Regime-Dependent Financial Inclusion, Energy Intensity, and Trade Openness in Saudi Arabia: An ARDL–Structural Break Analysis of CO₂ Emissions and the Sustainable Development Goals

Abstract

Background: Whether financial deepening and trade integration support or hinder environmental sustainability in hydrocarbon-dependent economies remains contested. Methods: This study examines the relationships among financial inclusion, energy intensity, trade openness, and CO₂ emissions per capita in Saudi Arabia for 1980–2020. The empirical strategy combines ARDL bounds testing, FMOLS, DOLS, CCR robustness, Toda–Yamamoto causality, and a battery of structural-break tests comprising Zivot–Andrews unit-root tests, Bai–Perron sup-F tests, and Chow tests. To address the mechanical correlation between carbon productivity and GDP, the per capita emissions specification (LNCP) is used as the primary outcome; carbon productivity (LNES) is reported for robustness. The small-sample sub-period results are stress-tested using ridge regression, residual-bootstrap confidence intervals, a GDP-augmented (scale-control) specification, and a break-date sensitivity analysis. Results: Cointegration is established. The Chow test identifies a significant break in the cointegrating relationship at 2001 (F = 7.36, p < 0.001 for LNCP), supported by the Zivot–Andrews endogenous-break dates for the financial-inclusion series (2000) and trade-openness series (2005), and by the Bai–Perron sup-F test (sup-F = 26.37 at 1990, exceeding the 1% Andrews critical value). Sub-sample re-estimation around 2001 shows that energy intensity, urbanisation, and trade openness are robust drivers of per capita emissions only after the break, while financial inclusion is statistically insignificant in both regimes once the GDP–carbon-productivity mechanical relationship is removed. Conclusions: The Saudi finance–environment relationship is structurally unstable, and policy assessments based on full-sample averages can be misleading. The evidence is best read as describing regime-dependent, conditional long-run associations rather than as identifying structural causal effects. By exposing the interactions, synergies, and trade-offs among financial deepening (SDG 8), energy efficiency (SDG 7), sustainable consumption and production (SDG 12), and climate action (SDG 13), the study shows how this descriptive quantitative evidence can inform—rather than directly identify—an instrument-level policy discussion. The findings are consistent with a Vision 2030 mix that prioritises energy efficiency and green-finance reform, with implications for SDG Targets 7.3, 8.10, 12.2, and 13.2 across oil-exporting economies.

1. Introduction

Achieving environmental sustainability in hydrocarbon-dependent economies poses a distinctive policy challenge. These economies must reconcile continued fiscal reliance on fossil-fuel exports with international climate commitments, growing demands for financial deepening, and the structural diversification required to outlast the carbon transition. Saudi Arabia exemplifies this challenge: the Kingdom is implementing the Vision 2030 diversification agenda, scaling the Saudi Energy Efficiency Program (SEEP), expanding financial-inclusion initiatives through the Saudi Central Bank, and meeting the climate-reporting commitments articulated in its First Biennial Transparency Report to the United Nations Framework Convention on Climate Change [1]. Whether these reforms collectively support or impede the country’s emissions trajectory remains an open empirical question with implications for the wider community of oil-exporting and emerging economies [2,3]. Translating the Sustainable Development Goals (SDGs) from ambition into measurable national outcomes requires not only new technologies but also rigorous, data-driven evidence on how the underlying economic drivers interact—where they reinforce one another and where they trade off. This study contributes to that science-and-evidence pathway for the SDGs by quantifying, with a robust econometric design, the interactions among financial inclusion (SDG 8), energy efficiency (SDG 7), trade structure (SDG 12), and per capita carbon emissions (SDG 13) in a major oil-exporting economy, and by mapping the resulting evidence onto concrete implementation instruments.

The most recent country-specific evidence reports that financial inclusion raises CO₂ emissions in Saudi Arabia, attributing this to credit-fuelled consumption and the limited channelling of financial resources toward renewable projects [4]. Earlier evidence from a panel of six oil-exporting countries found that the finance–environment relationship is conditional on information and communication technology (ICT) infrastructure [5]. Belloumi and Alshehry [6] document that trade openness exerts a long-run negative effect on environmental quality in Saudi Arabia using ARDL cointegration over 1971–2016, but the mechanism, persistence, and policy-regime dependence of this effect have not been fully examined. Earlier work by some of the present authors documented analogous regime sensitivity in the urbanisation–energy–CO₂ nexus for Saudi Arabia over 1968–2022 [7] and in the energy-consumption–growth–CO₂ nexus over 1985–2021 [8], suggesting that single-coefficient estimates may obscure economically meaningful regime variation.

Against this backdrop, the present study makes five contributions. First, it positions its findings directly against the closest Saudi comparators, explaining the apparent contradiction with [4] on financial inclusion and the broad consistency with [6] on trade. Second, the study uses energy intensity rather than energy consumption per capita as the proxy for energy efficiency, addressing the long-standing concern that the latter conflates scale and efficiency effects [9]. Third, the study uses per capita CO₂ emissions (LNCP) as the primary dependent variable, addressing the mechanical correlation that arises when carbon productivity is regressed on GDP. Fourth, the study identifies a statistically significant structural break in the cointegrating relationship at 2001 (Chow F = 7.36, p < 0.001), with the Zivot–Andrews endogenous-break dates for financial inclusion (2000) and trade openness (2005) clustering around the same window. The 2001 break falls within the 2000–2005 WTO accession negotiation period; 2005 (the year of formal accession) is used as a policy robustness check rather than as a separate statistical break. Fifth, the study grounds its policy discussion in Saudi institutions and instruments—SEEP, Public Investment Fund green-finance initiatives, and the National Renewable Energy Program—drawing out implications for SDG Targets 7.3, 8.10, 12.2, and 13.2 [10,11]. In doing so, the analysis addresses the trade-offs and synergies among these goals: expanding financial access (SDG 8) does not by itself reduce emissions and must be coupled with energy-efficiency (SDG 7) and trade-reorientation (SDG 12) measures to deliver climate gains (SDG 13).

The remainder of the paper is organised as follows. Section 2 reviews the theoretical and empirical literature. Section 3 describes the data, presents descriptive statistics, and details the econometric methodology. Section 4 presents the results, including the structural-break analysis and robustness checks. Section 5 discusses the findings and policy implications. Section 6 concludes.

2. Literature Review

2.1. Theoretical Framework

The conceptual basis of this study draws on three complementary frameworks. The Environmental Kuznets Curve (EKC) posits an inverted-U relationship between income and environmental degradation. The STIRPAT framework decomposes environmental pressure into population, affluence, and technology [12]. The scale–composition–technique decomposition explains how trade openness can simultaneously raise emissions through expanded activity and lower them through the adoption of cleaner technology [13]. Within these frameworks, financial inclusion operates through two opposing channels: broader credit-financed consumption raises emissions, while green lending and the diffusion of efficient technologies lower them [14,15]. Which channel dominates is an empirical question that depends critically on the institutional regime governing credit allocation.

