Impact of Risk Factor Proxies Used to Construct BSFIs

One ultimate aim of this thesis is to compare the choice of different risk factor proxies for constructing the BSFI in terms of their effect on the predictive power performances of the EWS models. To this aim we define twenty-five different BSFIs which form the dependent variables of twenty-five different EWS models. Accordingly, in order to measure the credit risk factor, we use either one of the two proxies as NPF and BC. Market risk is proxied by considering either the TIER or FL proxies. We proxy liquidity risk and profitability risk factors using DEP and ROE respectively. For example, the BSFIs of Model 1 and Model 5 are constructed based on those factors of liquidity risk, credit risk and market risk. In Model 1 and Model 5, while market risk and liquidity risk are measured by using the same proxies as TIER and DEP respectively; in Model 1 the proxy for the credit risk is chosen as NPF whereas, in Model 5 credit risk is proxied by BC. In other words, these two models differ only in the credit risk proxies. Likewise, in Model 5 and Model 15, while the credit risk and liquidity risk are measured by the same proxies in both models; in Model 5 the market risk is measured by TIER and in Model 15 it is proxied by FL. The BSFIs of Model 9 and Model 14 are constructed on credit risk, liquidity risk, market risk and profitability risk factors. In both models the same proxies of BC, DEP and ROE are used to measure the credit, liquidity and profitability risk respectively. However, in Model 9, the market risk is measured by TIER but in Model 14 it is proxied by FL. According to the results given in Table 16, despite the fact that the BSFIs are defined considering the same risk factors for Islamic banks (i.e. credit risk, liquidity risk, market risk, profitability risk), the predictive power of different EWS models is highly sensitive to the proxies that are used to measure those risks and their combinations.

In terms of the proxies that are used to measure the credit risk, we use two different proxies as NPF and BC. To be able to identify which proxy has reveal more successful results in terms of the prediction power of the EWS models, we compare those models that we constructed with the same variables other than the credit risk proxy. For instance, while Model 1 is constructed with NPF, DEP and TIER proxies, the BSFI of the Model 5 is constructed using BC, DEP and TIER proxies. In addition, Model 2 and Model 3, Model 4 and Model 15, Model 6 and Model 16, Model 7 and Model 9, Model 8 and Model 10, Model 11 and Model 14, Model 23 and Model 17 and Model 20 and Model 21 are constructed in the same vein as the pair of Model 1 and Model 5 in order to provide an opportunity to analyze the marginal impact of the BC and NPF proxies on the predictive power rates of the EWS models while keeping other proxies of the risk factors fixed.

When we compare the models, we observe that although using BC as a proxy for the credit risk does not make a clear difference, however in the overall, it has a weak positive effect on the predictive power of the EWS model. For instance, when Model 4 and Model 15, Model 11 and Model 14, Model 8 and Model 10 are compared pairwise, it is seen that using BC as a proxy rather than NPF increases the predictive power rate only by 3%, 1%

and 4% respectively. On the other hand, in some of the model comparisons (i.e. Model 1 and Model 5, Model 7 and Model 9 and Model 20 and Model 21), using NPF as a credit risk proxy instead of BC increases the predictive power by 1% and 3% respectively. In terms of Model 13 and Model 17, it is observed that both variables reveal the similar predictive power results. However, when Model 2 and Model 3 and; Model 6 and Model 16 are compared, it is observed that the using BC as a proxy for credit risk increases the predictive performance of the models by 10% and 6% respectively. While the BSFI of Model 3 is comprised of BC and TIER proxy, Model 16 is constructed with BC and FL proxies. Since both models are constructed with BC and market risk proxies, it can be concluded that if the fragility of Islamic banks is measured by considering credit risk factor and market risk factor, BC reveals more substantial results compared to the NPF proxy. This might stem from the strong link between the domestic credit to private sector and market risk. The domestic credit is triggered by markets risk factors as exchange rate changes and interest rate volatilities and it has significant effects on the economic activity

and financial stability (Yuafi and Bawono, 2017). Accordingly, the Islamic banks also become more fragile to crises. In the related literature NPF and BC are widely used and accepted as a prominent determinant for both conventional banking and Islamic banking system to measure the credit risk (Kibritçioğlu, 2003; Firmansyah, 2014; Salim et al., 2016; Khan et al.,2020). To sum up, credit risk can be proxied both by using CB or NPF variables in constructing a BSFI for Islamic banks. However, when the BSFI is constructed with a combination of credit risk and market risk, BC increases the predictive performance of the EWS rather than the NPF.

