The Impact of Climate Change on Financial Stability in South Africa

Abstract

This study investigates the dynamic relationships between climate change and financial stability in South Africa by employing a Bayesian vector autoregression model (BVAR). Using data from 1991 to 2022, we examine the impact of carbon emissions, adjusted savings, renewable energy consumption, lending interest rates, and unemployment on financial stability. Our findings indicate that carbon emissions, adjusted savings damaged by carbon dioxide emissions, renewable energy consumption, and unemployment significantly erode financial stability. Impulse response functions reveal that shocks to carbon emissions, lending interest rates, and unemployment have lasting effects on financial stability. Forecast error variance decomposition analysis shows that external factors, particularly carbon emissions and lending interest rates, increasingly drive uncertainty in forecasting financial stability over time. The study’s results support the Financial Instability Hypothesis and the Diamond–Dybvig model, highlighting the importance of considering climate-related risks in financial stability analysis. The findings have significant implications for policymakers and financial regulators seeking to promote financial stability and mitigate climate-related risks in South Africa.

1. Introduction

Climate change is increasingly recognized as a significant source of financial risk with the potential to destabilize global economies (Wu et al., 2023). Human activities, particularly burning fossil fuels and deforestation, have led to unprecedented changes in weather patterns, including rising temperatures, melting ice cover, and more extreme weather events driven by climate change (Climate Change Knowledge Portal, 2021). Over the past century, the world has experienced more frequent droughts, floods, and heatwaves, each contributing to disruptions across economic sectors, including agriculture, energy, and finance (Fabris, 2020).

South Africa, which is heavily dependent on rainfed agriculture and natural resources, faces a particular vulnerability to these changing climatic conditions and the effects of climate change (The World Bank Group, 2021). The Intergovernmental Panel on Climate Change (IPCC, 2021) confirms a significant increase in global average temperatures. Their comparison of Earth’s temperature between 1850–1900 and two recent periods (1995–2014 and 2011–2020) reveals a clear warming trend, with temperatures increasing from 0.85 °C to 1.09 °C. However, South Africa’s annual temperature has risen 1.5 times faster than the global average of 0.65 °C over the past 50 years, accompanied by more frequent extreme rainfall events and rising temperatures, both consequences of climate change (Ziervogel et al., 2014).

This accelerated warming trend underscores the urgent need to address the financial risks posed by climate change in South Africa. As the frequency and severity of extreme weather events, driven by climate change, continue to escalate, they threaten to disrupt key industries, including banking and insurance, with far-reaching consequences for financial stability (Pointner & Ritzberger-Grünwald, 2019; Wu et al., 2023). These events are not only physical in nature but also social in nature, exacerbating unemployment and income inequality, which are forms of climate-related social disruption, and, in turn, weakening the resilience of the financial system. Such economic instability hinders financial institutions’ ability to recover from shocks, reduces investor confidence, and may lead to a vicious cycle of economic decline (Brunetti et al., 2022).

Despite growing recognition of the impact of climate change on financial stability, the existing literature presents mixed conclusions globally. While some studies identify clear threats (Chabot & Bertrand, 2023; Z. Liu et al., 2024) and others highlight complex interactions (Z. Liu et al., 2021; Brunetti et al., 2022), a significant limitation is the scarcity of in-depth analyses focusing specifically on the South African context and the financial implications of its energy transition, a key aspect of managing the transition risks associated with climate change mitigation. Existing global studies, with their diverse methodologies, geographic scopes, and timeframes, may not fully capture the unique vulnerabilities and interconnectedness of South Africa’s fossil fuel-reliant economy and its financial system as it navigates the shift towards cleaner energy sources to mitigate carbon emissions.

Therefore, this study directly addresses this critical gap by providing a focused empirical investigation into the specific impacts of carbon emissions (CO2M) and the resulting transition policies, particularly renewable energy consumption, on South Africa’s financial stability. The use of carbon emissions as a proxy for physical climate risk and renewable energy consumption as a proxy for managing transition risk allows for a nuanced analysis. By employing a Bayesian vector autoregression (BVAR) model, this study offers a robust analytical approach within the South African context, allowing for the incorporation of prior information to provide a more nuanced understanding of these complex dynamics. Ultimately, this research aims to offer a robust, South Africa-specific analysis that will contribute both theoretically, by enriching our understanding of the financial stability implications of energy transition in a developing economy, and empirically, by providing crucial evidence for policymakers and financial institutions operating within the South African landscape as they navigate the challenges and opportunities of a shift towards renewable energy.

Therefore, this study will test the following hypothesis:

• Climate change has no statistically significant impact on South Africa’s financial stability.

The structure of this paper is as follows: Section 2 provides an overview of the relevant literature, Section 3 outlines the proposed model, Section 4 presents the results from the BVAR analysis, Section 5 Robustness analysis and Section 6 concludes the study with a discussion of policy implications.

2. Literature Review

Major theories of financial stability, such as Minsky’s (1992) instability theory, posit that financial markets and institutions are inherently unstable due to vulnerabilities arising from interactions between financial institutions and their debtors and creditors, particularly concerning liquidity and term mismatches. While Minsky’s original framework focused on internal dynamics, we argue that exogenous shocks like climate change and its associated risks act as significant amplifying factors. For instance, the increasing frequency and severity of extreme weather events driven by climate change can lead to unexpected asset devaluation, increased insurance payouts, and disruptions to economic activity, thereby straining the balance sheets of financial institutions and their borrowers, potentially triggering a shift towards more fragile financing structures as described by Minsky.

Similarly, Diamond and Dybvig’s (1983) theory on bank runs, which highlights the fragility of banks due to maturity transformation and asymmetric information, gains a new dimension when considering climate risks. Climate-related disasters can trigger widespread economic hardship and uncertainty, potentially leading to a loss of confidence in financial institutions and increasing the likelihood of coordinated withdrawals, even for solvent banks, due to the systemic nature of climate shocks. Our study builds upon these foundational theories, integrating the exogenous risks posed by climate change as a critical driver that can trigger and amplify the endogenous mechanisms of financial instability within the South African context.

Climate change introduces significant risks to financial stability, primarily categorized as physical and transition risks (Grippa & Demekas, 2021). Physical risks stem from the financial harm caused by the long-lasting effects of climate change. As Wu et al. (2023) observe, climate extremes, which are severe impacts of climate change, pose a substantial threat to human life and social wealth, potentially triggering climate-related social issues like unemployment and migration, which can destabilize economies and financial systems. This study uses carbon emissions as a proxy to measure physical climate risk, arguing that higher carbon emissions exacerbate these risks, leading to asset stranding and devaluation (Batten et al., 2016) and reduced economic growth and investment returns (Dietz et al., 2016). The empirical literature largely supports the view that physical climate risks have a negative impact on financial stability. For instance, Varntanian and Pancera (2020) argue that climate-induced disasters can damage the balance sheets of financial institutions, leading to widespread financial instability.

Transition risks arise from policy changes and technological advancements aimed at mitigating climate change, such as carbon pricing and investments in renewable energy. These can create financial risks for businesses and investors (Grippa & Demekas, 2021). Specifically, transition risks can impact financial stability by increasing credit and liquidity risks for firms heavily invested in carbon-intensive industries (Batten et al., 2016) and stranding assets, thereby reducing investment returns (Dietz et al., 2016). This study uses renewable energy consumption as a proxy to measure transition risks, reflecting the efforts to move away from carbon-intensive energy sources.

In addition to these direct impacts, climate change can also affect financial stability indirectly. Climate-related social disruptions, such as unemployment, can lead to social unrest, increased crime, and disruptions to economic activities (Wu et al., 2023). Interestingly, Mueller et al. (2020) observed that precipitation anomalies (a consequence of climate change) were negatively associated with unemployment in middle-income Africa. This suggests that during such anomalies, increased local labor demand reduces migration rates, which helps offset production risks and encourages workers to remain in local economies. This rise in unemployment then impacts banking stability through two primary channels identified by the International Monetary Fund (2008): the Non-Performing Loans (NPL) effect, where higher unemployment reduces households’ repayment ability and increases default rates; and the Demand for New Loans effect, where rising unemployment lowers consumer and business confidence, decreasing the demand for new loans. Both channels negatively affect banks’ profitability and liquidity.

