An Empirical Study on the Determinants of Dollarization in Cambodia * Socheat CHIM Graduate School of Economics, Osaka University 1-7 Machikaneyama, Toyonaka, Osaka, 560-0043, Japan E-mail: chimsocheat3@yahoo.com Abstract This paper provides an empirical study on the dollarization phenomenon for the case of Cambodia. Unlike the previous studies, we use Autoregressive Distributed Lag (ARDL) approach to cointegration to examine the determinants of dollarization. From the estimation results, we find that there exists a cointegration or long-run relationship among variables in consideration. However, the result shows that inflation is not the main determinant of dollarization in both short-run and long-run. Furthermore, Cumulative Sum of Recursive Residuals (CUSUM) and Cumulative Sum of Squares of Recursive Residuals (CUSUMSQ) tests provide the evidence supporting the stability of our estimated model. Finally we investigate the causality direction between the dollarization and the exchange rate by adopting the Granger causality test. The results indicate the uni-directional of causal relationship running from the exchanger rate to dollarization ration. Keywords: Dollarization; Currency Substitution; ARDL; Cointegration; Cambodia JEL Classifications: F31; F41; C13 * I would like to thank Prof. Shinji Takagi, Sovannroeun Samreth and Theara Horn for their helpful comments and suggestions. However, all the remaining errors that may appear in this paper are solely my own responsibility. The financial supports from Japanese government (Monbukagakusho, MEXT) are also gratefully acknowledged. 1
1. Introduction After decades of war and the transformation from the plan to market economy, in the early of 1990s, Cambodia started to rebuild the social and economic foundation by receiving grant aid from foreign countries and allowing foreign direct investment to flow into the country. With these currents, massive foreign currencies are seen circulating in Cambodian financial market, mostly in US dollar. In the market, due to the lack of trust in local currency (i.e. Riel), US dollar is also accepted for transaction. After the inflow of FDI, foreign aids and hyperinflation in the early 1990s, exchange rate was seen depreciated remarkably from around 1000 Riel per US dollar in 1992 to 3800 Riel in 1999 (IFS, 2007). The dollarization in Cambodia starts to increase drastically since early 1990s. The ratio of foreign currency deposit to money supply (M2) increases from approximately 50% in 1995 to 70% in 2006 (data generated by author from IFS, 2007). It is still skeptical that this high depreciation of exchange rate and hyperinflation can be the causes of the lack of trust on local currency. Dollarization makes the central bank impossible to exercise monetary and exchange rate policy (see Zamaróczy and Sa, 2002). Therefore, the study of dollarization is of interest due to the fact that it can 2
provide the policy makers with the insight of dollarization mechanism that can help them in shaping policy to de-dollarize the economy. Although Cambodia now is experiencing the high degree of dollarization, to the best of our knowledge, only few published works, such as Kem (2001), Zamaróczy and Sa (2002), Kang (2005) and Samreth (2008), can be found concerning the study on dollarization for the case of Cambodia. Zamaróczy and Sa (2002) indicate that Cambodia has a very high degree of dollarization and describe the cost and benefit of dollarization. The cost of dollarization includes loss of seigniorage, lower international reserve, loss of effective monetary policy and loss of exchange rate policy. The benefits include isolation from exchange rate fluctuation, economic and financial integration and fiscal discipline. Kem (2001) investigates the currency substitution (dollarization) in Cambodia by using OLS and shows that there is long-run relationship between exchange rate and US dollar holdings. Kang (2005), like Zamaróczy and Sa (2002), provides the explanation of cost and benefit of dollarization to Cambodia economy and suggests the de-dollarization method. The main reason is the huge loss of seigniorage which reaches approximately US$ 681 millions and the loss of monetary and exchange rate policy. Samreth (2008) provides the empirical study 3
