WPS'2L(O POLICY RESEARCH WORKING PAPER 2186 Contagion, Bank Lending A positive historical shock to external spreads can lead to Spreads, and O utput an increase in domestic Fluctuations spreads and a reduction in the cyclical component of output. Shocks to external Pierre-Richard Agenor spreads immediately after the Joshua Aizenman Mexican peso crisis had a Alexander Hoffmaister sizable effect on movements in output and domestic interest rate spreads in Argentina. The World Bank World Bank Institute Economic Policy and Poverty Reduction September 1999 POLIcy RESEARCH WORKING PAPER 2186 Summary findings Ag6nor, Aizenman, and Hoffmaister study how Using generalized impulse response functions, theN contagion affects bank lending spreads and fluctuations show that a positive historical shock to external spreads in output in Argentina. leads to an increase in domestic spreads and a reduction They analyze what determines bank lending spreads in the cyclical component of output. when verification and enforcement costs for loan Historical decompositions indicate that shocks to contracts are high. external spreads immediately after the Mexican peso They present estimates of a vector autoregression crisis had a sizable effect on movements in Output and model that relates bank lending spreads, the cyclical domestic interest rate spreads in Atgentina. component of output, the real bank lending rate, and the spread in external interest rates. This paper sa product of Economic Policy and Poverty Reduction, World Bank Institute - is part of a larger effort in the institute to analyze the real effects of financial sector inefficiencies. Copies of the paper are available free fro the World Bank, 1818 H Street NW, Washington, DC 20433. Please contact Tanya Shiel, room J4-282, telephone 202-473-6317, fax 202-676-9810, Internet address tshiel@worldbank.org. Policy Research Working Papers are also posted on the Web at http:,/www.worldbank.org/html/dec/Pu.blications/Workpapers/home.hcml. The authors may be contacted at pagenor@worldhank.org or j.aizenman@dartmouth.edu. September 1999. (31 pages) The Policy Research Working Paper Series disseminates the findings of work in progress to encorage the exchange of ideas about development issues. An ohjective ofthe series is to get the findings out quickly, even if the presentations are less than (esly polished. The papers carry the names of the authors and should he cited accordingly. The findings, interpretations, and conclusions expressed re pahper are entirely those of the authors. They do not necessarily represent the view of the World Bank, its Executive Directors, rthe countries they represent. aPr Produced by the Policy Research Dissemination Center Contagion, Bank Lending Spreads, and Output Fluctuations P. R. Ag6nor, J. Aizenman and A. Hoffmaister* Abstract This paper studies the effects of contagion on bank lending spreads and output fluctuations in Argentina. The first part presents the an- alytical framework, which analyzes the determination of bank lending spreads in the presence of verification and enforcement costs of loan contracts. The second part presents estimates of a vector autoregres- sion model that relates bank lending spreads, the cyclical component of output, the real bank lending rate, and the external interest rate spread. Using generalized impulse response functions, a positive his- torical shock in external spreads is shown to lead to an increase in do- mestic spreads and a reduction in the cyclical component of output- as predicted by our analytical framework. Historical decompositions indicate that shocks to external spreads in the immediate aftermath of the Mexican peso crisis had a sizable effect on movements in output and domestic interest rate spreads in Argentina. JEL Classification Numbers: E44, F36, 131 *World Bank Institute; Department of Economics, Dartmouth College, and NBER; and Research Department, International Monetary Fund. We are grateful to Luis Catio for sharing his data with us, and to Martin Eichenbaum, Andrew Powell, Alejandro Werner, Peter Wickham and participants at the conference on Integration and Contagion, held in Buenos Aires (June 17-18), for useful comments. Brooks Calvo provided excellent research assistance. The views expressed here do not necessarily represent those of the Fund or the World Bank. 1  1 Introduction Argentina faced a severe economic downturn in 1995 and early 1996. Output, domestic credit, and stock prices fell dramatically. A massive shift away from peso-denominated deposits was associated with large capital outflows, a sharp drop in official foreign reserves and a contraction of the monetary base. Unemployment peaked at almost 19 percent in May 1995 and remained high in subsequent months. The liquidity crunch led to a sharp rise in bank lending rates, on both peso- and US dollar-denominated loans. At the same time, the spread between the lending rates on peso- and US dollar-denominated loans widened significantly between February and May 1995 (as shown in Figure 1), reflecting an increase in the perceived risk of a collapse of the currency board regime introduced in 1991 and a subsequent large exchange rate depreciation. The spread between deposit and lending rates, both in pesos and in US dollars, also increased sharply. The timing and severity of the economic downturn in Argentina was asso- ciated with an adverse external financial shock-an abrupt change in market sentiment regarding the country's economic prospects, triggered by expec- tations that the currency board regime would collapse. Various observers attributed this phenomenon to a contagion effect triggered by the Mexican peso crisis of December 1994. Our analysis follows this perspective and mod- els