The empirical literature presents varying perspectives on the relationship between climate change and financial stability. Several studies suggest a negative relationship. For example, Ayele and Fisseha (2024) found that climate change exerts a significant negative effect on financial stability in Sub-Saharan Africa, with positive shocks to mean temperature (a manifestation of climate change) leading to substantial negative adjustments to financial stability. Similarly, Z. Liu et al. (2024) found that climate risk negatively affects financial stability across 53 countries. Other studies, such as those from Dafermos et al. (2018), Nieto (2019), Wu et al. (2023), Chabot and Bertrand (2023), and Le et al. (2023), also support this view, indicating that both physical and transition risks have the potential to undermine financial stability.

However, some research presents a more nuanced picture, suggesting a mixed relationship. Z. Liu et al. (2021) found that climate change could lead to both positive and negative climate shocks that could either harm or benefit financial stability. Meng et al. (2023) found that climate change policies in China had a negative impact on financial stability in the short run but a positive impact in the long run. These findings highlight the complexity of climate change’s economic effects and indicate that its relationship with financial stability is not always straightforward. Brunetti et al. (2022) also contribute to this view.

A third strand of the literature acknowledges the significant relationship between climate change and financial stability but does not specify a clear positive or negative direction. Pointner and Ritzberger-Grünwald (2019) found a significant relationship between climate change and financial stability in European countries, with Conlona et al. (2022) emphasizing that climate risk could contribute to higher tail risks and systemic risks in the banking sector. Curcio et al. (2023) found that while climate-induced disasters showed a significant correlation with financial risk in U.S. banks and insurers, they did not necessarily affect systemic risk, suggesting that the effects of climate risks on financial stability may vary based on specific events and their geographic location, a point also highlighted by Pagnottoni et al. (2022).

Historically, Carney (2015) outlined physical, liability, and transition risks as key channels through which climate change could affect financial stability, emphasizing the long-term risks to financial resilience. While Dafermos et al. (2018) supported this view with an ecological macroeconomic model, studies like those from Z. Liu et al. (2021) and Brunetti et al. (2022) suggest that climate change could have both positive and negative effects depending on the context. Recent studies continue to provide varying perspectives. While Wu et al. (2023), Chabot and Bertrand (2023), and Le et al. (2023) argue for a detrimental impact of climate change on financial stability, Curcio et al. (2023) and Pagnottoni et al. (2022) suggest that the effects of climate events on financial stability depend on the type and location of the event. Meng et al. (2023) reported a negative short-term and positive long-term impact of climate change policies, and Z. Liu et al. (2024) found that climate risk negatively affects financial stability across 53 countries.

In summary, the literature suggests that climate change, through physical and transition risks, can significantly impact financial stability. While some studies emphasize the negative impacts of climate change on financial stability, others suggest more complex relationships, with both positive and negative effects or significant relationships without a specified direction. The varying findings may be attributed to differences in methodologies, regional focus, and the specific types of climate events or policies examined. This study aims to contribute to this body of literature by examining the impact of carbon emissions (as a proxy for physical climate risk), adjusted savings, renewable energy consumption (as a proxy for transition risk management), lending interest rates, and unemployment (as a potential indicator of climate-related social disruption) on financial stability in South Africa.

Table 1. Climate change impact on the financial system.

	Credit Risk	Market Risk	Operation Risk
Physical Risk	- Increasing flood risk to mortgage portfolio - Declining agricultural outputs	- Re-pricing of sovereign debt	- Impact on business continuity
Transitional Risk	- Tightening technological standards affects company business - Carbon taxes lead to growing expenses - Prohibition of use of outdated technologies - Long-term investments become unprofitable - Innovations jeopardize companies’ business based on outdated technologies	- Tightening climate-related policy leads to re-pricing of securities	- Changing sentiment on climate issues leads to reputational risk
Indirect risk	- Losses for companies connected with firms affected by climate change	- Re-pricing of securities	- Low probability of negative impact (jeopardizing of a bank’s supply chains)

Source: Author’s modifications based on the Batten et al. (2016) Transition in thinking: The impact of climate change on the UK banking sector, Bank of England, London.

summarizes the various climate related risks and their effects on financial stability.

3. Empirical Approach

3.1. Sample Selection and Data Source

This study considers South Africa as the research target. Due to data availability constraints, the study will utilize a Bayesian vector autoregression (BVAR) model using annual data from 1991 to 2022. It is important to acknowledge that this time period may not reflect the most recent data.

To account for financial stability, the study employs the Financial Conditions Index (FCI), which is compiled by EasyData. The FCI aggregates a range of financial indicators, including interest rates, credit spreads, and stock market performance, to provide a comprehensive measure of financial conditions. It reflects the broader impact of monetary policy on the economy, signalling the overall health of financial markets, especially in times of high market volatility. By incorporating factors such as inflation, exchange rate movements, and credit demand, the FCI serves as a valuable tool for assessing financial stability and guiding monetary policy decisions (EasyData).

For climate change, we used Carbon Dioxide Emissions (CO2M). Control variables include renewable energy consumption (REC), Adjusted Savings: Carbon Dioxide Damage (ASCDD), lending interest rate (LIR), and unemployment (UN). Data was sourced from the World Bank, EasyData, and International Monetary Fund websites.

3.2. Justification of Variables

Carbon Emissions (CO2M): Carbon dioxide emissions are those stemming from the burning of fossil fuels and the manufacture of cement. They include the carbon dioxide produced during the consumption of solid, liquid, and gas fuels and gas flaring. This directly reflects the impact of climate change. Higher emissions contribute to climate change, potentially impacting financial stability through extreme weather events, resource scarcity, and disruptions to key economic sectors.

Renewable Energy Consumption (% of total final energy consumption) (REC): Renewable energy consumption is the share of renewable energy in total final energy consumption. This captures efforts to mitigate climate change and potentially improve financial stability by reducing dependence on fossil fuels and increasing energy security.

Adjusted Savings: Carbon Dioxide Damage (ASCDD): Cost of damage due to carbon dioxide emissions from fossil fuel use and the manufacture of cement, estimated to be USD 40 per ton of CO2 times the number of tons of CO2 emitted. This variable accounts for the potential economic costs associated with carbon emissions. By adjusting savings for the damage caused by emissions, this provides a more nuanced picture of financial health.

Lending Interest Rate (LIR): This is the rate that commercial banks charge their most creditworthy customers. This reflects the health of the financial system. Higher interest rates can indicate financial stress or attempts to control inflation, potentially impacting financial stability.

Unemployment (UN): High unemployment rates can lead to decreased economic activity and put a strain on social safety nets, negatively impacting financial stability.

For clarity and completeness,

Table 2. Variable Description.

Variable Name	Abbreviation	Unit of Measurement	Source	Variable Type
Financial Conditions Index	FS	Index	Statista	Dependent
CO2 Emissions	CO2M	Kt (kiloton)	The World Bank	Independent
Lending interest rate	LIR	percentage	The World Bank	Control
Unemployment rate	UN	percentage	The World Bank	Control
Renewable Energy Consumption	REC	percentage	The World Bank	Control
Adjusted Savings: Carbon Dioxide Damage	ASCDD	Percentage	The World Bank	Control

offers a comprehensive breakdown of the variables central to this investigation. Each row details the variable’s full name, its corresponding abbreviation, the unit in which it is measured, its primary data source, and its classification by variable type

3.3. Pre-Estimation Tests

• Unit Root Tests: Unveiling the Order of Integration

Two commonly employed formal methods for checking stationarity are used, namely, the Augmented Dickey–Fuller (ADF) test and the Phillips–Perron (PP) test (Enders, 2010). Identifying the order of integration using unit root tests is essential because it guides the selection of appropriate models for further analysis.