on currency substitution (dollarization) in Cambodia. He uses a model of money-in-the-utility function. His empirical result shows that the elasticity of substitute between domestic and foreign currency is very high. His analysis shows that high degree of currency substitution limits the ability of government to raise seigniorage revenue. In this paper, the difference from the previous studies is that we study the effect on both short-run and long-run effect of exchange rate and inflation on dollarization by benefiting from Autoregressive Distributed Lag (ARDL) method. Furthermore, unlike Kem (2001), the stability test, Cumulative Sum of Recursive Residuals (CUSUM) and Cumulative Sum of Squares of Recursive Residuals (CUSUMSQ), of estimated model are also conducted in this study. Finally, we investigate the relationship between dollarization and inflation and exchange rate by applying Granger causality. This paper is organized as follows. Section 2 is Estimation model. Section 3 presents the empirical analysis, in particular, the estimation model and methodology process. The estimation results are provided in section 4. Finally, in section 5 some conclusions are drawn. 4
2. Estimation Model In recent years, there are many literatures focusing on the study of dollarization in developing countries using various estimation models. Ize and Yeyati (2003) present the portfolio model of dollarization which includes exchange rate and inflation. In the model there are risk averse depositor and borrower who can choose both currencies. The depositors maximize return from his portfolio comprising domestic and foreign currency deposit by hedging against inflation and exchange rate depreciation. His finding shows that dollarization is the function of exchange rate and inflation. He states that portfolio dollarization increase with inflation volatility and decrease with volatility of exchange rate depreciation. Honohan (2007) states that the impact of depreciation of exchange rate has long-run impact on dollarization ratio. He uses data of all dollarization countries from 1993 to 2004. The error-correction process is employed. He finds that exchange rate depreciation increases dollarization share in the short-run. Özcan and Us (2007) analyze the dollarization in Turkish economy by analyzing Granger causality between dollarization and inflation volatility, output growth, exchange rate depreciation volatility and expected depreciation. The result shows that there is two-way causality between 5
dollarization and inflation volatility. Yinusa (2007) investigates the relationship between nominal exchange rate volatility and dollarization in Nigeria by applying Granger causality test for the period 1986 2003. Result of Granger causality test supports a bi-directional relationship. He also states that exchange rate depreciation makes domestic currency share lower and increases the share of foreign currency in portfolio balance. On the other hand, dollarization also makes exchange rate volatile. As mentioned above, many previous studies on the dollarization or currency substitution use exchange rate and inflation as the key factors to explain the dollarization in their empirical model. Following these, in this paper, we also apply the model of dollarization with exchange rate and inflation. In particular, our empirical model can be expressed as: FCD f ( EX, Inf ) M =, (1) 1 where FCD stands for foreign currency deposits in banking system. M 1 is the sum of domestic currency outside bank and demand deposits. EX is exchange rate, measured by Cambodian Riels per US dollar and Inf is inflation. 6
The reduced from of equation (1) for empirical study is: FCD = + + + ( ) ln b0 b1 ln EX b t 2Inf ut M1 t, (2) where, ln is logarithmic expression; u t is the error term. Additionally, we also include the dummy variable to take into account the political instability in 1997 and 1998. Equation (2) can be rewritten as: FCD = + + + + ( ) ln b0 b1 ln EX b t 2Inf b3 Dummyt ut M1 t, (3) where Dummy t =1 for 1997:07-1997:08 and 1998:06-1998:12; Dummy t =0 elsewhere. We expect b1 and b 2 to be positive because when currency depreciates dollarization ration is expected to increase. Specifically, Riel will lose its value and agent will choose to hold dollar; we expect b 3 to be negative because political upheaval makes depositors fear and withdraw their foreign currency deposits from banks and other financial institutions. 7