2.2. Empirical Literature

2.2.1. Financial Inclusion and Emissions

The empirical evidence on the financial-inclusion–environment nexus is genuinely mixed. Several panel studies report that financial inclusion increases CO₂ emissions [14,15,16]. For six oil-exporting countries, including Saudi Arabia, Damrah et al. [5] find that the finance–environment relationship is conditional on ICT infrastructure. Alshammari [4] shows that, for Saudi Arabia, financial inclusion is associated with higher CO₂ emissions, while FDI inflows mitigate them. Parveen et al. [17] emphasise that when financial inclusion is coupled with green finance and efficiency-enhancing technologies, it can support decarbonisation. The mixed character of this literature implies that the sign of the finance–environment effect should not be assumed in advance and may depend on the institutional regime governing financial access—a point central to the present analysis.

2.2.2. Trade Openness and Emissions

Cross-country studies of emerging economies report that trade openness can reduce CO₂ emissions through technology transfer [18]. In contrast, studies of Belt and Road economies find that trade can increase emissions when scale and dirty export effects dominate [19]. The most directly relevant Saudi-specific evidence comes from Belloumi and Alshehry [6], who found that trade openness has a long-run negative effect on environmental quality in Saudi Arabia, reflecting the dominance of hydrocarbon exports. For other oil-exporters, similar long-run patterns have been documented [20,21].

2.2.3. Energy Intensity, Urbanisation, and Emissions

Energy efficiency is consistently identified as a central mitigation lever, although the choice of proxy matters. Panel cointegration studies show that higher energy efficiency reduces emissions in the long run [22]. The rebound effect can partially offset these gains [23]. Patterson [9] argues that energy intensity (energy use per unit of GDP) more directly captures the productivity of energy use than energy consumption per capita, but cautions that a decline in energy intensity may reflect either genuine efficiency improvements, shifts toward less energy-intensive sectors, or higher energy prices reducing demand. The present study uses energy intensity as the principal proxy and acknowledges this composite interpretation. Urbanisation has been linked to higher emissions in rapidly developing settings. For Saudi Arabia, Houaneb et al. [7] document that urbanisation and economic growth jointly elevate energy consumption and emissions over 1968–2022, while Khan and Khan [8] confirm a stable long-run cointegration among energy consumption, economic growth, and CO₂ emissions in Saudi Arabia over 1985–2021 using ARDL.

2.2.4. Structural Breaks in Environment–Energy Regressions

The methodological literature emphasises that environment–energy–finance regressions estimated over long horizons can be biassed if structural breaks are ignored. Zivot and Andrews [24] propose a unit-root test with a single endogenous break; Bai and Perron [25,26] develop a framework for multiple endogenous breaks; and Pesaran et al. [27] note that ARDL bounds tests retain asymptotic validity only under stable cointegrating relationships. In the Gulf and oil-exporter context, the structural-break literature remains comparatively thin, even though the region experienced several macroeconomic and policy events—the 1986 oil-price collapse, WTO accessions in the early 2000s, the 2008–2009 financial crisis, the 2014–2016 oil-price decline, and the 2016 launch of Vision 2030—that plausibly altered the elasticities of environmental outcomes [28]. The present study contributes by systematically applying a battery of break tests to the cointegrating regression and by re-estimating the long-run model across the resulting regimes.

2.3. Research Gap and Contribution

Three substantive gaps remain in the Saudi-specific literature. First, the financial-inclusion result for Saudi Arabia is contested: Alshammari [4] reports a positive finance–emissions effect, while broader oil-exporter evidence finds a conditional effect [5]. The source of this disagreement has not been examined through formal structural-break analysis. Second, the most common dependent variable in this literature is carbon productivity (GDP/CO₂). However, when this is regressed on GDP as a control, a mechanical correlation arises because GDP appears on both sides of the equation. Few studies in this literature have explicitly addressed this concern. Third, policy discussions tend to be generic rather than tied to Saudi institutions and to specific quantitative targets. This paper addresses each of these three gaps.

3. Materials and Methods

3.1. Data and Variable Definitions

The study uses annual time-series data for Saudi Arabia from 1980 to 2020 (41 observations), drawn from the World Development Indicators (WDI) of the World Bank and the Financial Access Survey (FAS) of the International Monetary Fund. The sample begins in 1980 because consistent series for the financial-inclusion index, energy use, and trade flows for Saudi Arabia are first reliably available from that year. The sample ends in 2020 for two reasons. First, the FAS Financial Institutions Access Index for Saudi Arabia is consistent through 2020 only; updates beyond that year reflect revised compilation methods that are not fully comparable across the full sample window. Because Alshammari [4]—the closest comparator—uses an end period that is effectively similar, retaining 2020 ensures direct comparability with the most relevant published Saudi evidence. Second, ending in 2020 means that the analysis is not contaminated by the unique combination of the COVID-19 pandemic, the 2022 oil-price surge, and the post-pandemic recovery, all of which would constitute additional structural disturbances over a very short window. Future work should extend the analysis once the FAS series is updated. The 41-year span covers the 1986 oil-price collapse, WTO accession in December 2005, the 2008–2009 financial crisis, the 2014–2016 oil-price decline, and the launch of Vision 2030 in April 2016. A detailed data-construction and replication map linking each regression variable to its original source and transformation is provided in Appendix A.

All variables are expressed in natural logarithms. The primary specification uses CO₂ emissions per capita (LNCP) as the dependent variable. This choice addresses a mechanical-correlation concern: a common alternative dependent variable, carbon productivity LNES = ln(GDP/CO₂), shares the GDP numerator with the LNG regressor and with the LNEE = ln(energy/GDP) denominator. Estimating LNES on LNG and LNEE therefore induces a built-in correlation that has nothing to do with the underlying environment–growth or environment–efficiency relationship. To address this, LNCP is used as the primary dependent variable in Section 4.3 onwards, with LNG dropped from the primary specification (its effect can be inferred from the EKC literature). The estimating equations map onto the STIRPAT framework [12] as follows: expressing the outcome in per capita terms embeds the population (P) component, energy intensity (LNEE) proxies the technology (T) component, and the affluence (A)/EKC channel is carried by GDP, which is retained in the LNES robustness specification (Equation (2)) and re-introduced as an explicit scale control in Section 4.5.3; financial inclusion, trade openness, and urbanisation enter as institutional and structural conditioning variables. We acknowledge that, with a single short national series, this mapping is approximate rather than a structural test of STIRPAT or the EKC. Carbon productivity (LNES) is retained as a robustness specification, and full LNES results are reported in Section 4.6. Variable definitions are reported in

Table 1. Variable definitions and data sources.

Variable	Symbol	Definition	Source	Expected Sign
CO2 per capita	LNCP	ln of total CO2 emissions/population (primary dependent variable)	WDI	Dependent
Carbon productivity	LNES	ln of real GDP/CO2 emissions (robustness)	WDI	Alt. dependent
Financial inclusion	LNFI	ln of Financial Institutions Access Index	IMF FAS	±(regime-dependent)
Trade openness	LNTO	ln of (exports + imports)/GDP	WDI	±
Energy intensity	LNEE	ln of energy use per unit of real GDP (lower = higher efficiency)	WDI	+on LNCP (higher intensity → higher emissions)
Urbanisation	LNU	ln of urban population share (% of total)	WDI	±
Economic growth (LNES spec. only)	LNG	ln of real GDP (constant USD); not used in primary LNCP specification	WDI	+/−(EKC; LNES only)

Note: All variables are in natural logarithms. LNCP is the primary dependent variable, which avoids mechanical correlation between GDP-based variables. LNG is dropped from the primary specification and retained only in the LNES robustness specification (Section 4.6).

.

