Purpose: In this paper, we study the potential of using Artificial Neural Network (ANN) models to predict currency crises in emerging markets, with a specific focus on the South African economy. South Africa’s rand is one of the most volatile currencies in the world and is prone to crises.

Methodology: We built two ANN models, where Model 1 uses ten economic indicators and Model 2 uses four. These models were assessed for statistical significance (using probit analysis), and their performance in predicting major South African currency crises (e.g., 1998, 2001, and 2008) was tested with both in-sample and out-of-sample data.

Results: The first model was much more accurate than Model 2 in predicting early warning signs nearly two years before the currency crises occurred. Model 1’s higher accuracy is attributed to its inclusion of a greater number of economic variables. Both models occasionally produced false positives, though overall, they were very accurate in predicting crises.

Originality: Our paper highlights the importance of ANNs in capturing nonlinear patterns in economic data, demonstrating their strength as early warning tools for financial crises. We recommend that ANN methods continue to be researched and advanced to further reduce false positives and improve predictive performance.

Currency crises, Early Warning System, Artificial Neural Networks, South Africa, Emerging markets.

Currency crises are not only a problem of the past, but also of the present and future. Since the fall of the Bretton Woods system in 1973, emerging and developing markets have experienced an increased frequency of crises, which continue to pose a threat today (Bernoth & Herwartz, 2021). Financial crises have affected numerous countries in the past and are likely to happen again in the future, as predicted in a prescient study by Bordo, Eichengreen, Klingebiel, Martinez‐Peria and Rose (2001). Notable currency crises include the Latin American crises of the 1970s/80s, crises in Europe from 1992 to 1993 counteracted by the European monetary exchange rate mechanism (ERM), the Mexican peso crisis of 1994 to 1995 (also known as the tequila crisis), the East Asian crisis of 1997 to 1998 which affected Indonesia, Korea, Malaysia, Philippines, and Thailand, Brazil in 1999, Argentina in 2002, the Zimbabwean currency crisis from 1990 to 2009, the USA in 2007, the GFC, and the 2015 Greek currency crisis (International Monetary Fund Staff Report, 2009). Given the great impact of sovereign debt and currency crises on economic activity, the study of international financial crisis events, particularly their forecasting, has garnered substantial attention from economists in the last twenty years (Alaminos Aguilera et al., 2021).

South Africa, as the economic powerhouse of Africa, is an interesting case to examine. The country operates a floating exchange rate system, which makes it, in theory, less vulnerable to currency crises (Juhro et al., 2022; Guei & Choga, 2022). This is because market expectations for adjustments prevent the buildup of excessive pressure in its foreign exchange markets (Ghosh et al., 2015; Georgiadis & Zhu, 2021). South Africa’s foreign currency market is known for its volatility and frequent bouts of turbulence (Laubscher, 2020). Instances of currency crises were documented in 1996, 1998, 2001, and 2008. The 2007/2008 GFC was the most severe economic downturn the globe had encountered since the Great Depression of the 1930s (Mirzaei, 2023). The GFC had a significant negative impact on South Africa’s economy, revealing the country’s fragility, as mineral exports declined leading to a growing trade imbalances (Bhundia & Ricci, 2005; Raudzingana, 2020). Furthermore, the economy experienced a recession for the first time in 17 years. Despite the South African Reserve Bank’s confidence in the economy’s strong foundations, its macroeconomic predictions had to be significantly adjusted downwards (South African Reserve Bank, 2012; Beyers et al., 2024).

South Africa as an emerging market economy (EME) is gaining global importance due to its strategic partnership with the BRICS+ association. Since EMEs have historically been more susceptible to crises (Alonso-Alvarez & Molina, 2023), it is essential that fiscal and monetary policymakers understand the extent, to which their financial systems are vulnerable to disturbances that can easily spill over into the key partner economies, and be ready to respond appropriately. South Africa suffers from frequent strikes and violent protests (Kali, 2023); severe currency crises and economic hardships could trigger even more populist events. All this highlights the importance of Early Warning Systems (EWSs). We are living in an era of increased risks of future pandemics, climate change induced global boiling, and global violent conflict (Goniewicz et al., 2023). These challenges further raise the risk of global financial crises and contagion.