In terms of the market risk factor for Islamic banks, we consider FL and also, we try to measure this risk utilizing the size of the banks in terms of their eligible capital by using the TIER proxy. To compare the impact of the market risk proxies, the BSFIs are constructed by keeping the other risk factors and proxies rather than the market risk proxies as the same. This gives an opportunity to compare the EWS models that only differ in terms of the market risk proxy. For instance, while the BSFI of the Model 9 is constructed by using BC, DEP, TIER and ROE proxies, the BSFI of the Model 14 comprise of the BC, DEP, FL and ROE. Likewise, Model 1 and Model 4; Model 2 and Model 6; Model 3 and Model 16; Model 5 and Model 15; Model 7 and Model 11; Model 8 and Model 12; Model 9 and Model 14; Model 10 and Model 13 and; Model 22 and Model 24 can be compared pairwise to investigate the impact of the TIER and FL proxies on the predictive power rates of the EWS models while keeping other proxies of the risk factors same. According to our results, while including TIER variable into the BSFI as a measure for market risk decreases the predictive power of the EWS models; using the variable FL improves the ability of the models in predicting fragility and tranquil episodes correctly. For instance, when we change the market risk proxy to FL in Model 1, the predictive performance increases by 11% (that becomes Model 4). In addition, the predictive power of the EWS model 15 increases by 15% by including FL instead of TIER (that becomes Model 5). In a similar vein, while the predictive power of the Model 9 is 71%, it increases to 81% when we measure the market risk by FL. Furthermore, we observe the same results in all comparisons of the models.

Tier is used to calculate the eligible capital, i.e. minimum capital requirement to cover the market risk of banks. The eligible capital is the sum of Tier 1 (shareholders’ equity and retained earnings) and Tier 2 (supplementary capital) (BCBS, 2006). Foreign liabilities, on the other hand, represents the total amount of the banks’ liabilities in foreign currency items. Our results indicate that while constructing a BSFI to define the fragility and tranquil episodes of Islamic banking, measuring the market risk using FL rather than using Tier gives better prediction results. Within this framework, when the value of the assets of a bank falls below the value of its liabilities, the financial structure of that bank deteriorates. The currency mismatch that arises between a bank's foreign currency assets and its foreign currency liabilities is called a foreign currency position. In other words, if banks' short-term liabilities in foreign currency exceed their short-term assets in foreign currency, the banking system would be in a liquidity shortage on an international basis.

In this regard, exchange rate risk is an important source of the market risk of the banking sector which mainly arises from the investments made by banks in foreign exchange transactions. Transactions as acquiring and distributing funds in foreign currency can be affected by the exchange rates movements. In cases where it is not expected an upcoming devaluation in the domestic currency, banks tend to acquire funds from international financial markets. However, if domestic banks have high amount of unhedged foreign currency debt, a sudden devaluation can cause a significant reduction in the bank’s net worth and threatens its profitability by disrupting the financial structure of the banking system (Demirgüç-Kunt and Detragiache, 1998; Kibritçioğlu, 2003).

Furthermore, besides the banks’ own balance sheet, Mishkin (1999) explains the impact of the domestic currency depreciation through firms balance sheet. According to his view, in case of the depreciation, the financial structure of firms deteriorates since their burden of debt increases more than their assets. As a result, this leads to problems in the return of debt to banks by causing capital depletions in banks (Mishkin, 1999). The basis of the Asian crisis is the growth of foreign currency openings of these countries. One of the reasons for this situation is the fact that a large amount of loans was provided to Asian countries, especially from foreign commercial banks, in the 1990s. Five Asian countries from international commercial banks as Indonesia, Korea, Malaysia, Philippines, and Thailand provided loans of approximately 150 billion dollars in 1990, while

approximately 390 billion dollars were extended in 1997, the beginning of the crisis. Most of the loans extended by foreign banks consist of short-term loans. The fact that this capital was made available by banks in local currency as loans for sectors that do not create value added and thus, the non-repayment of these loans contributed to the growth of the crisis (Chang and Velasco, 1998; Kaplan, 2002).

Note that, even the BSFIs defined in the EWS models are based on the same risk factors, the proxy variables that are used to measure these risk factors differ in each index. Indeed, our results show even if the same risk factors are utilized, the predictive powers of the