3. Estimation Method and Data 3.1. Estimation Method Cointegration is the study to find long-run relationship between non-stationary variables. When conducting cointegration technique based on the approach suggested by Engel and Granger (1987) or the maxim likelihood-based approach proposed by Johansen and Juselius (1990) and Johansen (1992). Although their popularity, since they require that all variables have the same order of integration, they cause difficulty in practice. Due to this, Pesaran et al. (2001) proposed a cointegration approach known as ARDL approach that does not require the condition of the same integration orders of the variables in consideration. Thus, the unit root test is not required when adopting this approach. For this advantage, ARDL are popular recently. In this paper we also adopt ARDL approach to cointegration technique as the methodology. The specification of dollarization as a function of exchange rate and inflation in equation (3) can be written as unrestricted error correction representation below. 8
n n n FCD FCD ln = α+ ϕi ln + βi ln( EX) + γ t i i Inf t i M1 t i= 1 M1 t i i= 1 i= 1, (4) FCD + λ ln + λ ln( EX) + λ Inf + λ Dummy + ε 1 2 t 1 3 t 1 4 t t M1 t 1 To conduct the ARDL approach, as first step, F-test is conducted to verify the existence of the long-run relationship or the null hypothesis of no cointegration, H 0 : λ1 = λ2 = λ3 = 0, is tested against its alternative, H 1 : λ1 λ2 λ3 0, 0, 0. The critical values of the F-statistics of this test are available in Pesaran and Pesaran (1997) and Pesaran et al (2001), in which there are two sets of critical values computed with the assumption that all the variables in ARDL model are I(1) and I(0). If the F-statistic computed in the test is higher than the upper bound of critical values, the null hypothesis of no integration is rejected. If it is below the lower bound of critical values, the null hypothesis cannot be rejected and if it lies between the lower and upper bounds, the result is inconclusive. If the result of F-statistic in the first step supports the existence of cointegration relationship between variables, next, we select lag orders of the variables based on Akaike Information Criteria (AIC), then short-run and long-run models are estimated. Additionally, serial correlation, functional form and 9
heteroscedasticity tests are presented. To confirm the stability of estimated model, we also conduct CUSUM and CUSUMSQ tests. Since the information on the direction of causality relationships is useful for policy implication, a causality test has to be conducted to specify it. In this paper we adopt VAR (Vector Autoregression)-based Granger causality test as our methodology. 3.2. Data We use Monthly data from International Financial Statistics (IFS) published by International Monetary Fund (IMF) for our analysis. The sample covers the period from January 1995 to February 2006. The exchange rate is calculated as Riels per US dollar; end of period series are used. Inflation is calculated using CPI. Due to the fact that there is no data of US dollars outside banking system, dollarization ratio is proxied by the share of foreign currency deposit and Cambodian money supply, M1, which consists of Cambodia Riel circulating outside banks plus demand deposit. 1 1 Money supply M2, which is the sum of M1 and time and saving deposits are not considered in this paper because most of time and saving deposits in Cambodia are US dollars. 10
4. Estimation Results First, we present the results of the F-statistics to show whether there is cointegration or long-run relationship between variables. Table 1 provides the result of the F-statistics according to various lag orders. As we can see from the result of Table 1, all of the values of F-statistics are above the upper bounds of critical values of standard significant levels provided by Pesaran and Pesaran (1997) and Pesaran et al. (2001). These values support the existence of cointegration or long-run relationship between variables. Next, we estimate equation (4) and use AIC to justify the lag orders of each variable. The maximum lag order is set to 12. Using Microfit 4.0, based on AIC, ARDL(5,12,12) is obtained. Table 2 presents the ARDL estimation results. The table indicates that the overall goodness of fits of the estimated equations are very high with the result 2 R =0.9832. From the diagnostic tests, we can see that the model passes all tests, implying that our estimation model is appropriate. Based on the specified ARDL model, the short-run and long-run results are also estimated in this step. Table 3 provides the estimation results of error correction representations or the short-run model. From the table, it is evident that the error correction term ( ECt 1 ) has the correct sign and statistically significant. 11