Augmented Dickey–Fuller (ADF) Test

Developed by Dickey and Fuller in the 1970s (Gujarati & Porter, 2009), the ADF test is used to assess the presence of a unit root in time series data. It is considered an “augmented” version of the Dickey–Fuller test, offering a more robust approach for a wider range of time series models (Enders, 2010).

The ADF test operates within an autoregressive framework. It examines the current value of a variable (Y) about its lagged value (Y(t − 1)), a potential deterministic trend (represented by β), and past differences (∆Y(t − i)) to account for any non-stationarity of lower orders (represented by α). The number of lagged differences (m) is chosen based on statistical criteria. Finally, an error term (ϵt) captures any unexplained variations.

The null hypothesis of the ADF test posits the existence of a unit root (δ = 0). If the estimated value of δ is sufficiently negative compared to the established critical values, we reject the null hypothesis and conclude that the series is stationary. H0: P − 1 = 0.

The Phillips–Perron (PP) Test

The Phillips–Perron (PP) test, developed by Peter C. B. Phillips and Pierre Perron in the late 1980s (Gujarati & Porter, 2009), is another widely used statistical test for determining the presence of a unit root in a time series. Similar to the Augmented Dickey–Fuller (ADF) test, the PP test is a nonparametric test that does not rely on specific assumptions about the error term distribution.

The null hypothesis of the PP test is the same as the ADF test: the time series has a unit root, meaning it is non-stationary.

H0: 

The time series has a unit root.

The PP test calculates a test statistic that measures the significance of the estimated coefficient on the lagged dependent variable. Therefore, the test statistic is compared to critical values from the PP test distribution. These critical values are typically obtained from statistical tables or software packages. If the calculated test statistic is less than the critical value, we reject the null hypothesis and conclude that the time series is stationary. Otherwise, we fail to reject the null hypothesis, suggesting that the time series is non-stationary.

• Lag Length in VAR

Vector autoregressions (VARs) are a popular tool for analysing the dynamic relationships between multiple time series variables. A crucial step in building a VAR model is selecting the optimal lag length. The lag length refers to the number of past periods included for each variable in the model. Choosing the right lag length is a balancing act. Including too few lags can lead to an incomplete picture, potentially missing important past influences on the variables. Conversely, incorporating too many lags can introduce noise and spurious correlations, increasing model complexity and potentially leading to inaccurate forecasts (Lütkepohl, 2005).

Information criteria like the Akaike Information Criterion (AIC) offer a solution. These criteria penalize models for both complexity (number of estimated parameters) and goodness-of-fit. The model with the lowest AIC value is considered optimal. AIC is favoured for its focus on relative fit, meaning it prioritizes models that capture relevant dynamics without overfitting, and its widespread use and interpretability across various time series models (Enders, 2010). Although other information criteria such as SIC and HQIC exist, AIC remains a strong starting point due to these advantages.

• Random Forest Test

Random forests (RFs) are very popular ML models that can be applied to both regression and classification tasks. Like CART, they also rely on decision trees for training, but in a much higher amount, making up a “forest” of trees. Similarly, for CART, they do not need any data preparation prior to the application of the model. In addition, they perform implicit on-the-run feature selection and provide more accurate indicators of feature importance. Moreover, they are unlikely to perform overfitting and they are relatively quick to train and versatile (Ho, 1998).

Prior to the estimation of our Bayesian vector autoregression (BVAR) model, we conducted a Random forest (RF) analysis (Breiman, 2001) to gain initial insights into the relative importance of our chosen variables in predicting financial instability. This pre-estimation step helps us understand the potential influence of each variable before delving into their dynamic and interconnected relationships within the BVAR framework.

The RF analysis was implemented to identify the variables that contribute most significantly to the variance in our measure of financial stability (FCI). By training the RF model to predict FCI using the other variables (CO2M, ASCDD, REC, LIR, and UN), we obtained a ranking of variable importance based on their contribution to reducing prediction error. The results of this pre-estimation RF analysis, including the ranking of variable importance, are presented and discussed in the initial part of our Empirical Results section (Section 4.1).

3.4. Bayesian VAR Model: Model Specification

Bayesian Vector Autoregression (BVAR) Model.

To meet the study’s objectives, we implemented a Bayesian vector autoregression (BVAR) model, building on the methodology detailed by Kuschnig and Vashold (2021). Following their approach, our BVAR model incorporates hierarchical priors tailored to address the specific requirements of our research. A VAR model of finite order p, denoted as VAR(p), can be represented as follows:

y_t = α₀ + A₁y_t−1 + ··· + A_py_t−p + ϵ_t   with ϵ_t ∼ N (0, Σ)

(1)

where y_t = FCI_t; CO2M_t; UN_t; LIR_t; ASCDD_t; REC_t is a 6 × 1 column vector of 6 endogenous variables in the BVAR system, while α₀ denotes a 6 × 1 vector of the intercept. A_j (j = 1, …, p) denotes a 6 × 6 matrix of autoregressive coefficients of regressors, while p is the order of the BVAR, and, lastly, ϵ_t is a 6 × 1 vector of Gaussian exogenous shocks with a zero mean and variance covariance (VCOV) matrix Σ. The 6 × 6²p are the number of coefficients to be estimated, rising quadratically with the number of included variables and linearly in the lags order. Such parameterizations can lead to inaccuracies regarding structural inference and out-of-sample projecting, especially for higher-dimensional models. This phenomenon is normally referred to as the curse of dimensionality.

The strength of the Bayesian approach to VAR estimation lies in its ability to overcome the limitations of traditional methods. It achieves this by imposing additional structure on the model, notably through the use of priors. These priors have proven effective in mitigating the curse of dimensionality, enabling the estimation of larger models (Doan et al., 1984). They guide the model parameters toward a more parsimonious baseline, which reduces estimation errors and improves the accuracy of out-of-sample forecasts (Koop, 2013). This shrinkage is similar to the regularization techniques used in frequentist statistics (Mol et al., 2008). Our Bayesian approach provides substantial flexibility by allowing us to incorporate a wide array of prior information, informed by both economic theory and the features of the available data. Moreover, the hierarchical modelling framework enhances the robustness of our estimates by effectively accounting for parameter uncertainty.

3.5. Prior Selection and Specification

Properly informing prior beliefs is critical, hence it is the subject of much research. In the multivariate context, flat priors, which attempt not to impose a certain belief, yield inadmissible estimators (Stein, 1956) and poor inference (Sims, 1980; Bańbura et al., 2010). Other uninformative or informative priors are necessary. Early contributions (Litterman, 1980) set priors and their hyperparameters in a way that maximizes out-of-sample forecasting performance over a pre-sample. Del Negro and Schorfheide (2004) choose values that maximize the marginal likelihood. Bańbura et al. (2010) use the in-sample fit as a decision criterion and control for overfitting. Economic theory is a preferred source of prior information but is lacking in many settings for high-dimensional models. Acknowledging this, Villani (2009) reformulates the model and places priors on the steady state, on which economic theory often focuses and is hence better understood by economists.