3.2. Descriptive Statistics, Correlations, and Multicollinearity

Table 2. Descriptive statistics and Jarque–Bera normality tests (1980–2020; n = 41).

Variable	Mean	Std. Dev.	Min	Max	Skewness	Kurtosis	JB p-Value
LNCP	−10.914	0.069	−11.031	−10.795	0.210	−1.148	0.278
LNES	21.058	0.255	20.829	22.001	2.579	6.787	<0.001
LNFI	−1.302	0.435	−2.283	−0.578	−0.304	−0.493	0.562
LNTO	4.272	0.169	3.861	4.565	−0.088	−0.589	0.671
LNEE	−10.674	2.689	−14.103	−5.794	0.622	−1.087	0.104
LNU	16.330	0.567	15.198	17.096	−0.367	−0.995	0.268
LNG	26.717	0.367	26.185	27.342	0.362	−1.147	0.211

Note: JB = Jarque–Bera test of normality. All variables are in natural logarithms. The dependent variable, LNCP, exhibits near-normal behaviour, while LNES is highly leptokurtic and right-skewed.

reports descriptive statistics and the correlation matrix;

Table 3. Pearson correlation matrix and variance inflation factors.

	LNCP	LNES	LNFI	LNTO	LNEE	LNU	LNG
LNCP	1.000	−0.305	0.383	−0.218	0.479	0.254	0.333
LNES	−0.305	1.000	−0.698	0.396	−0.611	−0.791	−0.429
LNFI	0.383	−0.698	1.000	−0.146	0.819	0.795	0.687
LNTO	−0.218	0.396	−0.146	1.000	−0.191	−0.244	−0.081
LNEE	0.479	−0.611	0.819	−0.191	1.000	0.925	0.957
LNU	0.254	−0.791	0.795	−0.244	0.925	1.000	0.879
LNG	0.333	−0.429	0.687	−0.081	0.957	0.879	1.000
VIF (LNES spec.)	—	—	5.30	1.38	38.28	7.58	23.09
VIF (LNCP spec.)	—	—	3.41	1.07	8.18	7.86	—

Note: VIFs above 10 (bold) indicate severe multicollinearity and motivate dropping LNG from the LNCP specification. In the LNCP specification, all VIFs are below 10 and therefore conventionally acceptable; however, LNEE (8.18) and LNU (7.86) exceed 5, indicating moderate multicollinearity. The coefficients on these variables are stable across estimators (ARDL, FMOLS, DOLS, CCR; see Section 4.3 and Section 4.4) and across structural-break sub-samples (Section 4.5.1), suggesting that the moderate VIFs do not materially bias inference.

reports the variance inflation factors (VIFs) for the LNCP specification. Three observations are noteworthy. First, the dependent variable LNCP exhibits much milder skewness (0.21) and kurtosis (−1.15) than LNES (skewness 2.58, kurtosis 6.79); the Jarque–Bera test does not reject normality for LNCP (p = 0.28) but does reject for LNES (p < 0.001). This reinforces the decision to use LNCP as the primary dependent variable. Second, the correlation matrix shows that LNFI, LNEE, LNU, and LNG are all highly positively correlated (pairwise correlations ranging from 0.69 to 0.96), reflecting the joint upward trend of these variables in the modernising Saudi economy. Third, the VIFs confirm severe multicollinearity in any specification that includes LNG alongside LNEE and LNU (VIFs for LNG = 23.1 and for LNEE = 38.3). This pattern justifies excluding LNG from the primary LNCP specification, leaving LNFI, LNTO, LNEE, and LNU as the four regressors. For the four-regressor LNCP specification, VIFs fall to acceptable levels (all below 10). Figure 1 presents the time-series plots of all six variables, with the 2001 statistical break (Section 4.5) and 2005 WTO accession marked as vertical lines.

3.3. Econometric Specification

The primary specification is as follows:

LNCP_t = β₀ + β₁ LNFI_t + β₂ LNTO_t + β₃ LNEE_t + β₄ LNU_t + ε_t

(1)

The robustness (LNES) specification adds LNG as a control and uses carbon productivity as the dependent variable:

LNES_t = α₀ + α₁ LNFI_t + α₂ LNTO_t + α₃ LNEE_t + α₄ LNU_t + α₅ LNG_t + η_t

(2)

The unrestricted error-correction representation of the ARDL model used for bounds testing, illustrated for the LNCP specification, is:

ΔLNCP_t = γ₀ + Σ θ_1i ΔLNCP_t−i + Σ θ_2i ΔLNFI_t−i + Σ θ_3i ΔLNTO_t−i + Σ θ_4i ΔLNEE_t−i + Σ θ_5i ΔLNU_t−i + δ₁ LNCP_t−1 + δ₂ LNFI_t−1 + δ₃ LNTO_t−1 + δ₄ LNEE_t−1 + δ₅ LNU_t−1 + υ_t

(3)

where Δ is the first-difference operator, and the δ coefficients on the lagged level terms jointly characterise the long-run (cointegrating) relationship; the ARDL bounds F-test is a joint test of H₀: δ₁ = δ₂ = … = δ₅ = 0. Equation (3) is the unrestricted conditional error-correction form and therefore does not contain a separate error-correction term: the single-term restricted representation, ΔLNCP_t = γ₀ + Σ … + ϕ ECT_t−1 + υ_t with ECT_t−1 defined as the lagged residual from the long-run regression (1) and ϕ the speed of adjustment, is an algebraically equivalent reparameterisation in which the lagged levels are collapsed into ECT_t−1. The two forms are alternatives and are not estimated jointly; we report ϕ from the restricted form in Section 4.3. For ARDL bounds testing, the dependent variable must be I(1), and regressors may be a mix of I(0) and I(1) [27].

3.4. Estimation Strategy

The empirical procedure proceeds in eight stages: (i) unit-root tests using ADF and Phillips–Perron, supplemented by Zivot–Andrews tests with one endogenous break [24] for every variable; (ii) the ARDL bounds test with critical values from Narayan [29]; (iii) long- and short-run ARDL estimation with AIC-based lag selection; (iv) FMOLS, DOLS, and CCR robustness for the long-run coefficients; (v) Granger causality on first differences with VAR lag selection by AIC/BIC/HQ, supplemented by Toda–Yamamoto tests [30] in levels for variables with mixed integration orders; (vi) standard diagnostic tests (Breusch–Godfrey, Breusch–Pagan–Godfrey, Jarque–Bera, Ramsey RESET) and CUSUM/CUSUMSQ stability tests; (vii) a structural-break battery comprising the Bai–Perron [25,26] multiple-break sup-F procedure (with Andrews [31] critical values) and the Chow [32] test at policy-motivated dates; and (viii) sub-sample re-estimation around the dominant break date. All computations were carried out in Python 3.6; the reproducible analysis code is provided in the Supplementary Materials (revision_analysis.py).

3.5. Hypotheses

H1. 

The effect of financial inclusion on per capita CO₂ emissions in Saudi Arabia is regime-dependent and may differ across statistically detected structural-break sub-samples.

H2. 

Trade openness has a positive long-run effect on per capita CO₂ emissions, reflecting hydrocarbon-intensive export composition.

H3. 

Lower energy intensity reduces per capita CO₂ emissions.

H4. 

Urbanisation has a significant effect on per capita CO₂ emissions, with the sign depending on infrastructure and consumption channels.

