Causal Effect of Financial Market Frictions and Flight to Quality on Cost of Credit in Kenya

Financial market conditions have been declining over the past ten years globally as most developing countries continue to adopt more liberal financial policies, such conditions may amplify adverse shocks to the economy. The Kenyan Banking sector was highly profitable before the implementation of financial market frictions, with industry return on equity’s average of 20%. The ratio of credit supply to gross domestic product was 35%; and the economy grew by 5.6 %. Nonetheless, after its adoption, listed Banks recorded negative Earnings per Share growth of 8.2%, compared to an average positive growth of 14.1%, The Net Interest Margin declined to 8.4% from 9.4%. Studies relating to financial market frictions, flight to quality and Cost of Credit have produced mixed results. It was on this basis that this study sought to establish the effect of financial market frictions and flight to quality on cost of credit in Kenya. The study adopted correlational research design. Secondary data from the Kenyan Market for the period January 2009 to December 2019 was analyzed. Augmented Dickey Fuller and Philips-perron unitroot test was used to test the stationarity of the data. VECM was estimated to establish the speed of adjustment towards the long run equilibrium; Wald statistics was also estimated to establish short run causalities amongst the variables. Based on cointegrating equations, the error correction term indicated a negative sign and was significant at 5% level (C (1) = -0.153042, .0429 < 0.05), an indication that a long run relationship exists amongst the variables. Wald statistics revealed that the estimated coefficients in the VECM were insignificantly different from zero (.8417; .5603; .9188>p=0.05),however, Central Bank rate was found to be different from zero and significant at 5% level (.0163>p=0.05), an indication that there was a short run casualty running from the Central Bank rate to cost of credit. The study therefore recommends that for Micro finance institutions to maximize their profits they should adopt new technologies like Mobile Banking for their credit facilities, this does not require administrative and operation costs, in a bid to cope with the market shocks and frictions.


Kenya lies between latitudes
. The country covers 569,140 square kilometers of land and 11,227 square kilometers of water, with a total area of 580,367 square kilometers. Her population is approximately 47,564,296 according to the 2019 population census. The GPS coordinates of Kenya show that the country is bisected by the equator. Approximately half of Kenya is in the northern hemisphere.

Data Type, Sources and Collection Methods
The data used in this study was sourced mainly from financial statements of Banks, Bank Supervision reports, Official websites of commercial Banks, Think Business Banking Survey and publicly listed enterprises. Our time series data set covers the period from the first month of 2009 to the last month of 2019. We also used several macroeconomic series, which are mostly sourced from the Central Bank of Kenya website. This diminishes the problems associated with heterogeneous demand shocks, because the share of different types of loans in the banks' portfolios does not differ significantly.

Model Specification.
A general Vector Autoregressive Model (VAR) of order "P" below was used to generate VECM; VECM was applied to find long-run equilibrium associations. To assess the short-run and long-run coefficients of the variables, we developed the following from equation (1.0) to form a VECM model and is generated recursively as; It contains long-run information derived from the long-run cointegrating relationships. This study expresses the lagged OLS residual obtained from the long-run cointegrating equations as; (1.2) From equation (1.2) we can re-write Error Correction Term (ECT) as; (1.3) 2.5 Data Analysis. Augmented Dickey Fuller (ADF) unit-root test and Philips perron test (PP) was done to check the stationarity of the time series data on the basis of a null hypothesis that the time series were non stationary (i.e. δ = 0) and alternative hypothesis that the time series were stationary (i.e. δ ≠ 0). This was undertaken as a precautionary measure against estimation of spurious regression models (Sim et al., 1990). The ADF unit root test will take the form of; is the difference operator, 0 a is a constant, and α is the autoregressive lag coefficient. The ADF then tests the hypothesis; the null hypothesis for the test is given below; 0 : 0 = α H , there exists a unit root problem. If t-statistic > ADF critical value, accept the null hypothesis. If tstatistic < ADF critical value, reject the null hypothesis. If the null hypothesis is rejected, the data of the series is stationary and can be used for modeling without taking any difference of the series. The Dickey-Fuller test statistics have been criticized for their low power, especially in distinguishing between unit roots and near unit roots and in small sample data while Phillips-Perron (PP, 1988) test is more robust to serial correlation, time dependent heteroscedasticity and regime changes (Moosa & Bhatti, 1997).

Cointegration Test
Cointegration test was performed to confirm the long run relationships amongst the variables. Johansen (1988), Johansen and Joselius (1990) Cointegration test was adopted, the two proposed two different likelihood ratio tests: the trace test and maximum eigenvalue test, as shown in equations (1.5) and (1.6) respectively.  Normality test was then conducted using Jarque-Bera statistics and the results are presented in Figure 2.0. The results shows that the P-value for the Jarque-Bera statistics is more than 5% (i.e 0.677570 > p=0.05), an indication that the data used were normally distributed.