Giannone et al. (2015) propose setting prior hyperparameters in a databased fashion, i.e., by treating them as additional parameters to be estimated. In their hierarchical approach, prior hyperparameters are assigned their own hyperpriors. This can be expressed by invoking Bayes’ law as

p(γ|y) ∝ p(y|θ, γ) p(θ|γ) p(γ),

(2)

p(y|γ) = ∫ p(y|θ, γ) p(θ|γ)dθ,

(3)

where y = (y_p + 1, …, YT)^T, while θ denotes the variance and autoregressive parameters of the VAR model and γ represents the set of hyperparameters. The first part of Equation (1) is marginalized with respect to the parameters θ in Equation (2). This yields a density of the data as a function of the hyperparameters, p (y|γ), also known as the marginal likelihood (ML). This quantity is conditional on the hyperparameters γ but marginal with respect to the parameters θ. The marginal likelihood test provides a decision criterion for both the maximization process and the choice of hyperparameters. This approach aligns with an empirical Bayes method, offering a clear frequentist interpretation (Giannone et al., 2015). In our approach, the marginal likelihood is used to explore the entire posterior hyperparameter space, acknowledging the inherent uncertainty. Giannone et al. (2015) demonstrate that this method, when implemented efficiently, produces robust and theoretically sound results. They showed that the model performs competitively against factor models, achieving high accuracy in forecasts and impulse response functions. Since its introduction, their approach has been widely adopted in applied research (Altavilla et al., 2019). A key feature of their contribution is the emphasis on conjugate prior distributions, specifically the Normal-inverse-Wishart (NIW) family.

Conjugacy implies that the marginal likelihood (ML) can be expressed in a closed form, enabling efficient computation. The Normal-inverse-Wishart (NIW) family encompasses many of the most commonly used prior distributions, although there are some notable exceptions. These include the Dirichlet Laplace prior, the steady-state prior, and the Normal-Gamma prior. Recent work in this area has focused on incorporating heteroskedastic error structures that could enhance model performance. However, this is not feasible within the conjugate setup and would also complicate inference. Within the selected NIW framework, we address the model in Equation (1) by defining A = [α0, A1, …, AP]^T and β= νec(A). Then, the conjugate prior setup reads as

β|Σ ∼ N (b, Σ ⊗ Ω),
Σ ∼ IW(Ψ, d),

(4)

where b, W, Ψ, and d are functions of a lower-dimensional vector of hyperparameters γ. In their paper, Giannone et al. (2015) consider three specific priors: the so-called Minnesota (Litterman) prior, which is used as a baseline; the sum of coefficients prior; and the single-unit-root prior (see also Sims & Zha, 1998). The Minnesota prior (Litterman, 1980) imposes the hypothesis that the individual variables all follow random walk processes. This parsimonious specification typically performs well in forecasts of macroeconomic time series (Kilian & Lütkepohl, 2017) and is often used as a benchmark to evaluate accuracy. The prior is characterized by the following moments:

$$\mathrm{E}[({\mathbf{A}}_{\mathbf{s}}{)}_{\mathrm{ij}}|\mathsf{\Sigma }]=\left\{\begin{array}{l}1ifi=jands=1,\\ 0otherwise\end{array}\right.$$

(5)

$$\mathrm{cov}[({\mathbf{A}}_{\mathbf{s}}{)}_{\mathrm{ij}},{({\mathbf{A}}_{\mathbf{r}})}_{\mathrm{kl}}|\mathsf{\Sigma }]=\left\{\begin{array}{l}{\lambda }^2{\displaystyle \frac{1}{s^{\alpha }}}{\displaystyle \frac{\Sigma ik}{\psi j/(d-M-1)}}ifI=jandr=s,\\ 0otherwise\end{array}\right.$$

(6)

Here, λ is the key parameter that regulates the tightness of the prior distribution and thus determines the relative influence of the prior information versus the sample data. As λ approaches 0, the priors become more influential, effectively imposing them with greater precision. Conversely, as λ approaches infinity (λ → ∞), the posterior estimates converge towards the ordinary least squares (OLS) estimates, indicating that the data dominates the prior. The parameter Ψ, on the other hand, controls the prior’s standard deviation on the lags of variables other than the dependent variable. In addition to this, the Minnesota prior is commonly included in the model to mitigate the impact of the deterministic component that arises from the VAR model’s estimated conditioning on the initial observations. Furthermore, Doan et al. (1984) incorporated the sum-of-coefficients (SOC) prior based on the idea that a “no-change” forecast is optimal at the beginning of a time series. This is implemented using Theil’s mixed estimation method, which involves adding artificial dummy observations to the data matrix. These dummy observations are constructed as follows:

$$\stackrel{y^1}{M\times M}=diag\left({\displaystyle \frac{\underset{\mathrm{y}}{¯}}{\mu }}\right)+\begin{array}{c}{x}^{+}\\ M\times \left(1+Mp\right)\end{array}=[0,y^{+},\dots ,y^{+}]$$

(7)

In Equation (7), y represents an M × 1 vector of the averages calculated over the first p observations of each variable. The variance is governed by the key parameter µ. Consequently, as µ approaches infinity (µ → ∞), the prior becomes uninformative due to the prior’s reduced tightness. Conversely, as µ approaches zero (µ → 0), the model is drawn towards a specification with as many unit roots as there are variables and no cointegration. This concept underlies the single-unit-root (SUR) prior (Sims & Zha, 1998), which accommodates cointegration relationships within the data. The prior guides the variables either towards their unconditional mean or towards a state with at least one unit root. These types of priors are linked to the following dummy observation:

$${Y_{1\times M}}^{++}=\left({\displaystyle \frac{\underset{\mathrm{y}}{¯}}{\delta }}\right)+\begin{array}{c}{x}^{++}\\ 1\times (1+Mp)\end{array}=\left[{\displaystyle \frac{\underset{\mathrm{y}}{¯}}{\delta }},y^{++},\dots ,y^{++}\right]$$

(8)

Here, y is defined as before. Similarly, δ is the key parameter that governs the tightness of the SUR prior. Several heuristics have been suggested for setting this parameter, as discussed in various studies, including Doan et al. (1984) and Bańbura et al. (2010). Giannone et al. (2015) observed that, from a Bayesian perspective, this choice of parameters is theoretically the same as that for other parameters in the model. They demonstrated that the model can be treated as a hierarchical one, with the marginal likelihood (ML) of the data, given the prior parameters, available in closed form for VAR models with conjugate priors. Estimating these hyperparameters by maximizing the ML constitutes an empirical Bayes method, which has a clear frequentist interpretation (Giannone et al., 2015).

3.6. Impulse Response Function and Variance Decomposition

To analyse the dynamic interrelationships between variables, we employed impulse response functions (IRFs) and forecast error variance decomposition (FEVD) within our Bayesian vector autoregression (BVAR) framework. These techniques are widely used in VAR analysis to investigate the propagation of economic shocks and their impact on the system (Wang et al., 2021).

IRFs and variance decompositions work together to investigate these dynamic relationships. While causality tests within a VAR framework are limited to the sample period (Wang et al., 2021), variance decomposition allows us to explore potential out-of-sample causality. This technique helps us understand how shocks to one variable might influence others even after the sample period has ended.

An impulse response in a VAR system refers to a shock applied to an individual variable’s error term (Zhou et al., 2018). IRFs essentially measure how the dependent variables within the VAR respond to these shocks. Researchers often apply a unit shock (a one-time change) to each variable and observe its impact on the entire system (Zhou et al., 2018). This analysis plays a crucial role in modern macroeconomics, particularly with structural or semi-structural VAR models (Ivanov & Kilian, 2005). By relying on assumptions about the short-run and long-run responses to shocks, these models can be used to investigate the temporal effects of shocks on various variables within the system (L. Liu, 2019). In simpler terms, IRFs allow us to trace how variables react to both their internal shocks and shocks originating from other variables in the VAR model.

While traditional VAR models often rely on ad hoc identification schemes, Bayesian VAR models offer a more flexible and data-driven approach to identifying structural shocks. By incorporating prior information and hierarchical priors, these models can mitigate the challenges associated with structural identification and provide more robust and reliable inference.

3.7. Diagnostic Test

The BVAR framework offers various standard methods for objects of type “bvar” and derivatives, which support both preliminary assessments and detailed analysis. These methods include print, plot, summary, predict, and the irf function, as outlined by Kuschnig and Vashold (2021).