The South African economy needs an alternative early warning system (EWS) model, which is to be developed and evaluated. This model should accurately detect the early signs of crisis events before they occur, allowing authorities to implement optimal policies to prevent or mitigate their impact. One tool that has garnered significant interest from the International Monetary Fund (IMF), policymakers, and academics is the Artificial Neural Networks (ANN) approach (Gupta & Kumar, 2022a; Ataman & Kahraman, 2022). ANNs probability prediction is a type of dynamic filtering in which past values of one or more time series are used to predict future values (Pokou et al., 2024). ANN models are commonly applied in management fields such as sales and demand forecasting, logistics and supply chain (Khan et al., 2020; Huang et al., 2021; Elalem et al., 2023). However, there are increasing studies of ANN’s use to forecast economic growth, downturns, and financial crises (Tölö, 2020; Longo et al., 2022; Liu et al., 2022; Bluwstein et al., 2023; Barthélémy et al., 2024; Raj et al., 2024). Most of the recent literature in this field does not concern the developing world, let alone sub-Saharan Africa. We assert that using ANN as described in this paper is superior to other methodologies such as the conventional machine learning models, shallow neural networks or traditional econometric models, which cannot properly analyze the challenging multidimensional and highly volatile nonlinear financial market data (Heo et al., 2020; HongXing et al., 2022). To the best of our knowledge, this study is one of the first in South Africa to use the ANN methodology to evaluate its ability to forecast business cycles.

The ANN models used a backpropagation learning algorithm to determine model performance (Heng et al., 2022). For training¹ the network, this study considered two periods, an in-sample period from February 1993 to December 2004 and out of sample period from January 2005 to March 2017. Simulations were conducted on the models with various numbers of iterations to achieve the smallest training error. To this effect one hundred networks were simulated using one to 20 neurons in maximum iterations of 10000 (0.130), 20000 (0.131), 30000 (0.126), 40000 (0.131) and 50000 (0.137). With thirty thousand iterations the smallest error 0.126 was found using 19 hidden neurons. Following this, both models were trained using parameters set in Table 3 below.

The process began when both models stochastically assigned their initial connection weights between neurons and bias neurons in the input layer and the hidden neurons in the hidden layers. Feed-forward was employed to ensure that signals constantly flow in a forward direction from the input layer to the output layer through the hidden layer. Within the concealed neuron, the models aggregate all these signals and convert them into a numerical range of 0 to 1 by means of the logistic sigmoid activation function. The initial weights connecting the hidden neuron and bias neuron to the output neuron in the hidden layer are randomly assigned. Subsequently, the models transmit modified signals from the concealed neuron to the output neuron, employing the logistic sigmoid activation function once more. The learning stage commences by contrasting the output neuron with its designated target. The mean square error is propagated backwards from the output layer to the input layer through the hidden layer using the backpropagation learning technique. This mechanism enables the establishment of suitable connection weights for all neurons. Consequently, the models stop the learning process whenever they reach the minimum error or the maximum number of repetitions.

To establish the most statistically significant input variables, the study employed a popular method using connection weights and the feed-forward back-propagation algorithm (Yadav et al., 2022). This method determines the relative importance of the predictor variables of the model as a function of the neural network synaptic weights, according to the mathematical expression:

$R_{ij}=\sum _{H-1}^k{W}_{ik}·W_{kj}$ (15)

Where R_ij is the relative importance of the variable x_i with respect to the output neuron j, H is the number of neurons in the hidden layer, W_ik is the synaptic connection weight between the input neuron i and the hidden neuron k and W_kj is the synaptic weight between the hidden neuron k and the output neuron j (Nguyen et al., 2021).

Table 4 below presents the contribution of input neurons to the output neurons for Model 1. It shows that the most significant contributor is the ratio of budget deficits to GDP, followed by growth rate of the ratio of domestic credit to GDP and then a change in the international liquidity position. For more details on the other variables, see Table 4 below.