3.2.2: Test for Heteroskedasticity.
The study further tested for the Breusch-Pagan-Godfrey Heteroskedasticity effect, with the null hypothesis that the error term was not heteroskedastic. Since the estimated P-value(s) corresponding to the observed R-squared was more than 5% (0.1127> 0.05), the null hypothesis that the error term was not heteroskedastic was confirmed as seen in Table 2.

Unit Root test.
For stationarity of data to be achieved, the overall behavior of the data set should remain constant (Gujarat & Porter, 2009). Stationarity of the time series data is important in ensuring that an accurate forecasting of events is realised. Time series data was therefore, first subjected to stationarity test. Augmented Dickey-Fuller test (ADF) and Philips perron test (PP) was used to test for the stationarity. As a rule of thumb, the null hypothesis assumes the presence of unit root, and the p-value obtained should be less than the significance level (e.g. 0.05) while the absolute value of the test statistic should also be less than the critical value for the rejection of the null hypothesis. Referring to the above rule of thumb, the data sets for CC, CBR, NPL, PALL and TBLL in table(s) 4 and 5 have unit root. The ADF p-values obtained for each data set was greater than 5% (p=0.05 < .5545, .1201, .3655, .9327, .9428), this compares well with the p-values for PP (p=0.05 <.5126, .2535, .3659, .0809, .9472) which are also clearly greater than 5%. Similarly, the absolute values of the test statistics for each of the variables for both the ADF and PP are less than the corresponding absolute values of the test statistics at 5% level of significance. The study thus concludes that the series are non stationary at levels.   Table 6 and 7 shows the unit root test results for the series at first difference. From Tables 6 and 7 we can deduce that unit root does not exist in each of the series at first difference since the p-values for both the ADF and PP are less than 5% level of significance (p=0.05 < 0.0000). The deduction is further supported by the absolute value of the test statistics for each of the variables which are more than the corresponding absolute value of the test statistics at 5% level of significance. The study thus concludes that the series are stationary at first difference.   Table 8 shows VAR lag order selection criteria for Cost of Credit and the explanatory variables. Final prediction error (FPE), LR and Akaike information criterion (AIC) test statistic suggests lag 7 as the optimal lag. Schwarz information criterion (SC) and the Hannan-Quinn information criterion (HQ) suggest lag 1 as the optimal lag. Liew (2004) suggest that most economic sample data can seldom be considered large in size, AIC and FPE are therefore, recommended for the estimation of their autoregressive lag length, and since the observations in this study were relatively large, the Akaike information criterion (AIC) which suggested lag 7 at 93.85311* was chosen for the autoregressive lag length for cost of credit.

Cointegration Test
Data was then subjected to Cointegration test, Johansen (1988) and Johansen and Joselius (1990) two different likelihood ratio tests were adopted. Since the variables were stationary at first difference as shown in tables 6 and 7, cointegration test was therefore, necessary to establish a long run relationship. Results obtained from the Trace statistics and Maximum Eigenvalue Statistics as captured in Table 9 and Table 10 respectively, indicated that there is one (1) cointegrating equation or one error term >> At most 1, p=0.1740=17.4% and p= 0.2474 = 24.74% Statistics respectively at 5% level of significance, meaning all the variables are cointegrating. The null hypothesis that there is no Cointegrating equation is thus rejected. The results therefore, suggest that in the long run, the variables move together or have a long run association.   Table 11 shows normalized cointegrating coefficients. From the table, while it can be concluded that Central Bank Rate and Treasury bills, on average, had a positive effect on cost of credit in the long run, Ceteris Paribus, Non-performing loans and Provision in anticipation for loan losses, on average, had a negative effect on cost of credit, ceteris paribus. The coefficients are statistically significant at 5% level. Since the coefficients are just OLS estimates, they have to be interpreted as ceteris paribus effects, and the signs reversed in the long run (Green, (2003); Gujarat and Porter, (2009) ;Wooldridge, (2009). The null hypothesis that there is no Cointegrating equation is thus rejected. This means that there is a cointegrating relationship in the model.