4. Results

4.1. Unit Root Tests

Table 4. Unit root tests: ADF, Phillips–Perron, and Zivot–Andrews (one endogenous break, Model C).

Variable	ADF (Level)	ADF (Δ)	PP (Level)	PP (Δ)	Z–A t-Stat (Level)	Z–A Break Year	Order
LNCP	−2.163	−3.479 ***	−2.184	−3.523 ***	−4.190	2010	I(1)
LNES	−1.791	−5.066 ***	−1.853	−5.183 ***	−7.580 ***	2015	I(0) w/break
LNFI	−1.667	−4.523 ***	−1.781	−4.498 ***	−4.567	2000	I(1)
LNTO	−1.627	−4.184 ***	−1.210	−4.252 ***	−3.168	2005	I(1)
LNEE	−1.283	−3.264 **	−0.210	−3.880 ***	−3.238	1995	I(1)
LNU	−1.787	−3.142 **	−1.622	−3.087 **	−3.577	2015	I(1)
LNG	−0.067	−3.639 ***	0.262	−3.660 ***	−7.233 ***	2011	I(0) w/break

Note: ** and *** denote significance at the 5% and 10% levels. Zivot–Andrews critical values (Model C, Z–A 1992): 1% = −5.57; 5% = −5.08; 10% = −4.82. The primary dependent variable, LNCP, is I(1) under all three tests, satisfying the ARDL bounds-test precondition.

reports the ADF and Phillips–Perron unit-root tests under the intercept-only specification appropriate to the bounds test [27], together with Zivot–Andrews [24] tests with one endogenous break (Model C). Under the standard ADF and PP tests, all variables are non-stationary at the level and become stationary at the first difference, confirming an I(1) order and satisfying the precondition of the ARDL bounds test. The Zivot–Andrews tests provide additional information: LNES and LNG reject the unit-root null at the 1% level when an endogenous break is allowed (with detected breaks at 2015 and 2011, respectively), while LNCP, LNFI, LNTO, LNEE, and LNU do not reject and are therefore confirmed as I(1) even under break-augmented testing. The detected break dates for LNFI (2000) and LNTO (2005) cluster around the 2001–2005 window, which emerges as the structural-break date in Section 4.5, providing convergent evidence that the most important regime change in the Saudi finance–trade–emissions system occurred in the early 2000s. Because the primary dependent variable, LNCP, is I(1) under all three tests, the ARDL bounds testing approach remains valid for the primary specification.

4.2. ARDL Bounds Test

Table 5. ARDL bounds test for cointegration.

Specification	F-Statistic	t-Statistic (ECT)	5% Bounds (k as Shown)
LNCP (primary, k = 4)	4.050	−4.820	F: 2.86/3.49; t: −2.86/−3.99
LNES (robustness, k = 5)	4.500	−5.600	F: 2.62/3.79; t: −2.86/−4.19

Note: Critical values from Narayan [29] (Case II, restricted intercept and no trend). Both specifications confirm cointegration at the 5% level. The F-statistic and t-statistic are independent tests; rejection of both reinforces the cointegration conclusion.

reports the ARDL bounds F-statistic and the t-statistic on the lagged dependent variable for both the primary LNCP and the robustness LNES specifications. For the LNCP specification, the F-statistic of 4.05 (with k = 4 regressors) exceeds the upper-bound 5% critical value of 3.49 from Narayan [29], confirming cointegration. The corresponding ECT t-statistic of −4.82 also exceeds its 5% critical value of −3.99. For the LNES specification (k = 5), the F-statistic of 4.50 exceeds the 5% upper bound of 3.79, also confirming cointegration. The bounds test thus provides robust evidence of a long-run equilibrium relationship in both specifications.

4.3. ARDL Long-Run and Short-Run Estimates (Primary: LNCP)

Table 6. ARDL long-run and short-run estimates (primary specification: dependent variable LNCP).

Variable	Coefficient	Std. Error	p-Value
Long run
LNFI	0.013	0.030	0.665
LNTO	0.022	0.024	0.358
LNEE	0.092 ***	0.010	0.000
LNU	−0.412 ***	0.108	0.001
Short run
ΔLNFI	0.008	0.024	0.731
ΔLNTO	−0.094 *	0.049	0.064
ΔLNTO (−1)	0.123 ***	0.038	0.003
ΔLNEE	0.103 ***	0.014	0.000
ΔLNU	−0.831 ***	0.241	0.002
ECT (−1)	−0.610 ***	0.116	0.000
R2	0.872	—	—
Adjusted R2	0.823	—	—
F-statistic	17.8 ***	—	0.000
Durbin–Watson	2.04	—	—

Note: * and *** denote significance at the 1% and 10% levels. Standard errors are heteroscedasticity-robust (HC3). A positive coefficient on LNEE (energy intensity) means higher energy intensity raises emissions per capita; equivalently, lower energy intensity (greater efficiency) reduces them. ECT = error-correction term. Lag order selected by AIC.

reports the ARDL long- and short-run coefficients for the primary LNCP specification. The error-correction term is negative and highly significant at −0.61 (p < 0.001), indicating an annual adjustment speed of 61% toward the long-run equilibrium. In the long run, higher energy intensity (positive coefficient on LNEE = 0.092) is significantly associated with higher CO₂ emissions per capita at the 1% level, confirming H3 (equivalently, lower intensity reduces emissions). Urbanisation has a strong negative long-run effect on emissions per capita (−0.41, p < 0.001), consistent with the densification benefits that more concentrated urban living can deliver in modernising economies [7]. The long-run trade openness coefficient is small and not significant in the full sample. Financial inclusion is also statistically insignificant in the full-sample long run—a result that motivates the structural-break analysis in Section 4.5, where the full-sample average is shown to mask regime variation.

4.4. FMOLS, DOLS, and CCR Robustness

Table 7. Robustness check using FMOLS, DOLS, and CCR (primary specification, dep. variable: LNCP).

Variable	FMOLS Coef.	FMOLS P	DOLS Coef.	DOLS P	CCR Coef.	CCR P
LNFI	0.018	0.490	0.012	0.701	0.020	0.451
LNTO	0.027	0.241	0.031	0.213	0.028	0.228
LNEE	0.093 ***	0.000	0.078 ***	0.002	0.094 ***	0.000
LNU	−0.418 ***	0.000	−0.402 ***	0.000	−0.420 ***	0.000
Constant	−4.110	0.001	−3.890	0.002	−4.130	0.001

Note: *** denote significance at the 10% level.

reports the long-run coefficients from FMOLS, DOLS, and CCR for the primary LNCP specification. The energy-intensity coefficient is positive and highly significant across all three estimators, confirming H3. The urbanisation coefficient is negative and highly significant across all three. The financial-inclusion and trade-openness coefficients are not significantly different from zero in the full-sample long run—a result that, again, will be shown by the structural-break analysis in Section 4.5 to mask regime variation.

4.5. Structural Break Analysis

Two complementary structural-break tests are applied to the cointegrating regression from Equation (1). First, the Bai–Perron sup-F test [25,26] computes the maximum Chow F-statistic over the trimmed candidate-break range, compared with the Andrews [31] critical values. Second, the Chow test [32] is computed at six policy-motivated candidate dates. The Zivot–Andrews [24] endogenous-break tests reported in Table 4 (on each variable individually) provide auxiliary evidence on the timing of breaks in the input series, with detected break dates for LNFI (2000) and LNTO (2005) clustering around the 2001–2005 window.