contagion as a temporary increase in the risk premium faced by domestic borrowers on world capital markets-that is, an increase in external interest rate spreads. This view is, of course, also consistent with a more general interpretation of external shocks. It is reflected in the sharp increase in in- terest rate spreads (relative to US rates) on liabilities issued by private-as well as public-borrowers from Argentina in the immediate aftermath of the Mexican peso crisis (Figure 1). The real effects of this shock are analyzed both analytically and empirically, in a model that incorporates a link between bank credit and the supply side through firms' demand for working capital (an important feature of Argentina's financial system), domestic interest rate 2 spreads, and real lending rates.' In general, spreads between lending and deposits rates in most develop- ing countries tend to be relatively large for a variety of reasons-including high required reserve ratios, a limited degree of competition in the finan- cial system, low productive efficiency of financial institutions, and selective credit and interest controls that require these institutions to undertake a substantial amount of concessionary lending. Several studies, in particular, have emphasized the role of market structure.2 In a recent empirical study of the determinants of bank spreads in Argentina, for instance, Catao (1998) found-using aggregate monthly data for the period June 1993-July 1997-- that spreads are positively influenced by the degree of market concentration.3 He interprets this result as reflecting the fact that most peso borrowers in Argentina cannot arbitrage between domestic and foreign sources of funds, and thus become subject to the monopoly power of local banks. He also found that spreads are also responsive to operating costs and the share of non-performing loans, and to a lesser degree exchange rate risk and the cost of liquidity requirements. Our analysis, by contrast, focuses on the role of external factors, in addition to default risk. In contrast to existing studies, we focus on the role of domestic interest rates in the transmission process of external shocks to output. The remainder of the paper proceeds as follows. Section II presents the 'As documented for instance by Rojas-Surez and Weisbrod (1995), banks account for between 50 and 90 percent of the financing needs of firms in Latin American countries. Ag6nor (1998), Edwards and V6gh (1997), Greenwald and Stiglitz (1993), and Isard et al. (1996) also develop models which explicitly account for the link between firms' working capital needs and bank credit. 2Among recent studies of the determinants of bank spreads are Barajas, Steiner, and Salazar (1998) for Colombia, Demirgiig-Kunt and Huizinga (1998) for a large group of countries. Early studies include Ho and Sanders (1981), and Hanson and de Resende Rocha (1986). 3Catho uses, as we do in our empirical analsis, ex ante (or contract) interest rates, rather than effective interest rates (obtained from the income statements of commercial banks). As is well known, these two measures can differ markedly in a setting where the incidence of nonperforming loans is high and refinancing operations are widespread. 3 analytical framework, which describes the determination of domestic bank lending spreads in the presence of verification and enforcement costs associ- ated with loan contracts. The analysis shows how domestic financial inter- mediation spreads are related to default probabilities, underlying domestic shocks, and external spreads. Section III estimates a vector autoregression model using monthly data for Argentina (for the period June 1993-June 1998) that relates the ex ante bank lending spread, the cyclical component of out- put, the real bank lending rate, the effective reserve requirement ratio, and the external interest rate spread. Generalized variance decompositions are discussed in Section IV. Section V uses generalized impulse response func- tions to analyze the effects of a contagious shock, defined as an increase in the external spread. Section VI assesses the movements in output and inter- est rates in Argentina in the immediate aftermath of the Mexican peso crisis of December 1994. Section VII summarizes the main results of the analysis and offers some concluding remarks. 2 The Analytical Framework The credit channel provides a key transmission mechanism of macroeconomic shocks in developing countries. This channel impacts directly on producers who finance their working capital needs via the banking system. Banks engage frequently in costly monitoring and supervision of creditors' perfor- mance, to ensure the proper use of credit, and its timely repayment. As the frequency of costly monitoring increases in turbulent times, the credit chan- nel provides a natural way to model the effects of macroeconomic shocks and volatility on economic activity in developing countries. This section out- lines a simplified version of the analytical framework developed by Ag6nor and Aizenman (1998, 1999), which highlights the impact of productivity and external cost of credit shocks on domestic output.' 