The contributions of the input variables in the reduced model (Model 2) were calculated. As shown in Table 5, the primary contributor to the model’s predictions was the change in the international liquidity position, followed by the inflation differential to the USA and then the growth rate of foreign debt. The change in the domestic interest rate was found to be the least significant input variable in Model 2.

3.1. Predicting the South African Currency Crises

Both models were employed to simulate and evaluate their ability to predict the probability of the South African currency crises of July 1998, December 2001, and October 2008. To analyze the probability predictions, this section was divided into in-sample probability predictions (Figure 3) and out-of-sample probability predictions (Figure 4). The study utilized a 24-month crisis window. This means that any signals detected within the 24 months preceding a crisis were considered as valid indicators of an impending event.

Figure 3. 

The Artificial Neural Networks models: In-Sample Predictions, 1993/02 – 2004/12. Source: Authors own calculations from data set.

Figure 4. 

The Artificial Neural Networks models: Out-of-Sample Predictions, 2005/01 – 2017/03. Source: Authors own calculations from data set.

3.1.1. In-Sample Probability Prediction of Currency Crisis, 1993/02 – 2004/12

Figure 3 presents the in-sample probability predictions for both Model 1 and Model 2 in three separate graphs. While the two models’ predictions were initially plotted together in Figure 3(a), it became evident that their plots were not overlapping, unlike the Probit models. To facilitate analysis, the probability predictions were separated into individual graphs: Figure 3(b) for Model 1 and Figure 3(c) for Model 2. In Figure 3(b), Model 1 accurately predicts the July 1998 crisis. As early as July 1996 (24 months prior), warning signals were emitted, reaching a magnitude of approximately 42%. Although these signals subsided to around 5% in early 1997, they intensified again in the latter part of 1997, reaching approximately 36%. Following a brief decline in early 1998, the signals gradually increased until the occurrence of the crisis in July 1998, with a predicted probability of approximately 59%. Model 2 (Figure 3(c) also correctly identified the July 1998 crisis, with a predicted probability of around 66%. Similar to Model 1, Model 2 emitted warning signals as early as January 1996, although with a higher probability of almost 60%. While these early signals were outside the 24-month crisis window, the model continued to send signals throughout the period leading up to the crisis, albeit with lower magnitude compared to Model 1.

Regarding the second in-sample crisis of December 2001, Model 1 emitted warning signals from as early as the latter part of 1999. The signal strength reached approximately 58% in December 1999, which falls within the 24-month crisis window. After a decline to around 9% in the early months of 2000, the signals intensified again, reaching heights of approximately 70% in late 2000 and early 2001. By December 2001, the predicted probability had risen to 82%. Model 2 also sent signals indicating the December 2001 crisis, but the signal strengths were not as strong as those from Model 1. As December 2001 approached, Model 2’s predicted probability was approximately 62%, which is 20% lower than that of Model 1.

Despite these models’ abilities to correctly point to the two in-sample crises, many signals are also emitted outside the 24-month crisis window. This indicates that both models sent many false alarms.

3.1.2. Out-Of-Sample Prediction Currency Crisis, 2005/01 – 2017/03

Figure 4 below presents the out-of-sample probability predictions for currency crises in the period 2005/01 to 2017/03. In Figure 4 (a) both Model 1 and Model 2 predictions are plotted. Once again it is clear that the model plots do not overlap. Although it is observed that both models called the October 2008 GFC, Model 1 was sending more intense signals than model 2. For the analysis, the probability predictions of the models and model plots were separated and presented in Figure 4 (b), Model 1 and Figure 4 (c), Model 2. As depicted in Figure 4(b), Model 1 accurately predicts the October 2008 GFC. Early warning signals were emitted as early as the end of 2005 and the beginning of 2006. However, these signals were classified as false alarms because they occurred outside of the 24-month crisis window. Model 1 continued to send signals throughout the latter part of 2006 and early 2007 (within the 24-month window), reaching a peak probability of approximately 90%. Although the probability subsided to around 30%, it rose again to 90% in early 2008. By October 2008, the predicted probability of a crisis remained above 40%.