To assess model stability, we first generate an overview of the estimation using print. Next, we employ plot to evaluate the convergence of the MCMC algorithm, which is critical for model stability. The default plot method produces trace and density plots for the maximum likelihood estimates and the hyperparameters, excluding burnt draws and with parameter boundaries marked as dashed grey lines. The plot can be customized using the vars argument to focus on specific hyperparameters or autoregressive coefficients, and the Var response and Var impulse arguments provide alternative ways to retrieve autoregressive coefficients. Additionally, the type of argument allows for the selection of specific plot types.

As a supplementary step in convergence diagnostics, we also fit and analyse residual values to further ensure proper convergence.

4. Empirical Results

This section presents the empirical results of the study. We begin by examining the preliminary data analysis, including descriptive statistics and unit root test. Subsequently, we estimate and evaluate the BVAR model under various specifications to ensure robustness. The results are then discussed in detail, shedding light on the dynamic interrelationships among the variables and their implications for economic forecasting and policymaking.

4.1. Descriptive Statistics and Random Forest Analysis (RF)

Descriptive statistics are used in data analysis to help econometricians understand the fundamental characteristics of the data in their studies. These statistics provide meaningful summaries, revealing potential patterns within the data. Based on the Jarque–Bera test results presented in

Table 3. Descriptive statistics.

	FCI	CO2M	ASCDD	REC	LIR	UN
Mean	109.4245	361658.1	3.587648	12.17438	13.10747	21.68844
Median	107.4504	386590.7	3.421272	9.895000	11.50000	20.52650
Maximum	121.0472	448298.1	5.286674	18.59000	21.79167	28.84000
Minimum	92.19162	238780.6	2.221507	7.720000	7.041667	19.34200
Std. Dev.	8.324919	72326.47	0.887311	4.088553	4.243861	2.550015
Skewness	−0.100597	−0.463658	0.254646	0.501333	0.487856	1.457476
Kurtosis	1.806274	1.647791	1.971339	1.542004	1.981075	4.498750
Jarque–Bera	1.953948	3.584515	1.756696	4.174789	2.653629	14.32426
Probability	0.376448	0.166584	0.415469	0.124010	0.265321	0.000775
Sum	3501.583	11573060	114.8047	389.5800	419.4392	694.0300
Sum Sq. Dev.	2148.432	1.62 × 1011	24.40695	518.2043	558.3211	201.5798
Observations	32	32	32	32	32	32

, the variable UN is not normally distributed, while CO2M and REC appear to be normally distributed. In such cases, transformations like logarithmic or square root functions can be employed to achieve normality.

The variables in this study were transformed following a procedure similar to that from Kuschnig and Vashold (2021), which includes various transformations and, crucially, tests for stationarity. Stationarity considerations and model specification are vital prerequisites for Bayesian vector autoregression (BVAR) model estimation. Therefore, this study employs a machine learning (ML) approach, specifically Breiman’s (2001) random forest (RF) algorithm, to pinpoint the variable that most significantly contributes to financial instability. The outcome of this ML analysis then dictates the ordering of the variables within the BVAR model. To enhance the RF model’s prediction capability, the number of trees was increased to 10,000.

Based on these random forest (RF) results, ASCDD was identified as the most significant contributor to financial stability, followed by RES, LIR, and UN. Consequently, the variables were ordered accordingly for the BVAR model. Prior to model development, the data, comprising six variables (FCI, CO2M, ASCDD, RES, LIR, and UN), was structured into a rectangular numeric matrix, ensuring the absence of missing values. While most variables are expressed as rates, CO2M is measured in kilotons (kt) and FCI is an index. Following the methodology of Kuschnig and Vashold (2021), CO2M underwent a logarithmic transformation to facilitate the application of dummy priors, a transformation supported by McCracken and Ng (2016). In this transformation, CO2M was assigned code 4, while the remaining variables received code 1. It is important to note that the BVAR framework can accommodate variables with differing integration orders (I(0) and I(1)) through appropriate transformations, as detailed in McCracken and Ng (2016).

This table presents the key characteristics of the study’s six variables (FCI, CO2M, ASCDD, REC, LIR, UN) over 32 observations. The mean and median indicate central tendencies, with differences suggesting skewness. The maximum and minimum show the range, while the standard deviation quantifies data spread. Skewness reveals asymmetry (UN is notably positively skewed), and kurtosis indicates the shape of the tails (UN shows heavier tails). The Jarque–Bera test suggests that FCI, CO2M, ASCDD, REC, and LIR do not significantly deviate from a normal distribution, while unemployment (UN) is not normally distributed. The table also provides the sum, sum of squared deviations, and the number of observations for each variable.

4.2. Unit Root Tests

To determine the order of differencing required for the time series data, the study employed both Augmented Dickey–Fuller (DF) and Phillips–Perron (PP) unit root tests. The results, detailed in

Table 4. Unit root test.

South Africa 1991–2022: ADF Test
Variables	Lev	Prob	1st	Prob	Intr
FCI	−4.681018	0.0040	−5.190807	0.0015	I(0)
CO2M	−0.879152	0.9459	−7.471624	0.0000	I(1)
ASCDD	−3.047866	0.1367	−4.730924	0.0007	I(1)
RES	0.510173	0.9988	−8.156026	0.0000	I(1)
LIR	−3.972418	0.0210	−5.340142	0.0009	I(1)
UN	0.176748	0.9966	−7.124334	0.0000	I(1)
South Africa 1991–2022: PP Test
Variables	Lev	Prob	1st	Prob	Intr
FCI	−5.481249	0.0005	−11.81431	0.0000	I(0)
CO2M	−0.413672	0.9824	−6.870246	0.0000	I(1)
ASCDD	−2.217267	0.4640	−4.985761	0.0003	I(1)
RES	−0.328432	0.9860	−9.261465	0.0000	I(1)
LIR	−2.146104	0.5013	−5.131408	0.0002	I(1)
UN	−0.477767	0.9793	−11.77755	0.0000	I(1)

, indicate that only FCI is stationary at levels. All other variables were found to be non-stationary at levels but achieved stationarity after first differencing. Consequently, to prevent the transformation of FCI, which is stationary at levels, it was assigned code 1.

The estimation of models with variables of mixed orders of integration has been well-documented in the Kuschnig and Vashold (2021) article. For instance, Kuschnig and Vashold (2021) discuss methods for dealing with such mixed integration orders within a Bayesian vector autoregression (BVAR) framework.

For the other variables, code 2 was applied to transform them into first differences. This approach resulted in the selection of five log differences for CO2M, one for FCI, and two for the other variables. The lag order was set to one, though it may not fully capture the influence of recent events on South Africa’s financial stability. Therefore, the study utilized a BVAR model with a first-order lag to analyse the dynamic interactions between financial stability, carbon emissions, and control variables over time.

4.3. The Prior Setup and Configuration of the Model

TML-VAR models face two significant limitations, particularly when applied to data from middle- and low-income countries: issues with data quality and over-parameterization due to the inclusion of too many lags. To mitigate the latter issue, BVAR models rely heavily on prior selection.

To mitigate these issues, we follow the approach of Kuschnig and Vashold (2021) in specifying priors and configuring our model. This involves adjusting the Minnesota prior, where the hyperparameter λ is assigned a Gamma hyperprior and its Gaussian proposal distribution is bounded. The hierarchical treatment of α is initially omitted. Additionally, Ψ is automatically set to the square root of the innovation variance after fitting AR(p) models to each variable.

In addition, we incorporate a System of Equations (SoE) prior by constructing three dummy observation priors. Similar to λ, the hyperpriors of their key parameters are assigned Gamma distributions. This BVAR configuration, which includes setting up the model’s priors and the Metropolis–Hastings algorithm, is implemented by providing the character vector c (Lambda; Soc; Sur).

4.4. Model Estimated Threshold of the BVAR Model

As detailed in the methodology section, the BVAR model from Zungu and Greyling (2023) requires specific data preparations and transformations, including defining the lag order p. The BVAR function also needs a tailored setup to accommodate these parameters. We configure the initial burn-in iterations to 1,500,000 and set the number of draws to 500,000. In line with Kuschnig and Vashold (2021), we activate the verbose parameter to display a progress bar during the MCMC process. The resulting posterior mode estimates are shown in

Table 5. Posterior marginal likelihood.