4The Ag6nor-Aizenman framework combines the costly state verification approach pi- oneered by Townsend (1979) and the model of limited enforceability of contracts used in 4 We consider an economy where risk-neutral banks provide intermediation services. Agents (producers) demand credit from banks (lenders) to finance their working capital needs. Producers who lack access to the equity market rely on bank credit to finance the cost of variable inputs, which must be paid prior to production and the sale of output. Output is subject to random productivity shocks. The realized productivity shock is revealed to banks only at a cost. In the event of default by any given producer on its bank loans, the creditor seizes a fraction of the realized value of output. Seizing involves two types of costs: first, verifying the net value of output is costly; second, enforcing repayment requires costly intervention of the legal system. Future output of producer i is given by yi = M(1 + 60 + 6m + 6i), 0 <,3 < 1, Je < I < 1, ) where Mi denotes the variable input (which may consist of labor or raw materials) used by producer i, Ei is the realized i.i.d. productivity shock, 1+60 is expected productivity, and 6m is the realized common macroeconomic shock, which is assumed to distributed binomially: v probability 0.5 -v probability 0.5 The contractual interest rate on loans made to producer i is r. We assume that each producer must finance variable input costs prior to the sale of output, and that no one can issue claims on his or her capital stock. Consequently, producer i's variable costs are (1 + A )pmMi, where pm is the relative price of the variable input. We assume that the bank has information about the input choice of the producer and determines the interest rate such that the expected net repay- ment equals the cost of credit. Each bank is assumed to deal with a large the external debt literature, as in Eaton et al. (1986), Bulow and Rogoff (1989), and Helpman (1989). 5 number of independent producers, allowing the bank to diversify the idio- syncratic risk, ej. Henceforth we also assume that no default would occur in the good state of the macro shock, but that (at least) some producers will default partially in the bad state of the aggregate shock.' A producer will default if rMP + 6o - v + ei) < (1 + r')pmMj, (2) where r, is the fraction of realized output that the bank is able to seize in case of default. The left-hand side of equation (2) is the producer's repayment following a default, whereas the right-hand side is the contractual repayment. We denote by e? the highest productivity shock leading to default-that is, the value of Ej for which (2) holds as an equality: KM (1 + bo - v + ema') = (1 + r')PmMi. (3) If default never occurs, eT" is set at the lower end of the support (ETa -F). In case of default, the bank's net revenue is the producer's repayment minus the state verification and contract enforcement cost, assumed to be proportional to the cost of borrowed funds:6 ,MA4(1 + 6o + ej) - cipmMi(l + r*), (4) where 0 < ci < 1. 5The key results of our discussion hold even if this assumption is not valid. This assumption is equivalent to KMf (1 + 60 + V - F) > (1 + ri)Pm,Mi > KMi,3(1 + bc - r ), and will hold if the degree of volatility of the aggregate shock (as measured by v) is significant enough. 'The cost ci is paid by banks in order to identify the productivity shock ej, and to enforce proper payment. The analysis is more involved if some costs are paid after obtain- ing the information about ej. In these circumstances, banks will refrain from forcing debt repayment when realized productivity is below an "enforcement threshold." For simplicity of exposition, we refrain from modeling this possibility. We ignore also all other real costs associated with financial intermediation. Adding these considerations would not modify the key results discussed below. 6 We assume that banks have access to an elastic supply of funds, at a real cost of r*.' Assuming that banks are risk neutral and competitive, the contractual interest rate is determined by an expected break-even condition, derived in Appendix I. As also shown there, the contractual interest rate, rL, is determined by a mark-up rule. r' exceeds the bank's cost of funds, r*, by the sum of two terms: the first is the expected revenue lost due to partial default in bad states of nature, and the second measures the expected state verification and contract enforcement costs.8 In the particular case in which the aggregate shock follows a uniform distribution, the spread (A2) is characterized by a quadratic equation, which can be combined with (3) to derive a reduced-form solution for the probability of default and for the domestic interest rate. In general, the domestic interest rate/external cost of credit curve, plot- ted in the rL r* space, is backward-bending, and a given r* can be associated with two values of rL. This follows from the presence of a trade-off between the interest rate and the frequency of full repayment.9 The efficient point is associated with the lower interest rate, as more frequent default is associated with a lower expected surplus (see equation (A4) in Appendix I). Hence- forth we will assume that competitive banks choose the efficient point, and will ignore the backward-bending portion of the r'-r* curve. For an inter- nal solution where credit is supplied and where the probability of default is positive, the following proposition can be shown to hold: Proposition 1 A higher external cost of credit, r* raises domestic interest rates and the bank lending spread, and reduces expected output. 7This source of funds may be credit provided by foreign banks, as modeled by Ag6nor and Aizenman (1998). 8Appendix I also derives the producer's expected net income, and indicates that the optimal level of use of the variable input, Mi, is found by maximizing that expression. 9A higher interest rate would increase the probability of default, implying that the net effect of a higher interest rate on the expected repayment is determined by elasticity considerations. 7 As discussed in Appendix I, the magnitude of these effects increases with the responsiveness of the domestic interest rate to the cost of funds for banks, Or'/Or*, and are maximized as we approach the backward-bending portion of the supply of credit facing producers. 