The connection between FD and EF is multidimensional, with economists far from reaching a consensus as to whether FD exacerbates or improves EF (Awosusi et al., 2022). The view taken by some economists is that FD can reduce EF (UNEP, 2011; Majeed and Mazhar, 2019; Zaidi et al., 2019; Odhiambo, 2020; United Nations Environment Programme; Zhao and Yang, 2020; Akinsola et al., 2021; Hussain et al., 2022; Acar et al., 2023). This view is derived from many theories in this field, such as the Green Growth Theory, which suggests that FD supports and facilitates the expansion of green industries, thus implying that FD lessens EF in diverse sectors of the economy UNEP (2011). The view is also partly underpinned by the ecological modernisation theory; although it is not directly related to the connection between FD and EF, it does point to a certain degree of complementary harmony between FD and environmental protection. A view held by other economists is that FD intensifies EF due to increased economic activity (Adams and Klobodu, 2018; Nathaniel et al., 2019; Tsaurai, 2019; Khan et al., 2021; Yang et al., 2021; Dada et al., 2022a; Hussain et al., 2023). This view is partly premised on a consumerism theory whereby improvement in economic activities can induce consumer behaviour, which in turn exacerbates EF (Pepper et al., 2009). Another theory that sheds some light on this view is the pollution refuge hypothesis which highlights the close links between FD and EF. According to this theory, economies that lack environmental regulations tend to have a higher EF because local industries do not bother about regulations and the unintended consequences of the production process, thus stimulating higher EF (Cil, 2022). Last but not least, a well-known hypothesis that underpins the finance–emission nexus is the environmental Kuznets curve (EKC). The upward sloping portion of the EKC characteristically points to a positive relationship between economic development and environmental degradation. It is in this upward sloping part of the EKC that the financial sector appears to intensify/worsen environmental degradation by enabling more production in various sectors of the economy. The downward-sloping portion of the EKC is due to the financial sector embracing and using green and modern technologies.

There is an extensive recent literature that has empirically investigated the connection between FD and EF (see Table 1). Nevertheless, the findings have far from converged, which may be characterised by factors such as methodology, sample size, control variables, proxies, and countries. Thus, the empirical studies on this subject are in three distinct categories: 1) exploring the worsening effect of FD; 2) emphasising the mitigating effect; 3) analyzing nonlinear effects. These are explained below.

The first set of studies suggests a positive correlation between FD and EF (Adams and Klobodu, 2018; Baloch et al., 2019; Tsaurai, 2019; Godil et al., 2020; Dada et al., 2022c; Ju et al., 2023). Tsaurai (2019) assessed the influence of FD on CO₂ in West African nations employing the panel data method from 2003 to 2014. One of the conclusions centres on the pooled ordinary least squares (OLS) method whereby the local credit supplied by the financial sector led to the rise of CO₂. The impact of natural resources, FD, and economic success on the EF of Malaysia from 1980 to 2019 was evaluated by Khan et al., 2021. That study adopted a dynamic simulated auto-regressive distribution lag method to find that FD, economic prosperity, and natural resources are behind the rise in EF. Reaching a similar conclusion, Dada et al. (2022c) examined the link between FD and EF in Malaysia between 1981 and 2017, utilising ARDL, fully modified OLS (FMOLS), dynamic OLS (DOLS), and canonical cointegration regression (CCR). They found that FD leads to a long-term deterioration of EF. Cross-country research by Adams and Klobodu (2018) examined 26 African countries for 1985 to 2011, utilizing the chow test, cross-country regression, and generalized method of moments (GMM) to discover that FD propels environmental degradation.

The second collection of studies drew on claims that FD promotes EF (Shahbaz et al., 2013; Majeed and Mazhar, 2019; Zaidi et al., 2019; Odhiambo, 2020; Zhao and Yang, 2020; Adebayo et al., 2021; Akinsola et al., 2021; Kihombo et al., 2021; Hussain et al., 2022; Acar et al., 2023; Wang et al., 2023). Shahbaz et al. (2013) explored the effect of FD, economic prosperity, coal utilization, and trade liberalization on environmental performance in South Africa by employing ADRL bound test time sequence data for 1965–2008. They revealed that an upsurge of economic prosperity led to a rise in energy discharge, while FD lessened it. Habiba and Xinbang (2022) evaluated the stimulus of FD on CO₂ for advanced and emerging economies by considering a multivariate measure of FD, FM, and FI indices for 2000–2018. Their study employed two-stage GMM and panel data regression to reveal that FM indices reduce EF for both advanced and emerging economies, whereas FI exacerbates EF. Arogundade, Hassan & Bila (2022) explored the influence of the dispersion proceeds on the EF of 22 African nations using fixed effects, fixed effects instrumental variables, method of moments quantile regression, and the heterogenous Granger causality test. Some of their findings confirmed that FD has a remarkable influence in reducing the environmental deterioration impact of dispersion proceeds.