Table 8. Structural break test battery (primary specification, LNCP).

Test	Statistic	Break Year	Conclusion
Bai–Perron sup-F	F = 26.37	1990	Significantly above 1% Andrews CV (≈20.8); confirms a major break in the early 1990s
Chow in 1986 (oil-price collapse)	F = 17.29 ***	1986	Significant at 1%
Chow in 1995	F = 19.92 ***	1995	Significant at 1%
Chow at 2001 (close to Z–A breaks for LNFI/LNTO)	F = 7.36 ***	2001	Significant at 1%; primary break date for sub-sample re-estimation
Chow in 2005 (WTO accession)	F = 3.04 **	2005	Significant at 5% (policy robustness check)
Chow in 2010	F = 2.85 **	2010	Significant at 5%; coincides with CUSUMSQ excursion (Section 4.7)

Note: ** and *** denote significance at the 5%, and 10% levels. Andrews [31] sup-F 5% critical value for k = 5, trim = 0.15 is approximately 16.06; 1% is approximately 20.83. The Bai–Perron PELT algorithm with the conservative BIC penalty detects zero breaks (consistent with small-sample power limitations); with a moderate penalty, it detects three breaks (1985, 1989, 1995); these results are reported in the Supplementary Materials. An exploratory Zivot–Andrews test applied directly to the OLS residuals from Equation (1) yields t = −4.19 at a candidate break year of 2010 (marginally significant), but the standard Z–A critical values are derived for raw series rather than for OLS residuals; this result is therefore reported as exploratory only and is not used as a primary break-test in the present analysis.

summarises the break-test results.

Five findings emerge from Table 8. The corresponding cointegrating residuals are plotted in Figure 2. First, the Chow test at 2001 is significant at the 1% level (F = 7.36, p < 0.001), and this is the candidate break that combines statistical significance with substantive interpretability: 2001 falls in the middle of the WTO-accession negotiation period (2000–2005), is the year Saudi Arabia signed its Bilateral Investment Agreement with the United States [33], and is close to the Zivot–Andrews endogenous break dates for LNFI (2000) and LNTO (2005) reported in Table 4. We therefore use 2001 as the primary break date for sub-sample re-estimation. Second, the Bai–Perron sup-F statistic of 26.37 is achieved at 1990 (significant at the 1% Andrews critical value), indicating that an additional major break occurred during the post-Gulf-War period; we discuss why this is not used as the primary break point below. Third, Chow tests at multiple candidate dates (1986, 1995, 2001, 2005, 2010) are all individually significant, reflecting the genuine structural instability of the Saudi economy over the four decades studied. Fourth, the 2005 Chow test (significant at 5%) and the 2010 Chow test (significant at 5%) are reported as robustness checks. Fifth, the Bai–Perron PELT algorithm, with a conservative BIC penalty, detects zero breaks (consistent with small-sample power limitations), whereas with a moderate penalty it detects three early-period breaks; details are in the Supplementary Materials.

Why is 1990 not used as the primary break point? Although the Bai–Perron sup-F statistic is maximised at 1990 (Figure 3), this break is not used for sub-sample re-estimation for two reasons. First, the 1990 break coincides with the immediate aftermath of the Gulf War (1990–1991), which represents a transitory shock to oil markets and Saudi public finances rather than a persistent structural change; treating it as a breakpoint would conflate a short-lived disruption with the long-run cointegrating relationship of interest. Second, the 1990 break leaves only 10 observations in the pre-break sub-sample (1980–1989) and 31 in the post-break sub-sample, which is highly unbalanced and would not permit reliable sub-sample comparison. The 2001 break, in contrast, yields balanced sub-samples of 21 and 20 observations and is supported by independent evidence from the Z–A tests on LNFI and LNTO. We therefore use 2001 as the primary break for sub-sample re-estimation while reporting the Bai–Perron sup-F at 1990 as evidence of additional early-sample instability.

4.5.1. Sub-Sample Re-Estimation at 2001 (Primary) and 2005 (Policy Robustness)

Table 9. Sub-sample re-estimation around the 2001 (primary) and 2005 (policy robustness) break dates.

Variable	LNCP Pre-2001 (n = 21)	LNCP Post-2001 (n = 20)	LNCP Pre-WTO 2005 (n = 25)	LNCP Post-WTO 2005 (n = 16)
LNFI	0.088 (p = 0.249)	0.013 (p = 0.665)	0.047 (p = 0.474)	0.033 *** (p = 0.007)
LNTO	0.050 (p = 0.893)	0.036 (p = 0.018) **	−0.197 (p = 0.188)	0.054 (p = 0.004) *
LNEE	0.224 (p = 0.313)	0.108 (p < 0.001) *	0.021 (p = 0.783)	0.111 (p < 0.001) *
LNU	−0.415 (p = 0.159)	−0.680 (p < 0.001) *	−0.148 (p = 0.427)	−0.679 (p < 0.001) *
Constant	−1.529 (p = 0.814)	1.302 (p = 0.470)	−7.398 (p = 0.088)	1.234 (p = 0.251)
R2	0.389	0.985	0.292	0.996
Adjusted R2	0.236	0.981	0.150	0.994

Note: *, **, and *** denote significance at the 1%, 5%, and 10% levels. Standard errors are heteroscedasticity-robust (HC3). The pre-break sub-samples have low R² and no significant coefficients, consistent with the structural instability identified in Section 4.5. The post-break sub-samples show much tighter relationships, supporting the regime-shift interpretation. The implications of the poor pre-2001 fit are discussed in Section 5.1.

reports the long-run sub-sample re-estimates around the primary 2001 break and the 2005 WTO-accession robustness check. The pre-2001 LNCP regression has low explanatory power (R² = 0.39) and no individually significant coefficients, consistent with the high pre-break instability visible in Figure 2b and the multiple significant Chow tests in Table 8. The post-2001 LNCP regression has very high explanatory power (R² = 0.985) and tightly estimated coefficients: energy intensity (LNEE) is positive and significant at the 1% level (β = 0.108, p < 0.001), urbanisation (LNU) is negative and significant at the 1% level (β = −0.680, p < 0.001), and trade openness (LNTO) is positive and significant at the 5% level (β = 0.036, p = 0.018). Financial inclusion (LNFI) is statistically insignificant in both regimes once the LNCP specification is used. The qualitative pattern is similar at the 2005 robustness break, with slightly stronger LNTO and LNFI effects post-2005, reflecting the somewhat smaller sub-sample (n = 16).

4.5.2. Robustness for the Small Post-Break Sub-Sample

The post-2001 sub-sample contains n = 20 observations against 5 estimated parameters (constant + 4 regressors), yielding 15 residual degrees of freedom and 4 observations per parameter, below the conventional 10:1 rule of thumb [34]. Rolling-window OLS coefficient paths for the primary specification are plotted in Figure 4. Highly significant p-values in such samples must therefore be interpreted with caution. To address this concern, ridge regression with the regularisation parameter selected by leave-one-out cross-validation was conducted. The optimal regularisation parameter was α ≤ 0.01, and the ridge coefficients were within 5% of the OLS estimates, indicating that the OLS solution was not driven by overfitting. Detailed ridge results are provided in the Supplementary Materials (Table S1 and table_ridge.csv). The small sample size remains a binding limitation that will be relieved only when the FAS Financial Inclusion Index is extended beyond 2020 (Section 6.2).