3 VAR Estimation and Analysis We now apply the analytical framework developed above to an analysis of Argentina's experience in the immediate aftermath of the 1994 Mexican peso crisis. The model's explicit account of the role of external financial shocks in the determination of domestic interest rates and output makes it particularly suitable for that purpose. To implement our framework empirically we use vector autoregression (VAR) techniques and focus on the following variables: the external interest rate spread, ES, the domestic interest rate spread on peso-denominated assets and liabilities, DS, the real lending rate, RL, and two alternative measures of output: deviations of current output from its trend level, ln(y/yT), and the growth rate of output, ln(y/y-12). The trend component yT is obtained by applying the Hodrick-Prescott filter. We refer in what follows to the model with ln(y/yT) as Model A, and the one with ln(y/y-12) as Model B.o Both models are estimated with monthly data from January 1993 through June 1998. In addition to the variables listed above, we considered expanded VAR models with the average effective reserve requirement rate, in an at- tempt to control for changes in the cost of financial intermediation." Al- though reserve requirement rates did change significantly during the sample 1o Appendix II provides precise data definitions. The results of augmented Dickey-Fuller and Phillips-Perron unit root tests are mixed due to the relatively short time spun by the sample period over which they are done; the series are taken, nonetheless, to be stationary on economic grounds (see Campbell and Perron, 1991). " Of course, various other factors (such as changes in taxation of financial services) may affect domestic lending spreads, in addition to reserve requirement rates. Our analysis implicitly takes these factors as given. This assumption is appropriate to the extent that such factors fluctuate relatively little within the sample period. 8 period, the results obtained from this expanded model were not qualitatively different from the those obtained from the smaller version. Given the rela- tively short sample size, we opted to present the results based on the more parsimonious versions of the model. The number of lags included in the estimated models (as discussed in Appendix II) was set to three months. To provide empirical evidence on the analytical framework, we use gen- eralized VAR techniques that are based on reconsidering what impulse re- sponse functions (IRFs) and variance decompositions are meant to uncover. Koop, Pesaran, and Potter (1996) and Pesaran and Shin (1998) argue that the notion of IRFs (and variance decompositions) should be re-examined, and proposed to change the focus from "pure" structural shocks identified by orthogonalizing VAR innovations, to an understanding of what a histor- ical innovations suggests about the dynamics in the data. Typically, these historical innovations are not orthogonal but, contrary to standard VAR re- gression residuals (or innovations), they embody the full information of the contemporaneous correlation of these residuals. This makes them particu- larly useful for studying the implications of the analytical framework pre- sented above; using standard VAR techniques would require in the present context arbitrary timing assumptions about the contemporaneous interac- tions between the variables. Although our model does suggest that shocks to external spreads should be considered first in the ordering of the variables in a standard VAR system, it does not lend itself to a natural ordering for the remaining ones, because they are jointly determined. Moreover, as noted below, the historical shocks to the external spread are numerically equiva- lent to the standard VAR innovations when the external spread is placed first in the ordering. A further consideration is that generalized IRFs (and generalized variance decompositions) are unique, that is, not subject to or- dering/compositional effects of standard orthogonal analysis. To illustrate generalized VAR analysis, consider the moving average rep- resentation of the a VAR model: 9 Zt = C(L)-'pt, (5) where Zt = [ESt, DSt, RLt, ln(yt/YT)] and [tt is distributed (multivariate) normal, that is, N(O, Q). This implies that Zt is also normal with zero mean and covariance matrix C(L)-'QC(L)-'. Rather than orthogonalizing the VAR innovations in equation (5), generalized VAR analysis considers the conditional expectation of Zt given a specific shock to pt. For the sake of argument, consider the generalized (or average) effect on Zt of the historical shock that is of particular interest in this study, a shock to the ESt, specifically YES,t. This effect is obtained by taking the expectation of equation (5) conditional on the shock I'ES,t = GIR(Zt, PESt = v) = E[Zt I PES,t [= C(L)-E[y I AES,t = and given the properties of the multivariate normal distribution: GIR(Zt, PES,t = v) = C(L) -ES ' UES,ES 'V, (6) where QES is the column of Q corresponding to ES, and OES,ES is the variance of the innovation in ES. Note that although v could be any value, it seems natural to set it equal to its historical value: the standard error of the ES shock, c/2,ES In general, the GIR in equation (6) will differ from the standard impulse responses. However, the GIR in (6) will be numerically equivalent to the Choleski decompositions when the ES is placed first in the ordering, or in the special case when the innovations in pt are mutually orthogonal. Aside from these numerical equivalencies, generalized VAR analysis is a conceptually different construct. Generalized VAR analysis is intended to reveal to the analyst how the VAR model behaves following a specific historical shock. Likewise, generalized variance decompositions are intended to provide the "share" of the movements of a specific series associated with historical shocks. Neither generalized IRFs nor generalized variance decompositions intend to 10 uncover the effect of "structural" shocks as in standard VAR analysis, and thus historical shocks are not orthogonal. In this sense, generalized VAR analysis provides "stylized facts" about the VAR model that fully accounts from the dynamics and historical correlations present in the data, that in turn can be compared to the predictions of the analytical framework discussed above. Note that the fact that the generalized shocks are not orthogonal implies that the variance decompositions do not generally add up to 100 percent. 