Majeed and Mazhar (2019) assessed the effect of FD on the environment by utilizing an inclusive measure of EF for 131 nations for 1971–2017. A panel data method was employed, including DK standard errors and GMM. They discovered that FD proxies (local lending toward private sector, local lending to private sector by banks, and local lending supplied by financial sectors) play a critical function in enhancing environmental quality. Furthermore, their analysis revealed that local lending to the private sector is more robust when likened to former proxies of FD. Lastly, Wang et al. (2023) assessed the impact of clean energies, advanced technology industries, and the financial boom on the EFs of developed and emerging economies for 1990–2020. They used the GMM and panel quantile regression; their results suggest that FD improves EF in both types of economy.

The last group of studies produced mixed results when assessing the connection between FD and EF (Ashraf et al., 2022; Destek and Sarkodie, 2019; Saud et al., 2019. Saud et al. (2019) explored the influence of FD and globalization on the EF of 49 designated BRI nations for 1990–2014 using pooled means group long-term panel estimations. They found that EF increased because FD increased energy utilization, economic prosperity, and commerce in 30 nations. Inversely, FD decreased EF in 14 nations. Destek and Sarkodie (2019) examined the environmental Kuznet curve for EF in 11 newly industrialised economies over 1977–2013 using the heterogeneous panel causality technique and the augmented mean group. The findings confirmed that FD reduced EF only in China, Malaysia, and Singapore.

Although many of the above studies (Shahbaz et al., 2013; Adams and Klobodu, 2018; Baloch et al., 2019; Destek and Sarkodie, 2019; Majeed and Mazhar, 2019; Nathaniel et al., 2019; Saud et al., 2019; Tsaurai, 2019; Zaidi et al., 2019; Odhiambo, 2020; Shahbaz et al., 2020; Zhao and Yang, 2020; Adebayo et al., 2021; Akinsola et al., 2021; Khan et al., 2021; Kihombo et al., 2021; Yang et al., 2021; Dada et al., 2022b; Hussain et al., 2022; Acar et al., 2023; Hussain et al., 2023; Wang et al., 2023) analyzed the linear relationship between FD and the EF, very few studied the asymmetric nexus between these variables (Gill et al., 2022; Godil et al., 2020; Karasoy, 2019; Lahiani, 2019; Ling et al., 2021; Mensah and Abdul-Mumuni, 2022; Omoke et al., 2020; Shahbaz et al., 2016). The NARDL methodology eases the assumption of linearity and advocates for asymmetry among the measured variables (Shin et al., 2011). Other related empirical studies that have employed NARDL include the work of Omoke et al. (2020) in Nigeria, Lahiani, (2019) in China, Karasoy (2019) in Turkey, Gill et al. (2022) in Pakistan, and Mensah and Abdul-Mumuni (2022) in sub-Saharan Africa.

What emerges from the above literature review is that the findings on the finance–emission nexus have not yet converged, posing great difficulties in formulating a consistent environmental management policy for all countries. Part of the reason why these findings have not yet converged is that they assume a linear association between FD and emissions and use different sets of proxies and techniques for environmental quality and FD, thereby validating the need for a more nuanced account of the effect of FD on environmental quality. While this study is not the first to evaluate the asymmetrical influence of FD on environmental quality in emerging economies, the limitation of existing investigations (Lahiani, 2019; Tsaurai, 2019; Habiba and Xinbang, 2022; Hussain et al., 2022; Manigandan et al., 2022; Wu et al., 2022; Hussain et al., 2023; Ofori et al., 2023; Qalati et al., 2023; Raihan, 2023) is that they rely heavily on CO₂ emissions, which only account for air pollution but not problems for water, soil, and other aspects of the environment. Moreover, some of these studies also rely on traditional measures of FD, thereby ignoring multifaceted measures of FD which allow for aspects like depth, access, and efficiency across financial institutions and markets.

Considering this gap, this study advances the literature by examining a comprehensive measure of FD and disaggregates it into its sub-indices to provide a nuanced analysis, helping the reader to appreciate how components of FD affect EF. Second, this study used ARDL for the linear effects and NARDL for non-linear effects of FD to produce robust policy suggestions. In bridging these gaps, this empirical work aims to add to the present body of knowledge by raising the following questions: 1) Is the FD-EF nexus asymmetric in South Africa? 2) Does the disaggregated FD (FM and FI) have the same impact on EF?