4.5.3. Robustness to a GDP Scale Control, Bootstrap Inference, and Break-Date Sensitivity

In response to three concerns—(i) the omission of GDP from the primary specification, (ii) the reliability of inference in a 20-observation sub-sample, and (iii) the discretionary choice of the 2001 break date—this subsection reports three additional robustness exercises on the post-2001 LNCP model. The exercises are supplementary; the primary results remain those in Table 6 and Table 9.

First, we re-estimate the post-2001 model with the natural logarithm of real GDP (LNG) added as an explicit scale control, addressing the concern that omitting GDP could bias the remaining coefficients. As anticipated from the descriptive diagnostics (Table 3), including LNG reintroduces severe multicollinearity: the variance inflation factors rise to 42.8 for LNEE, 85.1 for LNU, and 70.3 for LNG. We therefore estimate the augmented model both by OLS and by ridge regression with the penalty selected by leave-one-out cross-validation (

Table 10. GDP-augmented post-2001 LNCP specification (n = 20). OLS standard errors are heteroscedasticity-robust (HC3); ridge coefficients use leave-one-out cross-validation (optimal penalty α = 0.001 on standardised data). R² = 0.985. ** and *** denote significance at the 5% and 1% levels. The insignificant LNG coefficient and the stability of the other coefficients indicate that omitting GDP does not materially bias the estimates.

Variable	OLS Coef.	OLS p-Value	Ridge Coef.
LNFI	0.013	0.695	0.012
LNTO	0.036	0.025 **	0.036
LNEE	0.108	<0.001 ***	0.108
LNU	−0.692	<0.001 ***	−0.688
LNG	0.015	0.859	0.015
Constant	1.083	0.711	—

). Two results follow. The LNG coefficient is small and statistically insignificant (OLS β = 0.015, p = 0.859), which is the expected pattern once the dependent variable is already expressed in per capita terms. More importantly, adding LNG leaves the other coefficients essentially unchanged relative to the baseline in Table 9—LNEE (0.108), LNU (−0.692), and LNTO (0.036) move by less than two percent—and the ridge estimates coincide with the OLS estimates. This indicates that the omission of GDP from the primary specification does not materially bias the reported elasticities; the scale information that GDP would add is already absorbed by the per capita transformation and by the energy-intensity and urbanisation regressors. We retain the parsimonious four-regressor LNCP model as primary, but the augmented results show that the substantive conclusions are not an artefact of excluding GDP.

Second, to provide inference that does not rely on small-sample normal approximations, we compute residual-bootstrap 95% confidence intervals for the post-2001 long-run coefficients (5000 replications; see Figure S1 in the Supplementary Materials). The intervals confirm the OLS inference: LNEE [0.095, 0.122], LNU [−0.834, −0.524], and LNTO [0.008, 0.063] all exclude zero, whereas LNFI [−0.026, 0.053] is tightly centred on zero and includes it. The financial-inclusion null result, therefore, reflects a genuinely near-zero, well-bounded estimate rather than merely low power, while the energy-intensity, urbanisation, and trade-openness effects survive resampling.

Third, because the Bai–Perron sup-F statistic is maximised at 1990 while we use 2001 as the primary break, we report a break-date sensitivity analysis (

Table 11. Break-date sensitivity: post-break long-run LNCP regression (LNFI, LNTO, LNEE, LNU) re-estimated at five candidate start years. Standard errors are heteroscedasticity-robust (HC3). ** and *** denote significance at the 5% and 1% levels. Energy intensity and urbanisation are significant at the 1% level for every candidate break; the substantive conclusions do not depend on the discretionary choice of 2001.

Post-Break Start	n	R2	LNFI	LNTO	LNEE	LNU
1990	31	0.929	−0.034	0.017	0.090 ***	−0.471 ***
1995	26	0.979	0.013	0.026	0.104 ***	−0.645 ***
2001 (primary)	20	0.985	0.013	0.036 **	0.108 ***	−0.680 ***
2005 (WTO)	16	0.996	0.033 ***	0.054 ***	0.111 ***	−0.679 ***
2010	11	0.993	0.005	0.004	0.115 ***	−0.774 ***

; visualised in Figure S2 in the Supplementary Materials) that re-estimates the post-break LNCP model at five candidate start years (1990, 1995, 2001, 2005, 2010). The qualitative conclusions are invariant to this choice: energy intensity is positive and significant at the 1% level at every break date (range 0.090–0.115), and urbanisation is negative and significant at the 1% level at every break date (range −0.471 to −0.774). Trade openness becomes statistically significant only from 2001 onward, consistent with the WTO-accession narrative, and financial inclusion is insignificant at every break except a marginal effect at 2005. The headline findings are thus robust to the break-date decision; what changes across breaks is confined to the trade-openness and financial-inclusion margins, exactly the coefficients we already report as the least secure.

4.6. Robustness: LNES Specification and Comparison with [4]

As a robustness check, Equation (2) replaces the dependent variable with LNES (carbon productivity) and adds LNG as a control. For comparability with Alshammari [4], the same five-regressor specification is estimated. Although the high multicollinearity (VIF for LNG = 23, LNEE = 38; Table 3) inflates standard errors and biases interpretation, the qualitative pattern is consistent with the LNCP findings: in the post-2001 regime, LNEE remains highly significantly negative, LNU remains negatively significant, and the LNFI coefficient flips sign across the break threshold. The LNES specification, therefore, confirms the regime-dependence story in directional terms but is methodologically inferior to the LNCP specification, which we view as the cleaner test.

4.7. Diagnostic and Parameter-Stability Tests

Table 12. Diagnostic and stability tests (primary LNCP ARDL specification).

Diagnostic Test	Test Statistic	p-Value	Inference
Breusch–Godfrey LM (lags = 2)	0.310	0.736	No serial correlation
Breusch–Pagan–Godfrey	2.140	0.084	No significant heteroscedasticity
Jarque–Bera	1.280	0.528	Approximately normal
Ramsey RESET (squared fitted)	1.620	0.214	Correct functional form
CUSUM	Figure 5a	—	Within 5% bounds
CUSUMSQ	Figure 5b	—	Brief 2010–2012 excursion (see Section 4.7)

reports the diagnostic battery for the primary LNCP ARDL model. The Breusch–Godfrey LM, Breusch–Pagan–Godfrey, Jarque–Bera, and Ramsey RESET tests are all consistent with the maintained assumptions. The CUSUM statistic (Figure 5a) remains within the 5% significance bounds throughout the effective sample. The CUSUMSQ statistic (Figure 5b) lies within the bounds for most of the sample but briefly approaches the upper bound around 2010–2012. We interpret this excursion as additional evidence of structural instability—consistent with the Chow test at 2010 (F = 2.85, p = 0.031; Table 8) and with the Zivot–Andrews break in LNCP and LNG at 2010–2011 (Table 4). The 2010–2012 excursion coincides with the post-2009 oil-price recovery (Brent crude rose from approximately USD 62 per barrel in 2009 to USD 111 per barrel in 2011), the regional uncertainty following the Arab Spring (December 2010 onwards), and the first major Saudi domestic stimulus package introduced in 2011 in response to these regional events. This period may therefore represent an additional minor structural break that our sub-sample analysis cannot separately identify, given the limited post-2001 observations; we identify formal modelling of this multi-break structure as a direction for future research with longer post-break samples.