Optimisation concluded.

Posterior marginal likelihood: −420.406

Hyperparameters: lambda = 0.27491

|===================================================| 100%

Finished MCMC after 47.33 min

.

The BVAR function produces a BVAR object containing various outputs, such as hierarchically treated hyperparameters, the variance–covariance matrix (VCOV), and posterior draws of VAR coefficients. The object also stores marginal likelihood values for each draw, specified prior settings, and initial values of prior hyperparameters derived from optimization, automatic settings, and the original BVAR function call (see

Table A1. BVAR results summary.

BVAR Results
Bayesian VAR consisting of 30 observations, 6 variables and 1 lags. Time spent calculating: 53.75 min Hyperparameters: lambda Hyperparameter values after optimisation: 0.27491 Iterations (burnt/thinning): 1,500,000 (500,000/1) Accepted draws (rate): 394,593 (0.395)
Numeric array (dimensions 7, 6) of coefficient values from a BVAR. Median values:         FCI  COM2  ASCDD  REC  LIR  UN constant 31.862 1.139 −1.245 −3.547 1.447 8.360 FCI-lag1 0.712 0.005 0.012 −0.006 −0.043 −0.556 COM2-lag1 −0.252 −0.040 −0.115 −0.036 0.047 0.085 ASCDD-lag1 −1.341 −0.587 0.151 0.090 0.186 0.007 REC-lag1 0.172 −2.734 0.020 0.233 −0.485 0.042 LIR-lag1 0.635 −0.093 −0.023 0.117 0.112 −0.062 UN-lag1 −0.761 −1.198 0.047 0.187 0.011 −0.162
Numeric array (dimensions 6, 6) of variance–covariance values from a BVAR. Median values:
	FCI	COM2	ASCDD	REC	LIR	UN
FCI	27.006	7.142	−1.546	−0.062	−1.648	1.628
COM2	7.142	315.906	−0.210	1.193	1.709	0.057
ASCDD	−1.546	−0.210	0.126	0.062	0.928	0.057
REC	−0.062	−0.031	0.060	0.252	0.005	0.070
LIR	−1.648	1.793	0.253	0.152	2.577	0.044
UN	1.628	0.700	−0.057	0.070	−0.044	0.554
Log-Likelihood: −274.7856

).

4.4.1. Result of the Convergence of Markov Chain Monte Carlo in a BVAR Model

This section assesses the stability of our estimation method, focusing on the convergence of the Markov Chain Monte Carlo (MCMC) algorithm.

Table 6. Summary of the BVAR model.

Bayesian VAR consisting of 30 observations, 6 variables, and 1 lags.

Time spent calculating: 47.33 min

Hyperparameters: lambda

Hyperparameter values after optimisation: 0.27491

Iterations (burnt/thinning): 1,500,000 (500,000/1)

Accepted draws (rate): 394,593 (0.395)

provides a summary of the Bayesian vector autoregression (BVAR) model, showing that the coefficient of lambda for South Africa (SA) is 0.27491. The MCMC process included 1,500,000 iterations, with 500,000 burn-in iterations and a thinning factor of 1, resulting in an acceptance rate of approximately 39.5% of the draws.

The var response and var impulse arguments streamline the extraction of autoregressive coefficients. We use the type of argument to select specific plot types, such as density and trace plots, for the hyperparameters. Figure 1 provides visual evidence of the convergence of key hyperparameters in the BVAR model for SA. Appendix A Figure A1 further presents the trace and density plots for all hierarchically treated hyperparameters and the Maximum Likelihood (ML), confirming the effective exploration of the posterior distribution with no significant outliers. The MCMC chain effectively explores the posterior distribution, with no significant outliers.

4.4.2. Impulse Responses of the Bayesian VAR

This study explores the relationship between climate change and financial stability in South Africa, focusing on whether climate change’s impact on financial stability is long-lasting from 1991 to 2022. To achieve this, a Bayesian vector autoregression (BVAR) model with a hierarchical prior selection method is employed. Figure 2 depicts the dynamic responses of financial stability to factors such as CO2M, ASCDD, REC, LIR, and UN, with tightened hierarchical prior distributions for enhanced precision. The shaded areas in the figure represent the 16% and 84% credible intervals. These intervals indicate the range where the true effect is estimated to lie with a certain probability. Specifically, there is an estimated 68% probability that the true value falls between these percentiles, showing the uncertainty around the estimated responses.

According to Figure 2 (Plot I), the analysis suggests that carbon emissions (CO2M) significantly heighten financial stability risk in South Africa. A 1% standard deviation shock to CO2M is estimated to result in a maximum impact of 0.6 after two years, before diminishing, returning to the steady state, and dissipating after three years. This aligns with the Financial Instability Hypothesis (Minsky, 1992) and findings from other studies (Dafermos et al., 2018; Wu et al., 2023; Chabot & Bertrand, 2023; Nieto, 2019), which indicate that climate change is a significant source of financial risk. The Financial Instability Hypothesis explains how economic instability can amplify these vulnerabilities, potentially causing financial crises.

In South Africa, a 1% standard deviation shock to the average annual growth rate of the South African Development Community (ASCDD) also contributes to a decrease in financial stability. This shock is estimated to have a maximum impact of 0.39 after two years, which gradually converges to the steady state over 11 years. These findings are consistent with research on climate transition risk and financial stability (Lee et al., 2024).

The study further reveals that financial stability decreases following a 1% shock to the real effective exchange rate (REC), reaching a maximum impact of 0.10 after two years and dissipating after 2.3 years. Investments in renewable energy have an asymmetric impact, reducing interbank connectivity (increasing bank failure probability), while higher energy taxes have the opposite effect (Safarzyńska & van den Bergh, 2017).

Surprisingly, financial stability increases in response to a 1% shock to the lending interest rate (LIR). This shock is estimated to have a maximum impact of 0.83 after two years, which gradually returns to the steady state over 12 years. This supports previous research (Mwangi, 2014).

Moreover, the analysis demonstrates that a 1% standard deviation shock to unemployment (UN) leads to a decrease in financial fragility. This shock is estimated to have a maximum impact of 0.10 after two years, converging after 2 years and persisting for more than 12 years. This aligns with the results reported by Friedman (2013).

Turning to the impact of policy shocks on carbon emissions, Figure 2 (Plot II) reveals that a 1% standard deviation shock to the lending interest rate (LIR) policy has the most pronounced effect on reducing South Africa’s carbon emissions. This impact is estimated to peak at 0.37 two years after the shock, then rapidly diminish and return to the original level within three years. Notably, this effect is statistically significant.

Similarly, the analysis shows that a 1% shock to the average annual growth rate of the South African Development Community (ASCDD) policy also contributes to a reduction in carbon emissions. This shock is estimated to reach a maximum impact of 0.58 two years after the shock, also dissipating relatively quickly within three years.

The study further indicates that a 1% shock to the real effective exchange rate (REC) leads to a decline in carbon emissions, with the impact peaking at 1.5 after two years. This effect is initially statistically insignificant, becoming significant after three years and diminishing over five years.

Counterintuitively, the analysis reveals that a 1% shock to financial stability (FCI) also results in a reduction in carbon emissions. This effect is estimated to reach a maximum of 1.5 immediately after the shock. However, this effect is not statistically significant and it rapidly converges and disappears within two years.

Finally, the study demonstrates that a 1% shock to the unemployment rate (UN) leads to a decrease in carbon emissions. This impact is estimated to peak at 0.7 two years after the shock and persist for over five years.

Figure 2 (Plot II) illustrates that the lending interest rate (LIR) policy has the most substantial impact on lowering South Africa’s carbon emissions. Specifically, a 1% standard deviation shock to LIR leads to a peak reduction in emissions of 0.37, observed two years after the shock, with this effect rapidly returning to the baseline level within three years. This reduction is statistically significant.