. Vector Error Correction Model for Cost of credit and its explanatory variables
The vector error correction estimates (Appendix 1) were estimated based on the existence of the cointegrating equations. From the Appendix 1, the long run model explains the error correction term that signifies the long run relationship among the variables. As may be concluded from the estimates, the model posits that Central Bank rate, provisions in anticipation of loan losses, treasury bills and nonperforming loans are important determinants of cost of credit in the long run (t-statistics 2 < -7.779867, 2 < 2.80890 and 2 < -3.44197 respectively), the null hypothesis that there is no long run relationship among the variables is rejected, Results in Appendix 1 shows that one unit change in Central Bank rate and treasury bills is associated with 15,334.03 units and 0.192070 units respectively, increase in cost of credit on average ceteris paribus in the long run. Both the Central Bank rate and Treasury bills were directly related to cost of credit. While on the other hand, one unit change in provisions in anticipation of loan losses and nonperforming loans is associated with 0.582190 units and 0.082959 units respectively, decrease in cost of credit on average ceteris paribus in the long run. Both the provisions in anticipation of loan losses and nonperforming loans were inversely related to cost of credit. The null hypothesis From the Appendix 1, the previous periods deviation from long run equilibrium is corrected in the current period at an adjustment speed of 15.3 % ( CointEq1 = -0.153042). Table 12 shows a make system approach, the results shows that Nonperforming loans, provisions in anticipation of loan losses and treasury bills are not important determinants of cost of credit in the short run, (t-statistics 2 > 0.798599; 2 > 0.988691 and 2 > -0.305125 respectively), and were statistically insignificant at 5% level in the short run (p=0.05 < 0.4267; p=0.05 < 0.3256; and p=0.05 < 0.7610 respectively), the null hypothesis that there is no short run relationship among the variables is accepted. Central Bank rate, however, returned as an important determinant of cost of credit (t-statistics 2< -3.585777) and was statistically significant at 5% level (p=0.05 > 0.0006), and it had an expected negative sign ( φ = -3953.296), which is a good sign, an indication that there is a short run relationship between Central Bank rate and Cost of Credit (Green, (2003); Gujarat and Porter, (2009); Wooldridge, (2009). The null hypothesis that there is no short run relationship between Central Bank rate and cost of credit is therefore, rejected. A Wald test statistic (table's 19a, b, c, and d) is further performed to confirm if indeed there is no short run relationship among the explanatory variables; Nonperforming loans, provisions in anticipation of loan losses, treasury bills and cost of credit. It was also used to confirm the short run causality between Central Bank rate and cost of credit.  Table 12 shows the VECM that was estimated based on the existence of the cointegrating equations. The dependent variable was Cost of Credit (CC) while the independent variables were Central Bank Rate (CBR), Non Performing Loans (NPLs), Provisions in anticipation for Loan Losses (PALL) and Treasury Bills (TBLL). The error correction term indicated the expected sign and was significant at 5% level (C (1) = -0.153042, p = .0.0429< 0.05), this means that the speed of adjustment towards long run equilibrium is negative and statistically significant. This implies that there is a long run causality running from Central Bank Rate (CBR) to cost of credit, in other words, Central Bank Rate has influence on cost of credit, this observation corroborates with Demetriades and Luintel (2001) who asserted that the use of interest rate ceilings, distorts the economy and inhibits financial deepening by depressing real rates of interest, consequently, MFIs levies other charges in a bid to recover their costs, in essence these additional charges increases cost of credit. Mohane et al, (2002) further supports this argument by stating that ceilings produces a series of adverse effects, they argue that since MFIs are not allowed to charge full cost recovery interest rates, they then drift up their operating costs thereby increasing their cost of credit. Table 18 also shows that there is a long run causality running from Non Performing Loans (NPLs) to cost of credit, this observation corroborates with Brown, Jappelli, and Pagano, (2009) whose estimates in their study showed that information sharing is associated with improved availability and lower cost of credit, Gaitho, (2013) observed that CRB reduces borrowing costs and loan delinquencies to a moderate extent.