4.8. Granger and Toda–Yamamoto Causality

Granger causality requires stationary series and a defensible lag choice. The four VAR lag-selection criteria computed on first-differenced data (1980–2020) give conflicting recommendations: AIC selects lag 4, FPE selects lag 2, and both BIC and HQIC select lag 0. The BIC selection of zero lags reflects the well-known small-sample tendency of BIC to over-penalise additional lags in samples with fewer than 50 observations and this would preclude causality testing entirely. On theoretical grounds appropriate to annual macroeconomic data, where one-year persistence is the minimum economically meaningful frequency, we use lag = 1 as the primary specification; this is also consistent with the FPE’s second-best choice of lag = 2 (rejecting lag = 0). To address the concern that Granger causality on first differences can lose information when variables are cointegrated, we also report the Toda–Yamamoto [30] causality test in levels with k + d_max lags (k = 1 chosen lag, d_max = 1 since all primary variables are I(1)).

Table 13. Granger and Toda–Yamamoto causality tests.

Null Hypothesis	Granger F (Lag 1)	Granger p	Toda–Yamamoto F	T–Y p
LNFI does not cause LNCP	0.188	0.668	2.077	0.159
LNCP does not cause LNFI	0.345	0.561	0.044	0.836
LNEE does not cause LNCP	0.084	0.774	0.923	0.343
LNCP does not cause LNEE	6.338 **	0.017	0.000	0.993
LNTO does not cause LNCP	0.084	0.773	0.800	0.377
LNCP does not cause LNTO	0.031	0.862	0.806	0.376
LNG does not cause LNCP	2.982 *	0.093	0.455	0.505
LNCP does not cause LNG	0.623	0.436	3.841 *	0.058
LNU does not cause LNCP	1.749	0.195	8.166 ***	0.007
LNCP does not cause LNU	0.310	0.581	0.112	0.740

Note: *, **, and *** denote significance at the 1%, 5%, and 10% levels. Granger results use lag 1 in the first-differenced VAR; this lag was chosen on theoretical grounds (annual data, minimum economically meaningful frequency) because BIC selects lag 0 (which would preclude testing), while AIC selects lag 4 and FPE selects lag 2. Toda–Yamamoto results use a level-VAR estimated with k + d_max = 1 + 1 = 2 total lags, of which only the first k = 1 lag of the cause variable is included in the Wald restriction; the additional d_max-th lag is included in the unrestricted regression to ensure the test statistic has a standard χ² distribution under cointegration.

reports both. We emphasise at the outset that the Granger and Toda–Yamamoto procedures test for predictive content (Granger non-causality)—that is, whether past values of one variable help forecast another—and not for structural causation; the results below should be read accordingly. Three findings are noteworthy. First, there is no significant predictive relationship between financial inclusion and per capita emissions in either direction—consistent with the regime-dependent nature of the relationship documented in Section 4.5. Second, urbanisation Granger-predicts per capita emissions in the Toda–Yamamoto specification (F = 8.17, p = 0.007), indicating that the urban–emissions association operates with a temporal lead, consistent with the urban planning literature [7]. Third, the differenced Granger test detects predictive feedback from per capita emissions to energy intensity (F = 6.34, p = 0.017): rising emissions in one year are followed by higher measured energy intensity in the next, plausibly reflecting that emissions-intensive output growth in oil-driven Saudi cycles is recorded in the energy-use numerator of LNEE before the GDP denominator catches up. The Toda–Yamamoto specification, which uses level data and is robust to integration order, does not reject the null for this pair, indicating that the differenced-data finding reflects short-run predictive dynamics rather than a stable long-run lead–lag relationship. These tests, by themselves, do not establish the direction of any structural causal mechanism.

5. Discussion

5.1. Resolving the Saudi Finance–Environment Puzzle

The most important methodological finding is that the apparent regime-dependence of the finance–environment relationship in Saudi Arabia reported in earlier work [4] reflects, in part, the specification choice of using carbon productivity (GDP/CO₂) as the dependent variable. Carbon productivity shares the GDP numerator with two important regressors (GDP itself and the GDP denominator of energy intensity), inducing a mechanical correlation that is hard to interpret. When the specification is reformulated using CO₂ emissions per capita as the dependent variable (eliminating this mechanical correlation), the financial-inclusion coefficient is statistically insignificant in both the pre-2001 (β = 0.088, p = 0.249) and post-2001 (β = 0.013, p = 0.665) sub-samples. The more cautious reading of the data support is therefore narrow: in our preferred LNCP specification, and using the 2001 break, the financial-inclusion coefficient is not statistically distinguishable from zero at conventional levels, and this near-zero estimate survives residual-bootstrap resampling (Section 4.5.3) rather than reflecting low power alone. We do not interpret this as evidence that finance is irrelevant to emissions. The IMF FAS index measures financial access; it does not measure how savings are allocated within the financial system. The relevant margin for emissions may lie in allocation and incentives—how institutions are induced to channel credit toward long-horizon, low-carbon investment—rather than in access per se. Evidence that incentive design and competitive pressure shape the behaviour of long-term financial institutions suggests that an “access” proxy is unlikely to capture this allocation channel, which is a more promising route for future Saudi-specific work than the access measure used here. This reading reconciles the apparent contradiction between [4] and the broader oil-exporter evidence [5]: both report effects that operate through the conditioning structure (specifically the inclusion of GDP and energy intensity together with a GDP-based dependent variable). Our finding is consistent with the more careful conclusion in [5] that the finance–environment effect is conditional on institutional infrastructure rather than direct.

A caveat applies to the pre-2001 sub-sample. The pre-2001 LNCP regression has R² = 0.39 and no individually significant coefficients (Table 9). This poor fit suggests that the linear cointegrating relationship assumed by the ARDL model may not hold before the 2001 structural break, which is consistent with the Bai–Perron evidence of multiple early-sample breaks (1986, 1995, and the sup-F at 1990; Table 8) and with the high pre-1995 volatility visible in the residual plots (Figure 2b). Inference on the pre-2001 period should therefore be treated as descriptive rather than as evidence of stable causal relationships. The post-2001 estimates, in contrast, exhibit high R² values, tight standard errors, and stable rolling-window coefficients (Figure 4), consistent with the interpretation that the cointegrating relationship became economically meaningful only after the 2001 structural shift. This temporal pattern of “instability→stability” is itself an important descriptive finding: the Saudi macroeconomic system transitioned from a regime of multiple short-lived shocks (oil-price collapses, Gulf War) to a more stable regime in which the long-run cointegrating relationships between financial, trade, and energy variables and per capita emissions could be reliably estimated.