4 Generalized Variance Decompositions Table 1 presents the generalized variance decompositions (GVDs) for the variables in the system, for both models. At a forecast horizon of less than twelve months, movements in the external spread are primarily associated with their own historical shocks for both models. At longer forecast horizons, historical shocks associated the domestic spread play a more substantial role in both cases; in addition, movements in output play a greater role in model A. These results are consistent with the analytical framework presented above and the extended framework developed by Ag6nor and Aizenman (1998), in which external spreads have an endogenous component-reflecting the probability of default of domestic producers on their liabilities to domestic banks and the risk of domestic banks defaulting on their foreign loans.12 More generally, the results obtained with Model A are also consistent with the view that domestic economic conditions affect movements in external spreads through their effect on market sentiment or expectations. At short horizons, movements of domestic spreads are also greatly influ- enced by their own historical shocks in both models. At a forecast horizon of '2Note also that the analytical framework predicts that shocks to output and domestic spreads are correlated, because the latter variable reflects the probability of default (which is itself related to output shocks). Recall, however, that since generalized VAR analysis focuses on nonorthogonal shocks, it is not valid to add up their shares to obtain a measure of their combined effect. 11 less than six months, these shocks explain the bulk of the movements in DS. At longer forecast horizons (beyond 6 months), historical shocks to the real lending rate play a greater role in explaining these movements, again in both models. In addition, this is also true for the external spread and movements in output in Model A-with each accounting for about the same share of the movements in the domestic spread at a horizon of 24 months. Again, these results are consistent with the theoretical framework discussed above and our main proposition; shocks to both external spreads and output affect the capacity of domestic firms to repay, thereby raising banks' perceived risk of default. As is the case with external and domestic spreads, movements in out- put are mostly explained by their own historical shocks at forecast horizons less than six months in both models, explaining in excess of 75 percent of its movements. At longer forecast horizons, historical shocks for both mod- els associated with the external spread play a substantial secondary role, accounting between 20 to 25 percent of cyclical movements in output at a forecast horizon of 24 months. Shocks to domestic spreads explain a similar portion of movements in output in Model B but account for about half as much as shocks to external spreads in Model A. A rather puzzling fact is that shocks to the real lending rate account for a relatively small proportion of cyclical movements in output in both models.13 One possible explanation is that our index of output (industrial production) reflects essentially output of traded goods; to the extent that producers of traded goods have a greater access to world capital markets (because of their ability to post collateral in foreign currency terms), one would expect a limited effect of the cost of borrowing on domestic capital markets. Movements in the real lending rate in both models are greatly influenced by their own historical shocks at all horizons, explaining in excess of 80 per- 13We attempted to measure the real lending rate by using various proxies for the ex- pected inflation rate (lagged, current and one-period ahead actual values). This did not change significantly our results. 12 cent of its movements. Shocks to domestic spreads and to the real lending rate play a secondary and tertiary role respectively in accounting for movements in the real lending rate, accounting respectively about 15 and 10 percent of the movements in the real lending rate and shocks to the cycle. 