The value of ECt 1 is small in absolute value, 0.1020, indicating the low speed of adjustment to equilibrium when there is short-run shock. To confirm the stability of the estimated model, the tests of CUSUM and CUSUMSQ are employed in this study. Figures 1 and 2 indicate that the plot of CUSUM and CUSUMSQ are completely stable within 5% of critical bands. Thus, we can conclude that the estimated model is stable. The results of long-run estimation of our analysis are presented in Table 4. They indicate that the signs of the estimated coefficients of ln( EX ) and Dummy variables are respectively positive and negative as expected. The result of negative effect of dummy variable should not be surprising because, as previously mentioned, political upheaval makes depositors fear and withdraw their foreign currency deposits from banks and other financial institutions. This leads to the decrease of FCD. For the estimated coefficient of inflation, it is M 1 negative but not significant. This result shows that inflation seems not to be a determinant of dollarization, although it is a main cause of dollarization at the outset of dollarization phenomenon in Cambodia. As for the estimated coefficient of exchange rate, it is statistically significant and positive, suggesting that when Riel depreciates against dollar the dollarization ratio will increase. 12
In sum, from the above results, it is evident that a stable cointegration relationship exists among the variables in consideration. This also implies that there is causality relationship between variables. However, to specify the causality direction, the Granger causality test has to be conducted. We adopt VAR-based Granger causality test as methodology. To apply VAR-based Granger causality test, first we have to conduct the unit root test to specify the integration order of the variables in consideration. In this paper, Augmented Dickey-Fuller (ADF) test is employed. The results are presented in Table 5. The results indicate that only dollarization ratio does not have unit in level but with only 10% significant level. Other variables, exchange rate and inflation are both non-stationary at level. All variables are stationary at first difference. Therefore, we can consider that all variables are I(1). Next, lag length selection for VAR estimation should be tested. The optimal lag length, suggested by various information criteria, is shown in Table 6. Based on AIC, the optimal lag length is shown to be 11. Based on the results of unit root test and lag length selection test, the VAR-based Granger causality test is carried out and its results are illustrated in table 7. The results indicate the uni-directional causality running from exchange 13
rate to dollarization ratio. We also confirm the result that inflation does not Granger cause dollarization. 5. Conclusion This paper investigates the determinants of dollarization in Cambodia, using Autoregressive Distributed Lag (ARDL) approach to cointegration. Cumulative Sum of Recursive Residuals (CUSUM) and Cumulative Sum of Squares of Recursive Residuals (CUSUMSQ) are also employed to test the stability of the estimated model. Since the information on the direction of causality relationships is useful for policy implication, in this paper, we also conduct the Granger causality test. VAR (Vector Autoregression)-based approach is adopted as our methodology. From estimation results, it is evident that there exists a stable cointegration or long-run relationship between variables in consideration. Although the short-run and long-run results of ARDL show that the exchange rate is the main determinant of dollarization, it seems that the inflation has no significant effect on dollarization process in Cambodia. However, the insignificant result of inflation coefficient should be interpreted with caution due to the fact that inflation played 14