This research utilized 38-year time-series dataset for 1980–2018. The rationale for using this timeframe is that 1980 marks the beginning of numerous variables for South Africa while 2018 was chosen as the end of the period due to the unavailability of data for some variables. EF presented as the dependent variable denoting “ecological footprint” (constant per capita) and assessed by means of a consumer-based method and measured in global hectares (gha) per capita (See Table 2). The core explanatory variables were financial market, financial institution development indices, and overall financial development (FM, FI, and FD) established by the International Monetary Fund (2019). The justification for this index is that it accounts for the complex multifaceted nature of financial development which traditional measures do not consider. The other control variables were chosen because they have been commonly used as covariates in preceding studies. We controlled for gross domestic product per capita—GDP divided by population (constant 2010 US$)—and ENUSE—energy use per capita (measured in kg of oil equivalent per capita) —following Adebayo et al. (2021), Arogundade et al. (2022), Nathaniel et al. (2019), Omoke et al. (2020), and Shabhaz et al. (2013). To obtain dependable results, all proxies were converted into natural logarithms in order to reduce variability and make data move closer to normal distribution. The EF variable is obtained from global footprint network (GFN), financial development from the International Monetary Fund (IMF), and gross domestic product per capita and energy use from the World Development Indicators.

The main intent of this research was to assess the asymmetric effect of FD on EF in South Africa. Several studies have used FD as the proxy for evaluating its influence on EF (Shabhaz et al., 2013; Nathaniel et al., 2019; Kihombo et al., 2021; Gill et al., 2022; Wu et al., 2022). The relationship between FD, FM, FI, GDP, ENUSE, and EF is presented in the following functional form.

Equation 1 regresses EF on the aggregated measure of FD, which combines FM and FI. Eqs 1a and 1b regress EF on the unbundled measure of FD (FM and FI). E F t = β 0 + β 1   F D t + β 2   G D P P C t + β 3   E N E U S E t + ε t . (1) E F t = β 0 + β 1   F M t + β 2   G D P P C t + β 3   E N E U S E t + ε t . (1a) E F t = β 0 + β 1   F I t + β 2   G D P P C t + β 3   E N E U S E t + ε t . (1b)

In this case, EF depicts the ecological footprint in time t, FD represents financial development in that time, GDPPC denotes GDP per capita, and ENEUSE shows energy use. β 0 is denoted as a constant term, while β 1 β 2   β 3   are expressed as coefficients of parameters. On the opposite, ε t is the error term. The study mimicked the ARDL estimation technique established by Pesaran et al., (2001) and specified below: Δ E F t = δ 0 + ∑ n = 1 p   δ 1 Δ E F t − k + ∑ n = 0 p   δ 2 Δ F D t − k + + ∑ n = 0 p   δ 3 Δ G D P P C t − k + ∑ n = 0 p   δ 4 Δ E N U S E t − k + π 1 Δ E F t − k + π 2 Δ F D t − k + + π 3 Δ G D P P C t − 1 + π 4 Δ E N E U S E t − 1 + e t . (2)

Equation 2 has the advantage of providing a combination of short- and long-term estimates in the same equation. In addition, ARDL estimates can be utilized even if the integration order is I (1), I (0), or a mix of both. The coefficients along with Δ demonstrate the short-term estimates, π 1 , π 2 ,   π 3 ,   π 4 symbolize long-term estimates, and e t represents a residual term. On the long-term estimation, the primary step is to assess the level of cointegration among variables. This study investigated the joint significance of lagged level variables by adopting the F-statistics test as recommended by Pesaran et al. (2001). When the calculated bound test value exceeds the critical values, it indicates that the variables are integrated, confirming the long-term relationship. Inversely, when the bound test F statistic is lower than the critical values, it suggests that there is no long-term relationship between variables. This study also applied the double unit root tests of augmented Dickey–Fuller (ADF) and Ng Perron to determine stationarity. Ng Perron was adopted to further augment the ADF, which assumes a constant or linear trend, whereas Ng Perron permits a more flexible modelling approach to detect mobility in time-series data.