In contrast, Model 2 (Figure 4(c)) also correctly predicted the 2008 crisis, but with an approximate probability of 52%. However, the signals emitted 24 months prior to the crisis only reached magnitudes of 30%. For the out-of-sample period of 2005/01 to 2017/03, specifically after the October 2008 crises, Model 1 sends signals of an impending crisis of between 20% to 40%. Model 2 emits signals during the same time period to just below 50%. During that period South Africa never formally reported any occurrence of a currency crisis: fluctuating signals could have simply been attributed to South Africa’s extremely volatile foreign market fluctuations or they could have pointed to some volatility of an impending crisis or simply regarded as false signals.

The primary objective of this study was to develop an early warning system (EWS) model capable of effectively predicting financial crises, rather than merely identifying tranquil periods. To achieve this, the study utilized various cut-off probabilities (20%, 30%, 40%, and 50%) as thresholds for crisis prediction. To evaluate the performance of the EWS model in predicting South African financial crises, both in-sample and out-of-sample, six assessment methods were employed (Berg and Pattillo, 1999):

(i) Percentage of Observations Called Correctly: Measures the overall accuracy of the model.

(ii) Percentage of Pre-Crisis Periods Called Correctly: Assesses the model’s ability to accurately identify periods leading up to crises. (iii) Percentage of Tranquil Periods Called Correctly: Evaluates the model’s effectiveness in distinguishing tranquil periods from those with potential crisis indicators.

(iv) False Alarms as a Percentage of Total Alarms: Measures the rate of incorrect crisis predictions.

(v) Diebold and Rudebusch’s (1989) Quadratic Probability Score (QPS): Quantifies the accuracy of probabilistic forecasts.

(vi) Global Score Bias (GSB): Measures the overall bias in the model’s predictions.

For a more detailed explanation of these methods, please refer to the appendix. The results of the performance evaluation are presented in Table 6. Following Berg and Pattillo (1999), the performance of the models can be seen from its ability to capture the whole observation when there is a crisis and when there is no crisis. Thus, from Table 6, it is seen that as the threshold increases to from 20% to 50%, the percentage of observations correctly called also increases in both models. With regard to the first out-of-sample probabilities, as the threshold increases from 20% to 50%, the percentage of observations correctly called decreases. The same is true for the second out-of-sample period. To conduct a more comprehensive evaluation, it is useful to consider the percentage of pre-crisis periods correctly identified as well as the percentage of tranquil periods correctly classified. In the in-sample periods, as the cut-off probability increases, a decrease in the percentage of pre-crisis periods correctly called is observed, while the percentage of tranquil periods correctly called follows a similar trend. For the first out-of-sample period, both Model 1 and Model 2 demonstrated the ability to correctly identify crises within the 24-month window prior to their occurrence. Model 1 exhibited a higher success rate, correctly calling crises in 96% of cases compared to Model 2’s 52%. However, both models also generated a high percentage of false alarms, indicating the need for further refinement. Comparable results were observed in the second out-of-sample period, although with slightly reduced performance. In terms of false alarms, both in-sample and out-of-sample periods showed a decreasing trend in false alarm rates as the cut-off probability threshold increased, aligning with the observations depicted in Figure 4. In summary, the non-linear relationship shows the delicate balance between correctly detecting crises (true positives) and not misclassifying calm times (false alarms). The percentage of false alarms goes down as the threshold goes up, but some crises are missed. This is why the percentage of false alarms goes down in a way that is not linear.

Both models have performed well on the whole, as indicated by the small values of the QPS and GSB in which a value of zero indicates perfection and a value of 2 indicates model failure (Handoyo et al., 2020; Gupta & Kumar, 2022b). Furthermore, based on these two measures it is obvious that the Model 1 outperforms Model 2. This confirms the results of Zhang et al. (2021), showing that an artificial neural network model with more input neurons is better than that with fewer input neurons.