an important role at the outset of dollarization phenomenon in Cambodia. The results of VAR-based Granger causality test indicate that there is uni-directional causality relation running from exchange rate to dollarization ratio. Although how to de-dollarize Cambodian economy is beyond the scope of our study, our finding results could provide some implications for policy makers in designing the de-dollarization policy. 15
References [1] Engle, R. F. and Granger, C. W. J. (1987) Co-integration and Error Correction: Representation, Estimation, and Testing, Econometrica 55: pp. 251-276. [2] Honohan, P. (2007) Dollarization and Exchange Rate Fluctuations, World Bank Policy Research Working Paper 4172. [3] Ize, A. and Yeyati, E. L. (2003) Financial Dollarization, Journal of International Economics 59: pp. 327-347. [4] Johansen, S. (1992) Testing Weak Exogeneity and the Order of Cointegration in UK Money Demand Data, Journal of Policy Modeling 14: pp. 313-334. [5] Johansen, S. and Juselius, K. (1990) Maximum Likelihood Estimation and Inference Cointegration-with Application to the Demand for Money, Oxford Bulletin of Economics and Statistics 52: pp. 169-210. [6] Kang, K. (2005) Is Dollarization Good for Cambodia?, Global Economic Review 34: pp. 201-211. [7] Kem, R. (2001) Currency Substitute and Financial Sector Developments in Cambodia, International and Development Economics, Working Paper 01-4, 16
Australian National University. [8] Özcan, K. M., and Us, V. (2007) Dedollarization in Turkey after decade of dollarization: A myth or reality, Physica A: Statistical Mechanics and its Applications 385: pp. 292-306. [9] Pesaran, M. H., and Pesaran, B. (1997) Microfit 4.0 (Window Version). New York: Oxford University Press. [10] Pesaran, M. H., Shin, Y. and Smith, R., J. (2001) Bound Testing Approaches to Analysis of Level Relationships, Journal of Applied Econometrics 16: pp. 289-326. [11] Samreth, S. (2008) Currency Substitution and Seigniorage-Maximizing Inflation: The Case of Cambodia, Applied Economics, forthcoming. [12] Yinusa, D. O. (2007) Between Dollarization and Exchange Rate Volatility: Nigeria s Portfolio Diversification Option, Journal of Policy Modeling 30: pp. 811-826. [13] Zamaroczy, M. and Sa, S. (2002) Economic Policy in a Highly Dollarized Economy: The Case of Cambodia, IMF working paper No02/92. 17
Tables and Figures Table 1: F-statistics of ARDL Test, 10%CV [2.711, 3.800], 5%CV [3.219,4.378], 1%CV [4.385,5.615] Lag order 1 2 3 4 5 6 F-statistics 4.1438* 4.8530** 5.0559** 5.9260*** 5.8377*** 5.6720*** lag order 7 8 9 10 11 12 F-statistics 4.7098** 4.8958** 5.2642** 6.7183*** 8.3181*** 8.2986*** Notes: The asterisks *** and ** are 1% and 5% of significant levels. Table 2: Autoregressive Distributed Lag Estimation Results (Dependent Variable: ln( FCD / M 1) Variable ( ) ( ) ( ) ( ) ( ) AIC-based ARDL(5,12,12) ln FCD / M1 t 1 0.8308(0.0988)*** ln FCD / M1 t 2 0.1934(0.1204) ln FCD / M1 t 3-0.0707(0.1206) ln FCD / M1 t 4-0.2617(0.1206)** ln FCD / M1 t 5 0.2062(0.0931)** ln( EX ) t 0.8709(0.3290)*** ln( EX) t 1-0.4483(0.5353) ln( EX) t 2-1.3537(0.5428)** ln( ) t 3 EX 0.6245(0.5587) ln( ) t 4 EX 0.7183(0.5569) ln( ) t 5 EX -0.3016(0.5463) ln( ) t 6 EX -1.2726(0.5386)** ln( ) t 7 EX 1.5276(0.5565)* ln( ) t 8 EX -1.0228(0.5827)* ln( ) t 9 EX 0.8039(0.6069) ln( ) t 10 EX -0.3569(0.5778) ln( ) t 11 EX 1.0949(0.4995)** ln( ) t 12 Inf -0.4713(0.4830) t EX -0.8157(0.3123)** Inft 1 0.1909(0.3825) Inft 2 0.3439(0.3504) Inft 3 0.0595(0.3608) Inf 0.0096(0.3622) t 4 Inft 5 0.3518(0.3480) 18