Furthermore, the research assessed the non-linear influence of FD on the EF to determine whether positive and negative shocks have the same impact. The study employed the non-linear ARDL econometric structure recommended by Shin et al. (2014). This methodology assesses the asymmetric effect which shows a combination of positive and negative shocks of the respective variable(s). The NARDL methodology has been extensively used in environmental economics (Omoke et al., 2020; Das and McFarlane, 2021; Ling et al., 2021; Dada et al., 2022c; Biyase and Naidoo, 2023). Moreover, a nonlinear auto regressive distributed lag model was chosen over other econometric methods since the former can capture the multifaceted, nonlinear relationships which are regularly detected in time-series data, thereby yielding a true picture of real-world dynamics compared to simple linear models. It can also cope with various econometric problems such as asymmetries and structural breaks, allowing the robust analysis of economic data.

In this study, FD was the core explanatory variable disaggregated into a positive and negative shock. The functional form is as follows: F D t + = ∑ j = 1 t ∆ F D t + = ∑ j = 1 t = max F D t + , 0 . (3) F D t − = ∑ j = 1 t ∆ F D t − = ∑ j = 1 t = min ∆ F D t − , 0 . (4)

Eqs 3, 4 demonstrate variables in partial sum, while Eq. 5 considers NARDL. To signify the NARDL equation, two partial variables are replaced in Eq. 2, resulting in Eq. 5 below. Δ E F t = δ 0 + ∑ n = 1 p   δ 1 Δ E F t − k + ∑ n = 0 p   δ 2 Δ F D t − k + + ∑ n = 0 p   δ 3 Δ F D t − k − + ∑ n = 0 p   δ 4 G D P P C t − k + ∑ n = 0 p   δ 5 Δ E N U S E t − k + π 1 Δ E F t − k + π 2 Δ F D t − k + + π 3 Δ F D t − k − + π 4 Δ G D P P C t − 1 + π 5 Δ E N E U S E t − 1 + e t . (5)

The study conducted diagnostic tests which ascertained the validity of the coefficients used, and furthermore determined whether the models are favourable for interpretation. The Breusch–Godfrey test was used to detect autocorrelation in the regression models. A normality test was conducted using Jarque–Bera. The BDS test was carried out to validate the non-linearity attributes of all variables. Furthermore, CUSUM and CUSUMSQ tests were performed to determine the stability of coefficient stability. Lastly, for robustness, this paper also adopted the fully modified ordinary least squares (FMOLS), dynamic OLS (DOLS), and canonical cointegrating regression (CCR).

The descriptive statistics given in Table 3 above show that most of the proxies used in this study—EF, FI, and FD index—were relatively stable, which is shown by the values of the standard deviations, which are close to 0. However, GDP demonstrates a higher variability since the standard deviations are distant from 0.

The research performed some pre-estimation tests which were necessary to make the estimations possible and the results dependable. The ADF and Phillips Parron tests were employed to check for stationarity. The null hypothesis for the tests was that there is a unit, meaning that the series is non-stationary against the alternative hypothesis that the variable is stationary. As a rule of thumb, a null hypothesis is rejected when the test statistic is above the critical values at 10%, 5%, and 1%. Table 4 summaries the test results conducted.

The outcomes of both tests are in harmony, ascertaining that the series in the study are both I (0) and I (1) variables. Moreover, the null hypothesis of the unit root was not rejected because all variables except GDPCC are stationary at level, and stationary variables become stationary after the first difference. The results of the unit root test validate the adoption of the ARDL model, as no variable is beyond I (1).

Table 5 depicts the F-statistics and findings of the diagnostic tests of the model specifications for assessing the impact of FD, GDP, and energy use on EF. Three specifications concerning linear ARDL bounds tests were undertaken in this study. The outcomes for all models reveal that the F-statistics for the models are significantly larger than the upper bound. Consequently, these findings point to the rejection of the null hypothesis and acceptance of the alternative, according to which there is a long-term correlation between EF, FD, FM, FI, GDP per capita, and energy use in South Africa.

Table 5 also exhibits outcomes of the diagnostic tests, demonstrating that the model has successfully passed all the tests: Durbin–Watson stat, Jarque–Bera normality test, BG serial correlation LM test (F-statistic), BG serial correlation LM test (R-squared), heteroskedasticity test: (ARCH), and Ramsey RESET test.

The non-linear ARDL F-statistics (Table 6) were also found to be higher than the upper bounds at all levels (1%, 5%, and 10%), indicating that there is a long-term relationship between EF, FD, FM, FI, and GDP per capita, including energy use in South Africa. To ascertain the linearity of the parameters, the study utilized the BDS test initiated by Broock et al. (1996) to capture non-linearity. The results for the BDS test are shown in Table 7 in the appendix, demonstrating that the variables have nonlinearities. These results allow us to employ non-linear ARDL to examine the relationship between financial development and ecological footprint.