5 External Spread Shock Figures 2 and 3 show the generalized impulse responses (GIRs) for the vari- ables of the two models to a positive historical shock in the external spread. As discussed in the introduction, this experiment can be viewed as one way of capturing "pure" (expectations-related) contagion effects, triggered by events taking place elsewhere in the region or the world. Of course, as also noted earlier, a more general interpretation of this experiment is possible; it can be viewed simply as reflecting an adverse external financial shock-related or not to contagion. GIRs and their one-standard error bands are shown for each variable.'4 As indicated earlier, GIRs are obtained as the conditional distribution of each shock in the system and thus provide the dynamic responses of the variables in the VAR model that accounts for all of the historical information in the data sample. As illustrated in equation (6), the GIR to an external spread shock shows the evolution of variables in the model corresponding to "historically correct" shock to external spreads that explicitly accounts for all the contemporaneous movements of the other shocks in the model. As noted in Section III, this is numerically equivalent to the traditional Choleski decomposition when the external spread "moves first," that is, when the external spread shock occurs before other shocks. "In all figures the dotted lines for the GIRs show one standard error band in each direction and are based on 1000 Monte Carlo replications. In each replication we sampled the VAR coefficients and the covariance matrix from their posterior distribution. Prom these replications we calculated the square root of the mean squared deviation from the impulse response in each direction. By construction, these bands contain the impulse response function but are not necessarily symmetric. 13 As shown in the figures, a one-standard deviation shock to external spreads of roughly 120 basis points leads in the next period to an increase in the domestic spread by only about 20 basis points in both cases. Whereas the response of the external spread lasts just over a year, the response of the domestic spread lasts for about half as long. The first finding is con- sistent with an extended version of the model presented in Section II to account for two levels of financial intermediation, along the lines of Ag6nor and Aizenman (1998). In that paper, the process of financial intermedia- tion is viewed as consisting of two stages: foreign banks provide credit to domestic banks, and domestic banks provide the intermediation services to domestic investors. The analysis shows that each spread is determined by similar considerations-it equals the expected revenue lost due to partial default, and the cost of financial intermediation, at the given level of inter- mediation. This extended model can explain the finding reported above, if the exogenous shock to the external spread indicates that the likelihood of ex- ternal default increases by more than the likelihood of internal default. This may be the case if the shock is due to contagion associated with asymmetric information-that is, if Argentina's perceived country risk by foreign lenders increased by more than the riskiness of business in Argentina for domestic lenders. Movements in output become significantly negative after 2 months and display a degree of persistence that is similar to that observed for the external spread in both cases." The response of the real lending rate is positive but 15Note that on impact in Model A, the response of the cyclical component of output is a perverse blip that is reversed very quickly. As noted in Section III, generalized IRFs embody the full information of the contemporaneous correlation in the VAR innovations and consequently and contrary to standard IRF's, none of these correlation are set to zero. Recall that generalized IRFs are numerically equivalent to traditional Choleski IRFs when the variable of interest is place first in the ordering. In this context, the perverse output blip on "impact" essentially reflects the positive contemporaneous correlation between the VAR innovations in ES and those in ln(y/yT) during the sample period (see Table Al) that is "picked up" by the generalized IRFs. Note, however, that this perverse blip is not observed with Model B. 14 imprecisely measured. The initial rise in that variable is consistent with an increase in the domestic spread that is brought about through a rise in the nominal lending rate that exceeds the rise in the nominal deposit rate, with inflation displaying some degree of inertia on impact. Alternatively, it is also consistent with a situation in which the fall in the cyclical component of output leads not only to a drop in both domestic rates (with the fall in the nominal deposit rate exceeding the fall in the nominal lending rate) but also to a drop in inflation, associated with a contraction in aggregate demand. 6 The Aftermath of the Peso Crisis: A His- torical Decomposition A useful application of the generalized VAR models estimated above is to as- sess the movements in output and domestic interest rate spreads in Argentina in the immediate aftermath of the Mexican peso crisis of December 1994. This can be done by using the historical decompositions of these variables for the period immediately following the collapse of the peso, specifically, from January 1995 to the end of 1996. Table 2 presents these results on a quarterly basis (obtained by averaging over the monthly decompositions) for both models. The results for both models indicate that the fall in output in the second quarter of 1995 (by about 3 percent with respect to trend in Model A, and by about 6 percent at an annual rate in Model B) was mostly associated with the adverse effect of higher external spreads-a result that is consistent with our analytical framework. This effect persists until the first quarter of 1996 in both cases-although the effect of shocks to output itself become important after the second quarter of 1995. Regarding the domestic spread, the conditional forecasts of the models (based on information available up to December 1994) appear to track the data fairly closely for the period under consideration. The results also suggest 15 that for the first half of 1995, external spread shocks raised the domestic spread by about 0.7 percentage points, compared to about 1.3 for domestic spread shocks. Note that during the same period, the effect of external spread shocks are twice as large as those of output shocks, and