The average annual growth rate of the South African Development Community (ASCDD) policy also influences carbon emissions. A 1% standard deviation shock to ASCDD results in a maximum decrease in emissions of 0.58, again occurring two years post-shock and dissipating quickly within three years.

The real effective exchange rate (REC) also affects carbon emissions. A 1% standard deviation shock to the REC leads to a reduction in emissions, peaking at 1.5 two years after the shock. However, this reduction is not statistically significant in the initial period; it only becomes significant after three years, and its influence diminishes over the subsequent five years.

In a somewhat unexpected finding, financial stability (FCI) also appears to have a reducing effect on carbon emissions. A 1% standard deviation shock to FCI results in a maximum emissions reduction of 1.5, which is observed immediately following the shock. Despite this immediate impact, the effect is not statistically significant and diminishes rapidly, disappearing within two years.

Lastly, the unemployment rate (UN) also plays a role in the level of carbon emissions. A 1% standard deviation shock to UN leads to a decrease in emissions, reaching a peak reduction of 0.7 two years after the shock. This effect on emissions persists for more than five years.

4.4.3. Forecast Error Variance Decomposition (FEVD)

Forecast error variance decomposition (FEVD) is a tool in econometrics used to assess the relative importance of different shocks in explaining the forecast error variance of a specific variable. It helps identify how much uncertainty in forecasting a variable is due to shocks to itself and how much is attributable to shocks from other variables in the model.

In

Table 7. Forecast error variance decomposition.

S’FCI
Period	FCI	CO2M	ASCDD	REC	LIR	UN
[1,]	100.0000	0.000000	0.000000	0.000000	0.000000	0.000000
[2,]	99.43515	0.265124	0.000273	0.009889	0.181493	0.108066
[3,]	98.52625	0.590214	0.000313	0.030422	0.615240	0.237560
[4,]	97.46115	0.864520	0.000247	0.055926	1.278548	0.339612
[5,]	96.31640	1.068595	0.000373	0.080893	2.128511	0.405228
[6,]	95.12921	1.212348	0.000888	0.101570	3.116539	0.439446
[7,]	93.92351	1.311424	0.001827	0.116169	4.196523	0.450543
[8,]	92.71796	1.379645	0.003108	0.124458	5.328676	0.446147
[9,]	91.52784	1.427431	0.004587	0.127192	6.480750	0.432204
[10,]	90.36537	1.462106	0.006105	0.125612	7.627823	0.412983

, each row corresponds to a forecast horizon (e.g., one period ahead, two periods ahead) and each column represents a variable in the Bayesian vector autoregression (BVAR) model. The variables considered are FCI, CO2M, ASCDD, REC, LIR, and UN. The numbers in the table represent the percentage of the forecast error variance of the variable in the row that is explained by shocks to the variable in the column.

At period one, 100% of the one-period-ahead forecast error variance of FCI is attributed to its own shocks, as expected. However, as the forecast horizon extends, the influence of external variables starts to increase. By period two, small contributions from CO2M (0.2651%), ASCDD (0.00027%), REC (0.00989%), LIR (0.1815%), and UN (0.1081%) are observed, indicating that shocks to these variables begin to affect the uncertainty in forecasting FCI.

At the three-period horizon, FCI’s own shocks still dominate, explaining 98.5263% of its forecast error variance. Yet, external variables such as CO2M (0.5902%), LIR (0.6152%), and UN (0.2376%) begin to play a more noticeable role in the uncertainty of forecasting FCI.

By period four, FCI’s own shocks account for 97.4612%, with external variables, particularly CO2M (0.8645%), LIR (1.2785%), and UN (0.3396%), contributing more to the forecast error variance. This trend continues at the five-period horizon, where FCI’s own shocks account for 96.3164%, while CO2M (1.0686%) and LIR (2.1285%) become more influential.

The forecast error variance decomposition at the ten-period horizon reveals that shocks to the Financial Conditions Index (FCI) itself explain the majority of its variance (90.37%). However, the influence of other variables increases over this longer term. Notably, the lending interest rate (LIR) accounts for a substantial portion (7.63%) of FCI’s variance, followed by carbon emissions (CO2M) at 1.46% and unemployment (UN) at 0.41%. Adjusted savings (ASCDD) and renewable energy consumption (REC) have a relatively minor impact on FCI’s forecast error variance at this horizon, accounting for 0.01% and 0.13%, respectively. These findings suggest that longer-term forecasts of financial stability in South Africa are increasingly sensitive to changes in lending interest rates, carbon emissions, and unemployment.

Overall, the FEVD results highlight a shift in the drivers of FCI’s forecast error variance over time. While FCI’s own shocks dominate in the short term, their influence diminishes in the long term, with external factors like CO2M, LIR, and UN becoming more significant in explaining forecast uncertainty. This suggests that the long-term behaviour of FCI is increasingly influenced by these external variables.

4.5. Diagnostics Test

4.5.1. Density Plot

The density plot of the autoregressive coefficient (AR (1)) for FCI (Figure 3) displays a well-shaped distribution, indicating reliable estimation. The coefficient’s mean and median range from 0.5 to 1.0, suggesting moderate to strong positive autocorrelation, meaning FCI values are positively correlated with their past values and show persistence. The peak around 3.0 indicates the most likely value for the coefficient, with a rapid decrease in probability for extreme values.

The majority of coefficient values fall within the 0.0 to 1.5 range, suggesting a relatively narrow uncertainty interval. The plot shows no visible multimodality, confirming a well-behaved posterior distribution. These results imply that FCI shocks may have lasting effects with moderate to strong persistence.

Overall, the diagnostic results support the appropriateness of the Bayesian vector autoregression (BVAR) model. However, further checks, such as convergence tests and residual autocorrelation assessment, are necessary to ensure model robustness. Additionally, residual plots can provide further insights into the relationships between FCI and other variables.

4.5.2. Residuals Plot

The performance of the Bayesian vector autoregression (BVAR) model is comprehensively evaluated through both its fitted values and residual analysis. The model’s capacity to replicate the observed data is evident from the close alignment between the fitted values and actual series, as presented in Appendix A, Figure A2.

Complementary to this, Figure 4 presents the residual plot analysis, offering significant insights into the performance of the Bayesian vector autoregression (BVAR) model. Specifically,

Absence of Systematic Patterns

The residuals exhibit no discernible trends, cycles, or structures, suggesting that the BVAR model has effectively captured the underlying relationships between Financial Conditions Index (FCI) and carbon emissions (CO2M). This finding supports the assumption of no systematic patterns in the residuals.

Randomness and White Noise Errors

The residuals appear randomly distributed, consistent with the assumption of white noise errors. This indicates that the BVAR model has accounted for the underlying dynamics between FCI and CO2M.

Homoscedasticity and Absence of Clustering

The variance of residuals remains constant over time, supporting the assumption of homoscedasticity. Furthermore, the residuals do not exhibit clustering, indicating no evidence of autocorrelation or serial correlation.

Collectively, these findings suggest that the BVAR model is adequately specified, capturing the complex relationships between FCI and CO2M. The relationships between FCI and CO2M are well-represented by the model and the residuals are consistent with white noise errors, indicating no significant omitted variables or model misspecification.

4.6. Discussion of the Bayesian VAR Results

The empirical analysis, leveraging a Bayesian vector autoregression (BVAR) model, reveals significant and complex relationships between climate change and financial stability in South Africa. Our findings indicate that carbon emissions (CO2M) exert a lasting and detrimental impact on financial stability. This robust result empirically supports Hyman Minsky’s Financial Instability Hypothesis, suggesting that prolonged environmental stressors can induce systemic vulnerabilities within the financial system. It also aligns with recent international studies (Dafermos et al., 2018; Wu et al., 2023; Chabot & Bertrand, 2023) that identify climate change as a substantial and growing risk to global financial stability.