5.2. Energy Intensity, Urbanisation, and Trade Openness as the Active Channels

In the post-2001 LNCP regression, three coefficients are quantitatively and statistically robust: energy intensity, urbanisation, and trade openness. The energy-intensity coefficient of +0.108 implies that a one-per-cent fall in energy intensity (greater efficiency) is associated with approximately a 0.11% fall in per capita CO₂ emissions, holding other factors constant. This is the most robust mitigation lever in the analysis, and it aligns with the broader literature [22]. The urbanisation coefficient of −0.680 implies that as urbanisation rises, per capita emissions fall, consistent with the urban-densification benefit channel emphasised in the recent Saudi-specific evidence [7]. The trade-openness coefficient of +0.036 (significant in the post-2001 LNCP regression) is positive, consistent with the hydrocarbon-export composition channel emphasised by [6]. This interpretation should, however, be advanced with caution. Our measure is gross trade openness (exports + imports)/GDP, which is not a sufficient proxy for the carbon content of trade: a rise in openness can reflect higher hydrocarbon exports, downstream petrochemical exports, greater import penetration, or cleaner capital-goods imports, and these channels carry very different emissions implications. Consumption-based accounting shows that emissions are reallocated across borders in ways that territorial-production measures miss [35], and emissions transfers embodied in trade grew materially over recent decades [36]. We therefore treat the positive openness coefficient as a regime-specific association consistent with—but not direct evidence of—a hydrocarbon-composition mechanism; isolating that mechanism would require trade-decomposition or embodied-carbon data that lie beyond the present single-country series (Section 6.2). Importantly, energy intensity itself is partly a composite of efficiency gains, sectoral shifts toward services, and energy-price responses [9]—a caveat we acknowledge and that motivates the Divisia-type decomposition discussed in Section 6.2.

5.3. Implications for Saudi Vision 2030

Before presenting the policy mix, we stress an important boundary on its status. The quantitative targets cited below (e.g., the 30% energy-intensity reduction, the USD 50 billion green-bond figure, the 50% non-oil export share) are drawn from existing Saudi strategy documents and are not estimated in this paper. Where we combine such a target with an estimated coefficient to produce an illustrative emissions figure, that figure is a conditional extrapolation—it holds only if the estimated post-2001 elasticity is treated as constant un to 2030, if the targeted change is realised, and if no offsetting structural change occurs—and it inherits the small-sample uncertainty quantified in Section 4.5.3 (for the LNEE elasticity, a bootstrap 95% interval of roughly [0.095, 0.122]). These projections are therefore offered to make the direction and rough order of magnitude of the channels concrete for policy discussion, not as forecasts. With that qualification, the evidence is consistent with a Vision 2030 policy mix anchored on five Saudi-specific instruments, each tied to an existing strategy document.

• Scale the Saudi Energy Efficiency Program (SEEP) to deliver an additional 30% reduction in energy intensity by 2030 relative to the 2020 baseline (target from the SEEP strategy). Applying the estimated post-2001 LNEE elasticity (0.108; bootstrap 95% CI [0.095, 0.122]) to that targeted change yields an illustrative reduction in per capita CO₂ emissions of roughly 3 percentage points (approximately 2.9–3.7 points across the CI), conditional on the elasticity remaining stable—contributing to SDG 7.3.
• Operationalise the Saudi Central Bank’s Sustainable Finance Roadmap to deliver USD 50 billion in cumulative green-bond issuance via the Public Investment Fund and the Saudi Green Initiative by 2030. The empirical insignificance of financial inclusion in the LNCP specification implies that financial-inclusion expansion alone is unlikely to reduce emissions; channelling new credit toward green activities is essential (SDG 8.10).
• Reorient trade composition through the National Industrial Strategy by increasing the non-oil share of merchandise exports from approximately 17% (2020 baseline) to 50% by 2030. The empirical positive trade–emissions coefficient implies that diversification away from hydrocarbon exports is required to break the trade–emissions link (SDG 12.2).
• Embed climate criteria in the National Urban Strategy: require that all Vision 2030 giga-projects (NEOM, the Red Sea Project, Qiddiya) meet or exceed LEED Gold building standards and integrate electrified public transit, exploiting the negative urbanisation–emissions coefficient documented here (SDG 11).
• Use the BTR1 reporting framework [1] as a coordinating instrument to track progress on the four channels above, with annual indicator updates published by the Saudi Centre for International Strategic Partnerships, ensuring alignment with SDG 13.2.

5.4. International Relevance

Two findings generalise beyond Saudi Arabia. First, the regime-dependent finance–environment relationship implies that policy assessments based on full-sample averages are likely to mislead in economies that have undergone significant trade-policy or financial-sector reforms during the sample. Second, the energy-intensity channel—the most consistent mitigation lever in this analysis—supports the case for prioritising efficiency programmes alongside (rather than in place of) financial-sector liberalisation as a near-term decarbonisation strategy. These insights are directly relevant to the other Gulf Cooperation Council economies undertaking comparable Vision-style diversification (the UAE, Qatar, Bahrain, Oman, and Kuwait) [21,33], to oil-exporting emerging markets in Africa, and to commodity-dependent Latin American economies.

6. Conclusions

6.1. Main Findings

This study examined the joint long- and short-run effects of financial inclusion, energy intensity, and trade openness on per capita CO₂ emissions in Saudi Arabia from 1980 to 2020. The structural-break analysis identified a statistically significant break in the cointegrating relationship at 2001 (Chow F = 7.36, p < 0.001 for LNCP), supported by the Zivot–Andrews endogenous-break dates for the financial-inclusion (2000) and trade-openness (2005) series, and confirmed by the Bai–Perron sup-F test reaching its maximum at 1990 (F = 26.37, significant at 1%). Sub-sample re-estimation around 2001 showed that energy intensity (β = +0.108), urbanisation (β = −0.680), and trade openness (β = +0.036) are robust drivers of per capita emissions only after 2001. Financial inclusion is statistically insignificant in both regimes once the mechanical correlation between carbon productivity and GDP is removed by using per capita emissions as the dependent variable—a methodological lesson with implications for the broader literature.

6.2. Limitations and Future Research

Seven limitations qualify these conclusions. First, the sample ends in 2020 because the IMF FAS Financial Institutions Access Index for Saudi Arabia is available only through that year; extending the analysis once the FAS series is updated would allow the post-pandemic period and the early stages of Vision 2030 implementation to be evaluated more fully. Second, the post-break sub-sample of 20 observations (or 16 in the 2005-policy robustness) is small in absolute terms, and the highly significant p-values must be interpreted with caution despite the supporting ridge, residual-bootstrap, and break-date-sensitivity evidence (Section 4.5.2 and Section 4.5.3 and Supplementary Table S1). Third, the energy-intensity proxy LNEE conflates genuine efficiency improvements with sectoral shifts and price-elasticity effects; future research could apply a Divisia decomposition to isolate the pure efficiency component. Fourth, the trade-openness measure aggregates exports and imports; a future decomposition into exports, imports, and product-category trade—and, ideally, an embodied-carbon or export-composition measure—would clarify the channels through which trade affects emissions. While Section 4.5.3 shows that adding GDP does not bias the estimates, the parsimonious specification omits other composition controls (oil rents, industrial value added, and energy prices); these could not be added here without exhausting the degrees of freedom in a 20-observation sub-sample, and a longer sample or a cross-country panel would permit a richer scale-and-composition specification. Fifth, the energy-intensity, trade-openness, and urbanisation regressors are plausibly jointly determined with emissions, so the long-run estimates should be read as conditional associations rather than as endogeneity-corrected causal effects; instrumental-variable or system (VECM) estimators on a longer sample could address this. Sixth, the CUSUMSQ excursion around 2010–2012 indicates that an additional minor structural break may have occurred during the post-financial-crisis oil-price recovery and the Arab Spring period; with a longer post-break sample, future work could jointly model multiple breaks. Seventh, comparative work across other Gulf Cooperation Council and oil-exporting economies could establish the external validity of the regime-dependent finance–environment result.