These observations however, contradicts Khandare and Alshebami, (2015); Miller, (2013) argument that for MFIs to remain sustained and active in the market, it is mandatory for them to cover their costs; cost of borrowing, cost of operation, inflation cost, cost of default loans and other costs of delinquencies must be levied on the borrower and expenses incurred when carrying out their activities not forgetting to add their profit margin. There was also a long run causality running from Provisions in anticipation for Loan Losses to cost of credit, this corroborates Hela, Senda, Younes & Collins (2016), they concluded that IFRS 9 adoption represents a key determinant of information asymmetry reduction, they argue this contributes significantly to decrease in cost of credit for post IFRS 9 period. This however, contradicts Chen et al, (2013), who indicated that IFRS adoption led to higher interest rates, greater likelihood of demand for collateral and shorter maturities, Gehrig and Stenbacka, (2007) who also contradicts this narrative stated that issues to do with lower credit ratings pay higher interest rates embodying larger risk premiums than higher rated issuers. Lastly, Table 18 shows that there exists a long run causality running from Treasury Bills to Cost of Credit; this observation corroborates Gubareva and Borges (2013) who noted that flight-to-quality events can be observed while correlation between safe and risky assets performance holds and, in some cases with increasing prices of risky assets, Jones, (2012) offered a further opinion arguing that flight to quality a cross financial markets have a strong negative interaction in sovereign debt markets, the structure of collaterals offered by individual micro enterprises led higher cost of credit. Their argument however, contradicts Gatev and Strahan (2006) who studied Banks' balance sheets and found out that when the spread between treasury bills and high grade commercial paper increases, Banks tend to experience inflows of deposits and decreased cost of funding.     (2020) 3.8. Short run Causalities 3.8.1. Short run casualties for cost of credit and its explanatory variables.
The study further employed Wald statistics to test whether or not the estimated coefficients in the VECM were significantly different from zero (i, e.) C(9)=C(10)=C (11) Table 13a is less than 5% (.0163<p=0.05). Thus, the null hypothesis of C(9)=C(10)=C(11)= C(12)=C(13)=C(14)=C(15)=0 is rejected, implying that there is short run causality running from Central Bank rate to cost of credit and is significantly different from zero. The Chi-square probability corresponding to the null hypothesis on other variables as presented in Table 13a-d were more than 5% (.8417; .5603; .9188>p=0.05). Thus, the null hypothesis of C(16)=C (17) Table 13b. Table 13c shows a similar observation from Provision in anticipation of loan losses to cost of credit. And lastly, Table 13d, indicates that there is no short run causality running from Treasury Bills to cost of credit. This is interdem with Kashyap, Stein, and Wilcox, (1993) who stated that, in the short run, the relation among the variables would be unstable and may fail to support the expectation hypothesis due to variable term premiums which are not under the control of the monetary authorities -especially in recent times when there are frictions after in the financial markets.     Table 14 shows Breusch-Godfrey Serial Correlation LM Test for cost of credit that was conducted on the data post the analysis to assess any possibility of serial correlation. The test yielded an observed R2 of 0.100120 P = .9512>0.05, suggesting lack of serial correlation.  (2020) The study further tested for the Autoregressive Conditional Heteroskedasticity (ARCH) effect on cost of credit, with the null hypothesis that there was no ARCH effect. Since the estimated P-value corresponding to the observed R squared was .8241> 0.05, the null hypothesis that there was no ARCH effect was confirmed as seen in Table 15.