Specifically, following a 1% standard deviation shock, the impulse response functions reveal that carbon dioxide-induced damage to adjusted savings, renewable energy consumption, and unemployment are significant factors contributing to reduced financial stability in South Africa. These results resonate with prior research on climate change’s financial repercussions. For instance, our findings on the overall impact of climate change on financial stability align with Ayele and Fisseha (2024)’s work in sub-Saharan Africa, emphasizing a regional vulnerability. The identified role of transition risks, particularly concerning renewable energy, finds parallels in Lee et al. (2024)’s analysis of climate transition risks in France, although the specific mechanisms may differ given South Africa’s unique energy landscape. The interplay of investment and environmental factors also touches upon themes explored by Safarzyńska and van den Bergh (2017) regarding renewable energy investments and Friedman (2013) on broader economic impacts.

Crucially, our findings contradict some of the existing empirical literature, particularly the second and third ranges often observed in reviews. The second range suggests both positive and negative impacts of climate factors on financial stability (e.g., Z. Liu et al., 2021; Brunetti et al., 2022; Meng et al., 2023), while the third range posits a significant but highly context-dependent relationship (Pointner & Ritzberger-Grünwald, 2019; Conlona et al., 2022; Pagnottoni et al., 2022; Curcio et al., 2023). Our BVAR analysis, however, consistently points towards a predominantly negative and significant relationship in the South African context. This divergence is likely attributable to several key factors specific to South Africa: its high dependence on carbon-intensive industries (e.g., mining, coal-fired power), its vulnerability to extreme weather events that directly translate into physical risks for infrastructure and economic sectors, and a financial system that is still developing robust mechanisms to price and manage climate-related risks. Unlike more diversified or climate-resilient economies, these factors may amplify the negative transmission channels, leading to a more consistent detrimental impact on financial stability.

Economically, the pathways through which climate change impacts financial stability are multifaceted. High carbon emissions, largely driven by human activity and industrialization, generate significant physical risks. These risks manifest as supply chain disruptions, damage to physical assets, and reduced productivity, directly translating into impaired asset values for financial institutions, increased loan defaults, and heightened credit risk across various sectors. The resulting reduced national savings, increased healthcare and environmental expenditures, and decreased economic growth directly undermine the foundational health of the economy, placing additional strain on government finances and ultimately eroding the resilience of the financial system.

Furthermore, the damage to adjusted savings stemming from carbon dioxide emissions directly compromises financial stability. This occurs by diverting capital away from productive investments in human capital and critical infrastructure, funnelling it instead towards remediation of environmental damage. This reallocation of resources inhibits long-term economic growth potential and increases the cost of capital for businesses, which in turn can lead to higher non-performing loans and lower profitability for financial institutions, thus weakening their balance sheets.

Surprisingly, renewable energy consumption in our model negatively impacts financial stability. This counter-intuitive finding warrants deeper economic consideration. While renewable energy is crucial for climate mitigation, its transition costs can be substantial, especially for a developing economy like South Africa. High upfront costs for renewable infrastructure, coupled with potential market disruptions as traditional energy sources are phased out, can lead to stranded assets in carbon-intensive sectors. Additionally, increased energy price volatility during the transition period, driven by reliance on intermittent sources or evolving policy frameworks, can introduce revenue uncertainty for businesses and pose liquidity challenges for financial institutions exposed to these sectors. However, it is important to acknowledge that this result may also be influenced by model specification, data quality, or omitted variables specific to the nascent stage of South Africa’s renewable energy transition, warranting further research.

Finally, unemployment plays an imperative role in weakening financial stability through various channels. High unemployment directly reduces consumer spending and overall economic activity, leading to lower corporate revenues and increased credit defaults for banks. This economic slowdown exacerbates financial fragility. Moreover, unemployment significantly heightens climate-related financial risks by reducing household resilience to climate shocks; unemployed individuals or households have fewer resources to cope with climate-induced damages or disruptions. This, in turn, increases the burden on government support programs, potentially straining public finances and indirectly impacting sovereign creditworthiness, which is a key pillar of financial stability.

Our results show that South Africa’s financial stability is highly vulnerable to climate-related risks. The identified dynamic relationships are complex and reveal bi-directional links between environmental degradation, economic disruption, and the stability of the financial system. These linkages manifest through increased credit risk, asset devaluation, increased market volatility, regulatory changes imposing costs on financial institutions, and broader social instability, all of which can severely undermine the resilience and integrity of the financial sector. Our findings thus provide empirical evidence for policymakers in South Africa to proactively manage climate-related financial risks.

5. Robustness Analysis

To ensure the stability and reliability of our findings, a robust analysis was carried out by augmenting the estimated BVAR model with the GDP growth rate variable. This variable was included under the assumption that it represents a highly exogenous macroeconomic factor, given its broad influence on the South African economy and its potential to independently impact financial stability.

The BVAR was re-estimated using the same lag length and prior specifications as the baseline model. The primary objective was to analyse whether the inclusion of GDP growth rate would lead to relevant changes in the dynamics of the impulse response functions (IRFs). These IRFs capture the dynamic effects of climate change shocks on financial stability in South Africa.

Upon comparing the IRFs from the augmented BVAR (Figure 5, (a)plot I) with those from the baseline BVAR (Figure 2, specifically (a) Plot I), it was observed that the inclusion of the GDP growth rate variable did not result in significant changes in the estimated dynamic responses. This consistency suggests that the relationships identified between climate change and financial stability are not materially influenced by the broader macroeconomic performance, as captured by GDP growth.

Accordingly, the BVAR model demonstrates robustness, and the results presented in the current study remain relevant for understanding the dynamic relationship between climate change and financial stability in South Africa.

6. Conclusions and Policy Recommendations

This study demonstrates the significant impact of climate change on financial stability in South Africa. Utilizing a Bayesian vector autoregression (BVAR) model, our findings reveal that elevated carbon emissions, the economic damage from these emissions, the dynamics of renewable energy consumption, and unemployment levels act as critical factors undermining financial stability. Impulse response functions indicate that shocks to these variables exert lasting effects, with carbon emissions showing the most pronounced and persistent influence.

Furthermore, our analysis highlights the complex interplay between renewable energy consumption and financial stability. While the strategic scaling of renewable energy is vital for long-term sustainability, policymakers must carefully manage the potential short-term financial risks associated with the energy transition through well-designed investment strategies and supportive financial policies.

Based on these findings, we recommend the following policy actions for South Africa:

Aggressively Reduce Carbon Emissions: Implement stringent carbon pricing mechanisms, enforce stricter emission standards across all sectors, and incentivize the adoption of low-carbon technologies.

Strategic Investment in Renewable Energy: Develop clear and consistent long-term policies that attract investment in diverse renewable energy sources, coupled with infrastructure development and financial incentives to ensure a smooth transition.

Prioritize Employment and Economic Diversification: Foster inclusive economic growth and implement targeted employment programs to build resilience against climate-related economic shocks. Further economic diversification into new industries such as new technology-based production like artificial intelligence (AI) and robotics will increase resilience.

Integrate Climate Risks into Financial Oversight: Financial regulators should develop and actively incorporate climate-related risks into macro-prudential frameworks and stress-testing exercises to safeguard financial system stability. For instance, climate risk considerations should be included by the Intergovernmental Fintech Working Group (IFWG) in its regulatory sandbox testing of various new technologies.

Develop Climate-Aware Monetary Policy: The South African Reserve Bank should explore and integrate climate change considerations into its monetary policy decisions, acknowledging the potential macroeconomic impacts of climate risks.

This research provides crucial empirical evidence regarding the climate–financial stability nexus in South Africa. Future research should delve deeper into the specific transmission mechanisms of these impacts and explore sector-specific policy interventions for greater precision. The study’s focus on South Africa, the 1991–2022 timeframe, and data limitations suggests avenues for broader comparative studies and the incorporation of more granular climate risk data as it becomes available.