CHAPTER 2 - EARLY WARNING SYSTEMS AND THE BACKGROUND
2.4. ESTIMATION METHODOLOGY
Jing et al. (2015) find that using nominal interest rate rather than real interest rate in the IMP increases the number of correctly predicted crisis episodes. Moreover, they show that their modified index gives better predictive power results with fewer false alarms than Von Hagen and Ho (2007).
Therefore, for constructing an EWS, there is no optimal number of independent variables.
For this reason, determining the optimum number and list of explanatory variables is an important and challenging criterion in the construction of an EWS model.
The non-parametric approach, namely the Signal Extraction Approach, is developed by Kaminsky et al. (1997). The logic behind the Signal Extraction Approach depends on the different behavior of the indicators in the crisis and tranquil episodes. Therefore, the Signal Extraction Approach is based on monitoring the behavior of a large set of indicators in crisis and tranquil periods in order to predict a possible future crisis. More precisely, in the first part of this approach, the explanatory variables, which are expected to have significant impact on the occurrence of a crisis, are selected. Correspondingly, an arbitrarily determined threshold value is determined for each variable and the values of these indicators are compared in crisis and tranquil episodes. Within this context, each indicator is analyzed separately and if an indicator deviates from its normal level and exceeds the threshold value, it is interpreted as an early warning signal of a possible crisis that will occur in the following 24 months.
The primary issue in the signal approach is to determine the optimal threshold value since the signals given by the model depend on these threshold values. If a low threshold value is determined, the system gives more crisis signals which increases the number of wrong signals (Type 2 error). Higher threshold value, on the other hand, sends less crisis signals that leads to missing actual crisis episodes (Type 1 error).
Table 2: Signal Matrix
Yi,t=1
Crisis within 24 months
Yi,t=0
No crisis within 24 months Signal
A
Good signal of crisis event
B Type 2 Error
False Signal No signal
C Type 1 Error Missing Signal
D
Good signal of non-crisis event
According to the table, A denotes the number of months that the indicator sends good signals. In other words, good signals mean that the indicator sends crisis signals and crisis occurs in the following 24 months. B is the number of months that the indicator issues a bad signal. This means that the indicator sends crisis signal but no crisis observed in the next 24 months which is also referred as type 2 error. C denotes for number of months that the indicator does not issue any signal but a crisis occurred which is the type 1 error.
D is the number of months that there the indicator does not send any signal and no crisis event experienced. Kaminsky et al. (1997) defines the ratio of false signals to good signals as noise to signal ratio. Accordingly, the authors adjusted the optimal threshold value for each indicator that minimize the noise to signal ratio, which can be formulized as:
Noise to signal ratio = [ 𝐵
𝐵+𝐷]/[ 𝐴
𝐴+𝐶] (2)
The non-parametric EWS, the signal extraction approach, is widely accepted by the academics and the policy makers since it gives an opportunity to analyze wide range of explanatory variables by revealing the leading indicators of a forthcoming crisis by giving the prediction powers of each indicator. For instance, Borio and Drehmann (2009) conduct their analysis with signal extraction approach and find reasonable evidence about the impact of credit and asset prices on the probability of banking distress. In addition, their model correctly predicts the crisis by 77% with a noise to signal ratio varying between 6% and 14%. Furthermore, the signal extraction approach also implies to EWS of Islamic banking crises by Kusuma and Asif (2016) and Wiranatakusuma and Duasa (2017).
However, this method is frequently criticized due to some of its drawbacks. For instance, as Frankel and Rose (1996) state, the signal extraction approach reveals the individual contribution of the indicators rather than the marginal contributions thus the relation between the explanatory variables is neglected. Furthermore, determining the optimal threshold value is another crucial point since the prediction power and the leading indicators are directly related with this value. If a low threshold value is chosen, the system gives more crisis signals which increases the number of false alarms (Type 2 error). Higher threshold value, on the other hand, sends less crisis signals that leads to missing actual crisis episodes (Type 1 error). For this reason, the threshold value is set regarding to minimize the Type 1 and Type 2 errors which appears as a crucial and challenging aspect of the model. Furthermore, this approach does not give an opportunity to evaluate the amount of the deviations of the indicators from the threshold value and to test the statistical significance levels of the indicators.
In this respect, some of the drawbacks in the signal extraction approach have been solved within the framework of the parametric approach, namely with limited dependent variable approach. It is a regression-based approach in which binary models with logit or probit functions are used and the probability of the crises is estimated as a function of various explanatory variables. While the signal method transforms each variable into a binary variable by restricting to observe the relationship between the indicators, the logit and probit methods analyze all variables simultaneously and reveal the marginal contribution of each variable. Additionally, they give an opportunity to test the statistical significance of the indicators by providing the magnitude of each variable. As in signal extraction methodology, the parametric approach has also some drawbacks. For instance, while it provides an opportunity to find if a variable is a significant indicator of a banking crisis or not, it is not possible to determine how successfully the individual variable predicts the crisis episodes.
The first attempt to employ logistic methodology in an early warning model of banking failures is made by Martin (1977). The author considers 5700 Federal Reserve member banks in US between the period of 1990 and 1976. According to the results of the study, the logit model reveals substantial results by correctly classifying the failed and non-failed banks by 87% and 88.6% respectively. After Martin (1977), the logit model has become a widely used methodology in predicting banking crises in the related literature (Demirgüç-Kunt, 1989). For instance, Lestano and Kuper (2002) construct and EWS model by conducting logistic methodology to examine the significant indicators of currency, banking and debt crisis on a panel of six Asian countries15 over 1970-2001 period. The authors find that GDP per capita and the ratio of M2 to international reserves are crucial determinants of the banking crisis.
Furthermore, there are vast number of studies in the literature that attempt to compare the parametric and nonparametric EWS.
For instance, Berg and Patillo (1999) compare the signal extraction approach to probit approach and find that probit approach gives superior prediction results compared to the signal approach. By comparing parametric and non-parametric EWS, Beckmann et al.
(2007) show parametric EWS is more successful than non-parametric EWS in terms of identifying the correct crisis episodes. Moreover, Davis and Karim (2008) construct an
15 Malaysia, Indonesia, Philippines, Singapore, South Korea, and Thailand.
EWS model for banking crisis and compare the results of the logit model EWS and signal extraction model EWS. The authors conclude that while the logit model shows a better performance in predicting global EWS, the signal extraction model is more appropriate to anticipate the country specific banking crises. Additionally, Comelli (2014) compares the logit and probit EWS. The author states that while both models reveal similar outcomes, logit EWS gives slightly better prediction results compared to those from probit EWS.
In sum, there is a wide range of studies that construct early warning systems to anticipate the banking crises for different countries relying on various explanatory variables, estimation methods and time periods. Moreover, most of these studies consider only the conventional banking system. To the best of our knowledge this study is the first attempt to investigate (i) the impact of banking crisis definition variations on significant indicators of the banking crises of Islamic banks (ii) the choice of the proxies that are used to construct BSFIs on the predictive power of EWS of Islamic banks, (iii) the impact of banking risk factors that are used to construct BSFIs on the predictive power of the EWS of Islamic banks, (iv) the impact of the profitability risk factor on the predictive power of EWS models and, (v) a BSFI definition specific to Islamic banks. In addition, this study extends the related literature by establishing an EWS model for Islamic banking system that covers all the leading countries in terms of Islamic banking assets rather than focusing on a country as well as examining a wide range of banking-specific and macroeconomic factors.