The study has developed and estimated an artificial neural network (ANN) model as an alternative EWS for predicting currency crises in South Africa. Two models were constructed: a comprehensive model (Model 1) incorporating ten input variables and a reduced model (Model 2) using only four statistically significant variables. The purpose of estimating the reduced model was to investigate the impact of the number of input variables on model performance, addressing a common point of contention in the literature. Performance evaluations confirmed that models with more input variables outperform those with fewer. The study also identified the key contributors to South Africa’s currency crises: the ratio of budget deficits to GDP, followed by the growth rate of the ratio of domestic credit to GDP, and then changes in the international liquidity position.

The ANN models demonstrated their effectiveness by accurately predicting currency crises up to 24 months in advance, both in-sample and out-of-sample. Both models successfully identified all three notable currency crises in South Africa during the study period. However, Model 1, with its larger number of input variables, was able to provide earlier warning signals, up to 24 months prior to crisis occurrences, compared to Model 2. Additionally, both models correctly predicted the absence of currency crises after October 2008, a period during which South Africa did not report any such events. A particularly noteworthy feature of the ANN model was its ability to predict the probability of the out-of-sample GFC without requiring prior training with a target output. This demonstrates the model’s capacity to identify emerging trends and patterns that may indicate potential crises. While the ANN models demonstrated promising results, they also generated a considerable number of false alarms during the 1998 and 2001 periods. This suggests a potential limitation of the models, as they were unable to effectively differentiate between currency crises and other economic vulnerabilities. Further research is needed to explore potential relationships between various economic vulnerabilities and their impact on currency crises. Overall, the study concludes that ANNs offer a promising alternative to traditional EWS tools for predicting currency crises in South Africa. The models’ ability to provide early warnings and accurately identify the key contributing factors highlights their potential value in risk management and policymaking.

Fernandes-Gondoza, G: Conceptualization, Formal analysis, Data curation, Investigation, Methodology, Writing – original draft, review & editing. Ncwadi, R: Project administration, Supervision, Validation, Writing – original draft. Fernandes, J: Data curation, Software, Resources, Methodology, Data Visualization, Writing – original draft. Nyika, F: Writing – Review & editing, Updating of manuscript literature.

Artificial Intelligence (AI) tools were not used to write this manuscript.

All authors declare no conflicts of interest in this article.

No funding was used in the writing of this article.

In evaluating the performance of the four EWS models, suggest the following assessment criteria: the percentage of observations correctly called; the percentage of pre-crisis periods correctly called; the percentage of tranquil periods called and the percentage of false alarms to total alarms. In order to carry out these assessment methods, use is made of the two-by-two matrix in Table A1, whereby the crisis signal was classified into four categories, namely, A, B, C and D (See Table A1 below for explanation of categories). The assessment methods are given by the following equations:

If the values in Equations 16, 17 and 18 are high, the model performs better. In contrast, in Equation 19, the lower the ratio the better the indictor. The QPS and GSB allow for the assessment of the average closeness of the realisation of a crisis (R_t) and the probability prediction of a crisis (P_t) during the signalling horizon. Recall that if the event is a crisis, it scores 1 and 0 otherwise. The QPS statistic lies in the range between zero and two. The performance quality of the composite index is better the closer the test statistic is to zero. The performance quality of the composite index can be affected by the sample size. The larger the sample size, the more robust the test statistics (Diebold & Rudebusch, 1989).

The ten leading indicators used in this study were selected using the signal approach. The performance of an indicator in predicting a crisis can be shown in the value of its noise-to-signal ratio (NSR). The NSR can be presented by taking the ratio of the percentage of bad signals over the percentage of good signals (Kaminsky et al., 1998),

$NS{R}_t=\left[\frac{B_t}{B_t+D_t}\right]/\left[\frac{A_t}{A_t+C_t}\right]$ (15)

In Equation 15, if the NSR ≥ 1 for an indicator, this indicates excessive noise and thus contributes less to the probability prediction of a currency crisis. The desired outcome is to have an NRS that is less than 1. Kaminsky et al. (1998) specified that the lower the value of the NRS, the more powerful the indicator in predicting crises.