Table 2 (Continued): Autoregressive Distributed Lag Estimation Results (Dependent Variable: ln( FCD / M 1) Variable AIC-based ARDL(5,12,12) Inft 6-0.3786(0.3280) Inft 7 0.1866(0.3521) Inft 8-0.0483(0.3451) Inft 9-0.1155(0.3408) Inft 10 0.2893(0.3526) Inft 11 0.09745(0.3428) Inft 12-1.2106(0.3335)*** Constant -0.4675(0.2887) Dummy -0.0774(0.0271))** 2 R 0.9832 DW-statistics 1.9135 SE of regression 0.0393 Serial Correlation: F(12,77)=1.2987[0.237] Diagnostic Tests Functional Form: F(1,88)=0.0431[0.836] Heteroscedasticity: F(1,120)=0.0375[0.8471] Notes: 1. *, ** and *** are respectively the 10%, 5% and 1% of the significant level 2. The number in parentheses are standard errors 3. The number in the brackets are p-value of the test. Table 3: The Error Correction Representation for the Selected ARDL Model (Dependent variable: ln( FCD / M 1) Variable AIC-based ARDL(5,12,12) ln FCD / M -0.0671(0.0928) ( 1 ) t 1 ( / M 1 ) t 2 ( / M 1 ) t 3 ( / M 1 ) t 4 ln FCD 0.1263(0.0863) ln FCD 0.0555(0.0956) ln FCD -0.2062(0.0931)** ln(ex ) t 0.8709(0.3290)*** ln( EX ) 0.3540(0.3344) t 1 ln( EX ) -0.9996(0.3463)** t 2 ln( EX ) -0.3750(0.3451) t 3 ln( EX ) 0.3432(0.3511) t 4 ln( EX ) 0.04165(0.3309) t 5 ln( EX ) -1.2310(0.3299)*** t 6 ln( EX ) 0.2966(0.3633) t 7 ln( EX ) -0.7261(0.3224)** t 8 19
Table 3 (Continued): Correction Representation for the Selected ARDL Model (Dependent variable: ln( FCD / M 1) Variable AIC-based ARDL(5,12,12) ln( EX ) t 9 0.0777(0.3787) ln( EX ) t 10-0.2792(0.3076) ln( EX ) t 11 0.8157(0.3123)** Inf t -0.4713(0.3840) Inft 1 0.9933(1.5557) Inft 2 1.3373(1.4521) Inft 3 1.3968(1.3361) Inft 4 1.4065(1.2047) Inft 5 1.7583(1.0890) Inft 6 1.3797(0.9772) Inft 7 1.5664(0.8560)* Inft 8 1.5181(0.7497)** Inft 9 1.4025(0.6345)** Inft 10 1.1132(0.4991)** Inft 11 1.2106(0.3335)*** Constant -0.4675(0.2887) Dummy -0.0774(0.0271)** ECt 1-0.1020(0.0242)*** 2 R 0.6142 ECt 1 = ln( FCD / M1) t 1-0.6714 ln( EX ) t 1 +12.4860 Inft 1 +0.75Dummy+4.5831C Notes: 1. *, ** and *** are respectively the 10%, 5% and 1% of the significant level 2. the number in parentheses are standard errors 3. The number in the brackets are p-value of the test. Table 4: Long-Run Estimation Result (Dependent Variable: ln( FCD / M 1) Variable AIC-based ARDL(6,12) ln( EX ) t 0.6714(0.2842)** Inf t -12.4860(17.1488) Dummy t -0.7594(0.2530)** Constant -4.5831(2.3595)* Notes: 1. *, ** and *** are respectively the 10%, 5% and 1% of the significant level. 2. The number in parentheses are standard errors. 20
Table 5: Unit Root Test Variables Level First Difference Intercept Intercept Intercept Intercept and Trend and Trend ln( FCD / M 1) -1.3651-3.1595* -12.4585*** -12.4185*** ln( EX ) -2.3208-1.6928-2.8921** -3.3220* Inf -1.7927-2.0198-9.3996*** -9.3404*** Notes: The asterisks *, ** and *** are 10%, 5% and 1% of significant level respectively. Table 6: VAR Lag Order Selection Criteria Lag LogL LR FPE AIC SC HQ 0 829.7273 NA 2.33E-10-13.66491-13.5956-13.63676 1 866.5438 71.19891 1.47E-10-14.12469-13.84742* -14.01208 2 887.4005 39.30033 1.21E-10-14.32067-13.83545-14.12360* 3 894.3807 12.8066 1.25E-10-14.28728-13.59411-14.00576 4 903.6715 16.58523 1.25E-10-14.29209-13.39097-13.92611 5 913.7811 17.54557 1.23E-10-14.31043-13.20136-13.85999 6 925.0499 18.99856 1.19E-10-14.34793-13.03091-13.81304 7 931.9175 11.23804 1.24E-10-14.31269-12.78771-13.69334 8 942.0865 16.13578 1.22E-10-14.33201-12.59908-13.6282 9 953.107 16.94065 1.19E-10-14.3654-12.42453-13.57714 10 959.2732 9.172874 1.26E-10-14.31857-12.16973-13.44584 11 987.9844 41.28725* 9.18E-11* -14.64437* -12.28759-13.68719 12 993.2759 7.346821 9.9E-11-14.58307-12.01834-13.54144 Notes: 1. * indicates lag order selected by the criterion. 2. LR: sequential modified LR test statistic (each test at 5% level). 3. FPE: Final prediction error. 4. AIC: Akaike information criterion. 5. SC: Schwarz information criterion 6. HQ: Hannan-Quinn information criterion. 21
Table 7: VAR Granger Causality Test Null Hypothesis 2 χ Probability Result ln EX does not Granger cause ln( FCD / M1) 33.0413 0.005 H 0 rejected Inf does not Granger cause ln( FCD / M1) 17.2675 0.1002 H 0 accepted ln( FCD / M1) does not Granger cause ln EX 14.9477 0.1849 H 0 accepted Inf does not Granger cause ln EX 22.7399 0.0192 H 0 rejected ln( FCD / M1) does not Granger cause Inf 21.6654 0.0271 H 0 rejected ln EX does not Granger cause Inf 15.5718 0.1578 H 0 accepted Figure 1: Plot of Cumulative Sum of Recursive Residuals (CUSUM) 30 25 20 15 10 5 0-5 -10-15 -20-25 -30 1996M1 1996M11 1997M9 1998M7 1999M5 2000M3 2001M1 2001M11 2002M9 2003M7 2004M5 2005M3 2006M1 The straight lines represent critical bounds at 5% significance level 22
Figure 2: Plot of Cumulative Sum of Squares of Recursive Residuals (CUSUMSQ) 1.5 1.0 0.5 0.0-0.5 1996M1 1996M11 1997M9 1998M7 1999M5 2000M3 2001M1 2001M11 2002M9 2003M7 2004M5 2005M3 2006M1 The straight lines represent critical bounds at 5% significance level 23