Summary and Conclusion.
The study investigated the long-run and short-run relationships among financial market frictions, flight to quality and cost of credit using Johansen's methodology of multivariate cointegration analysis and Vector Error Correction Model. The objective was to determine the effect of Central Bank rate on cost of credit in Kenya. Correlation results shows that Central Bank rate was positively associated with the cost of credit and was significant at 5% level ( r = .805565; .0000> p=.05); vector error correction estimates indicated that Central Bank rate is an important determinant of cost of credit in the long run (t-statistics 2 < -7.79867). Vector error correction term coefficient shows that one unit change in Central Bank rate was associated with 15334.03 units increase in cost of credit on average ceteris paribus in the long run. The null hypothesis that there is no long run relationship between Central Bank rate and cost of credit is therefore, rejected and the alternative accepted. Wald statistics results shows that there is a short run causality running from Central Bank rate to cost of credit and was significantly different from zero at 5% level (C(9)=C(10)=C(11)=C(12)=C(13)= C(14)= C(15)= 0; (.0163 < p= 0.05 ). The null hypothesis that there is no short run relationship between Central Bank rate and cost of credit is therefore, rejected and the alternative accepted. This finding invalidates this study's null hypothesis that Central Bank rate does not affect cost of credit in Kenya. This study concludes that market frictions i.e ceilings that are set too low are problematic especially in cases where they do not cover fees and commissions. World over, even if such ceilings are intended to reduce usury and exorbitant lending by MFIs who charge very high interest rates, they are very difficult to enforce. Moreover, lending practices without prudent regard for repayment capacity of micro enterprises, deceptive terms, and unlawful collection techniques causes more damage to microenterprises than do high interest rates. This is in agreement with Onyango and Odondo, (2018) ;Acclassato, (2006); Mohane et al, (2002) who elaborated that when interest rate ceilings are implemented, MFIs may be forced to impose additional charges which are not part of interest rates to cover administrative costs, this not only reduces transparency about the borrowers true cost of borrowing but also camouflages the actual interest rates charged by MFIs even if on surface, the cost may appear to be reducing like in the case of interest rate ceilings as shown in this study. The second objective was to establish the effect of provisions in anticipation of loan losses on cost of credit in Kenya. From the research findings, correlation results revealed that provisions in anticipation of loan losses was negatively associated with cost of credit and was significant at 5% level ( r = -.296420; .0006> p=.05); vector error correction estimates denoted that provisions in anticipation of loan losses is an important determinant of cost of credit in the long run (t-statistics 2 < 2.80890). Vector error correction term coefficient suggested that one unit change in provisions in anticipation of loan losses was associated with 0.582190 units decrease in cost of credit on average ceteris paribus in the long run. The null hypothesis that there is no long run relationship between provisions in anticipation of loan losses and cost of credit is therefore, rejected and the alternative accepted. Wald statistics results shows that there is no short run causality running from provisions in anticipation of loan losses to cost of credit and was not significantly different from zero at 5% level C(23)=C(24)=C(25)=C(26)= C(27)=C(28)= C(29)=0; (.5603> p= 0.05 ). The null hypothesis that there is no short run relationship between provisions in anticipation of loan losses and cost of credit is therefore, accepted and the alternative rejected. This study concludes that provisions in anticipation of loan losses affects cost of credit in the long run, provisions mitigates the MFIs against losses occasioned by loan default, costs that arise due to loan loss and default such as; costs associated with monitoring loans in arrears, post disbursement visits, costs of hiring external debt recovery experts, and other costs of delinquencies are greatly reduced since IFRS 9 is forward looking and loans are properly appraised and provided for at all times, this consequently reduces cost of credit.
As depicted in the research findings, correlation results evidenced that non-performing loans was positively associated with cost of credit and was significant at 5% level (.0388> p=.05). The null hypothesis that there is no long run relationship between non-performing loans and cost of credit is therefore, rejected and the alternative accepted. Wald statistics results shows that there is no short run causality running from non-performing loans to cost of credit and was not significantly different from zero at 5% level C(16)=C(17)=C(18)=C(19)=C(20)=C(21)= C(22)=0; (.0539> p= 0.05 ). The null hypothesis that there is no short run relationship between non-performing loans and Cost of Credit is therefore, accepted and the alternative rejected. The results of this study draw the conclusion that MFIs holding greater levels of NPLs are claimed for increasing profitability by equity investors in the long-run.
Investors perceive these MFIs as riskier than their counterparts or other assets, claiming greater returns on the equity holdings of these MFIs and hence inducing an increase in their cost of credit.
The last objective was to determine the long run relationship between flight to quality and cost of credit in Kenya. From the research findings, correlation results revealed that flight to quality was negatively associated with cost of credit and was significant at 5% level (r = -.687075; .0000> p=.05); vector error correction estimates elucidated that flight to quality is not an important determinant of cost of credit in the long run (t-statistics 2 > -3.44197). The null hypothesis that there is no long run relationship between flight to quality and cost of credit is therefore, rejected and the alternative accepted. Wald statistics results shows that there is no short run causality running from flight to quality to cost of credit and was not significantly different from zero at 5% level C(30)=C(31)=C(32)=C(33)=C(34)=C(35)= C(36)=0; (.9188> p= 0.05 ). The null hypothesis that there is no short run relationship between flight to quality and cost of credit is therefore, accepted and the alternative rejected. This study concludes that the increasing Treasury Bills take up by MFIs, is not a good sign to micro borrowers either, as it simply shows that there could be a flight to quality effect on Micro lending. This is the usury argument, most MFI's will tend to invest to a more profitable and risk free non funded income ventures like treasury bills. The findings in this is a confirmation of this argument, this however, is not a good news to borrowers as MFI's as tend to shrink lending in favour such investments, adverse selection and stringent measures are therefore employed to identify the qualified borrowers, those with unidentifiable credit worthiness and risk are denied access, with the reduced credit supply, the cost of delinquencies associated with lending also deceases with the decreasing credit supply.

Recommendation.
In view of the findings, the explanatory variables for financial market frictions and flight to quality significantly affects cost of credit in the long run, based on the findings, this study recommends that pegging interest rates on Central Bank rate is good as it protects unsuspecting individuals from being exploited by the MFIs, however, it also comes with usury charges to cover their administrative and operating costs, as such the Government should incorporate MFIs opinions and views in a way that will allow them charge interests which are neither high or low but enough cover their costs to remain in business, MFIs are also advised to invest in non funded income to maximize their profits. It is also prudent for MFIs to invest in online and mobile lending in order to reduce administrative and operating costs.