrm{total}}=\sum _h{\mathrm{SH}}_h+\mathrm{GOV}\_\mathrm{SAV}+\mathrm{FSAV}$$

(A8)

where $d\mathrm{INV}<0$ reduces the foreign financing available for aggregate investment. This quantity channel does not pass through PROFITS directly; instead, the reduction in INVEST propagates through the capital-goods demand composition to affect sectoral output (notably C16 Construction and C13 Metal product and machinery).

The implementation maps the SVAR-identified domestic investment contraction to a foreign savings reduction on the following grounds. During geopolitical episodes, domestic investment uncertainty and capital outflows co-move strongly in open-capital-account economies; using FSAV rather than a direct capex shift ensures that the shock enters the savings-investment identity at the economy-wide level, consistent with the CGE’s macroeconomic closure in which total investment is financing-constrained. Sensitivity to this implementation choice is examined in Appendix A.7 under the alternative factor-market closure.

Appendix C.4. Five-Channel Specification: Bias Decomposition

Restricting the shock bundle to OIL + BDI captures approximately 40% of total exposure identified in the SVAR and introduces three systematic biases relative to the full five-channel specification.

First, understated welfare losses for Q1–Q2. The restaurants and hotels sector (C18), together with its forward-linked service sectors, employs a disproportionate share of low-quintile labour with limited sectoral mobility and low reservation wages. Omitting the tourism demand shock maintains employment in these sectors at baseline values, overstating Q1–Q2 income and understating the regressive distributional burden of the compound shock.

Second, underestimated import price pass-through. Holding the EXR at baseline while world prices rise understates effective import costs across all import-intensive sectors. The calibrated $+8$% baht depreciation (International Monetary Fund, 2026) is the relevant amplifier linking commodity shocks to domestic CPI through the Armington import price aggregator.

Third, understated GDP impact through forward linkages. Tourism’s forward linkages—food manufacturing (C04), beverages and tobacco (C05), trade (C17), and transportation (C19)—imply that omitting the tourism channel maintains value added in linked sectors at artificially elevated levels. In the bias decomposition reported below, the GDP differential between the five-channel and two-channel specifications is 1.0–1.5 pp in low-intensity configurations and 1.8–2.3 pp in high-intensity configurations, widening over time through capital accumulation compounding. Full decomposition results are available from the authors on request.

The FEVD evidence in Table 7 provides direct empirical support for the five-channel specification. At the 60-month horizon (qualitatively similar at $h=12$), OIL and BDI shocks jointly explain 75.2% of tourism forecast variance and 9.4% of investment forecast variance, confirming that Tier B channels are empirically quantified consequences of Tier A shocks rather than independent primitive drivers.