almost four times as large as shocks to the real lending rate. The relatively limited impact of external spread shocks on the domestic lending rate is consistent with the possibility that credit rationing translates into larger movements in the volume of credit, as opposed to prices. However, in the absence of disaggregated data on credit flows and pools of borrowers (based on their creditworthiness, for instance), it is hard to assess the importance of this effect. Nevertheless, it remains true that during the first part of 1995 (that is, in the immediate aftermath of the Mexican peso), external shocks had sizable effects on the behavior of output and domestic bank lending spreads in Argentina.6 7 Summary and Conclusions The purpose of this paper has been to study the effects of external shocks on domestic bank lending spreads and output fluctuations in Argentina. The analytical framework, which was presented in Section II, analyzed the determination of bank lending spreads in the presence of verification and enforcement costs of loan contracts. Section III presented estimates of a vector autoregression system that relates the ex ante bank lending spread, movements in output (measured as deviations of output from trend and the growth rate of output), the real bank lending rate, and the external interest rate spread. Generalized variance decompositions, presented in Section IV, showed in particular, that at short horizons (less than 6 months) movements of domestic spreads are greatly influenced by their own historical shocks. At ANote also that movements in output in this context can be consistent with a demand channel. Again, identifying more precisely the supply-side effects emphasized in this paper would require more disaggregated data. 16 longer forecast horizons, the external spread and the cyclical component of output played a greater role in explaining these movements. The effects of an external shock, modeled as a positive historical shock in external interest rate spreads, were analyzed in Section V using generalized impulse response functions. The results indicated that such a shock led to an increase in do- mestic spreads and a reduction in the cyclical component of output. Both results are consistent with the predictions of our analytical framework. The results also showed that the response of the domestic spread with respect to the foreign spread is well below one; we argued that this prediction is con- sistent with an extended version of the model presented here (Ag6nor and Aizenman, 1998). Finally, Section VI used the generalized VAR models to assess the effects of historical shocks to external spreads on movements in output and domestic interest rate spreads in Argentina in the immediate af- termath of the Mexican peso crisis of December 1994. The results indicated that such shocks played an important role in the behavior of these variables. The experience of the emerging markets in the nineties provides new challenges for economists, requiring us to reassess our understanding of the transmission mechanism from financial markets to real economic activity. The empirical results of our paper are consistent with the notion that finan- cial volatility has adverse consequences in economies where banks and debt contracts are widely used to finance investment. Our results provide tenta- tive support for the predictions of models based upon the notion of costly financial intermediation. Further research is needed to validate these results for other countries, and to identify their policy implications. 17 Appendix I The Bank Lending Spread and the Effect of an External Shock As noted in the text, we assume that banks have access to an elastic supply of funds, at a real cost of r*. With competitive and risk-neutral banks, the contractual interest rate is determined by the expected break- even condition:17 (1 + r*)pmM =0.5 ( + +r)pmMi + [(I + r )pmMi] f (e)de (Al) + 'i [VM (1 + 6o - v + e) - cipmMi(l +r*)]f(e)de where f(e) is the density function. Using (3) and (Al), the interest rate spread can be shown to be given by ImO axm 0.5 J [KMf (m - e)]f(e)de 0.5cipMi(1 +r*) fJ f (e)de L PmMi PmnMi (A2) The contractual interest rate, ri, is determined by a mark-up rule. r exceeds the bank's cost of funds, r*, by the sum of two terms: the first is the expected revenue lost due to partial default in bad states of nature, and the second measures the expected state verification and contract enforcement costs. The producer's expected net income equals S(1 + r')pm,Mi + f, . [(1 + r')pmMi]fl(,-)d (1 + 6)M - 0.5 + .6+ax2[K Mj]f1 )d6 (A3) Using (Al), we can simplify (A3) to (1 +o)M - (1 + r*)pmMi - 0.5cipmMi(l + r*) f (e)de. (A4) "In what follows we drop the subscript i on e to simplify notations. 18 The optimal level of use of the variable input, Mi, is found by maximizing (A4). In the particular case in which the aggregate shock follows a uniform distribution, -F < E < F, the spread (A2) is characterized by a quadratic equation, given by r r*=2F A + ci(1 + r*)4i, (A5) PmMi where 4i = (F+ EmaT)/41 is the probability of default. Combining the above equation with (3) one can infer a reduced form solution for the probability of default and for the domestic interest rate. To establish the derivations in Proposition I proceeds as follows. Using (3) and (A5), we infer that the probability of default is determined by 2FV,Mf O + jcj(1 + r*)pmM - 4KMFj 1. + (1 + r')pmM (A6) -nlMf (1 + 60 - V - F) = 0. This is a quadratic equation, yielding 2 interest rates in the relevant range. Henceforth we assume that competitive forces induces banks to offer the lower interest rate, leading to a probability of default of 4Dj = (A7) where H = 4rM/F - c,i(1+r*)pmM, Z= H2 - 8rM FA, A = (1 +r*)pmM, - nM ( (+0 bo - F). Using (A6) and (3), we infer that drl/dr* =4.MF/ Z. (A8) 19 Hence, we operate on the upward-slopping portion of the supply of credit as long as H > VY and Z > 0. We approach the backward-bending part of the curve as Z -- 0. Henceforth we assume that this condition holds. The first-order condition determining the demand for the variable input is inferred from (A4) as dH = (1 + 60)3M' - (1 + r*)pmcj[ij + Mi( )] = 0. (A9) dMi OM) Applying the implicit function theorem to (A9), and using the second order-condition for profits maximization, we infer that dM. d2I/(dxdMi)I d2II (1 dr* d2I/dMj dr-dMkIlO This result implies that, to establish that dMi/dr* < 0, it suffices to show that d21/ i(dxdMi) < 0. Applying (A9) we infer that d2II (1 + 60)3MI-I0a482 ±- f+ r*)pmc4[ + M(- )]. (All) dr*dMi 1+ r* Or* M8ar* Applying (A7), and collecting terms, it follows that - M~ [ci(H~ Z ) M(lIc2(A) = [1- + --,( NL] = A(1 + ci ( ). (Al2) Or* 41M rz 82( 1(- 0Ci)i Mi(8Z/aM.) cH ci (1 [1+ ]+ -+ eP - aMiOr* Z 2Z 4,.M 4M Mr Thus, 02.1 +M( )= ar* aMiar* M (1+ r*)c1 Ms(8Z/&M2) cH S2 +(2 -i )c +c@ [ -] _ 2Z [I+ ci V/Z_ 4rM r 2Z 4rM f Using (A7) it can be shown that Mj(aZ/I)/2Z < 1 and cjH/4Mff > ci4. Applying these 2 results to the above equation it can be verified that 'D+ M( ) 0, Or* OMOr*) 20 from which we infer that, indeed, dIll/dr*dMi < 0. An Appendix (available upon request) establishes that lower expected productivity, 6o, and higher volatility of macroeconomic shocks, v, raise domestic interest rates and the bank lending spread, and reduces expected output. 21 Appendix Il Data Sources and VAR Estimation Data. The data used in this study are at a monthly frequency and cover the period 1993:M6-1998:M6. The variables are measured as follows:18 * ES is the external spread of Brady par bonds over U.S. Treasury bills. The series is virtually indistinguishable from spreads on Brady dis- counted bonds, and its movements are highly correlated with external spread on sovereign bonds (as shown in Figure 1). Data were obtained from Merryll Lynch. * DS is calculated as the difference between the nominal lending rate on peso-denominated loans and the deposit rate on peso-denominated deposits. The series were obtained from the Fund's International Fi- nancial Statistics (line 60p and line 601) and from Catiio (1998). * RL is calculated as the nominal lending rate on peso-denominated loans at a monthly rate minus monthly inflation, measured by the consumer price index. Raw series were obtained from the Fund's International Financial Statistics. (lines 60p and 64) * In(y/yT) measures deviations of industrial output, y, from trend, yT. yT is estimated with the Hodrick-Prescott filter, using a value of A = 16000 for the smoothing parameter. ln(y/y-12) is the growth rate of output. The industrial output index was obtained from FIEL. VAR estimation. To determine the number of lags to include in the VAR models, we started by calculating standard lag-length tests, that is Akaike Information Criteria (AIC), Hannan-Quinn (HQ), and Schwarz. These 18The effective reserve requirement rate, which was used in our preliminary ex- periements, was calculated by subtracting line 14a in the Fund's International Financial Statistics from line 14 and dividing by the sum of lines 24 and 25, minus line 14a. 22 tests compare the cost of increasing the lag length (reduced degrees of free- dom) to the benefit (increased information extraction from the data). Using a maximum lag length of six, all three tests suggested using six. This presents a problem due to the size of the sample: using the six lags means that each of the five equations would contain 31 (6*5+1) coefficients to estimate with 66 monthly observations (January 1993-June 1998). This translates into unacceptably low degrees of freedom and consequently low precision in the estimation. Rather using the six lags as suggested by the tests, we use three lags based on two considerations. First, it is the smallest lag length where the reduced-form innovations are white noise judging by Ljung-Box Q tests for serial correlation (up to order 12). This ensures that the white noise assumption implicit in the estimation procedure is not violated. Second and more importantly, the GIRs and GVDs using three lags are qualitatively the same as those using six lags. Thus, using the shorter lags does not affect the main qualitatively results presented in the paper. Table Al presents a summary of the estimated VAR equations that underlie the empirical results in the paper. 23 References Ag6nor, Pierre-Richard, "Borrowing Risk and Contagion," unpublished, World Bank Institute (October 1998). Ag6nor, Pierre-Richard, and Joshua Aizenman, "Contagion and Volatility with Imperfect Credit Markets," IMF Staff Papers, 45 (June 1998), 207-35. "Volatility and the Welfare Costs of Financial Market Integration," in Financial Crises: Contagion and Market Volatility, ed. by Pierre-Richard Ag6nor, Marcus Miller, David Vines, and Axel Weber (Cambridge University Press: 1999). Barajas, Adolfo, Roberto Steiner, and Natalia Salazar, "Interest Spreads in Banking: Costs, Financial Taxation, Market Power, and Loan Quality in the Colombian Case 1974-96," Working Paper No. 98/110, International Monetary Fund (August 1998). Bernanke, Ben S., and Mark Gertler, "Agency Costs, Net Worth and Business Fluctuations," American Economic Review, 79 (March 1989), 14-31. Campbell, John Y., and Pierre Perron, "Pitfalls and Opportunities: What Macroeconomists should Know about Unit Roots," in NBER Macroeconomics Annual 1991, ed. by Olivier-Jean Blanchard and Stanley Fischer, MIT Press (Cambridge, Mass.: 1991). Catdo, Luis, "Intermediation Spreads in a Dual Currency Economy: Argentina in the 1990s," Working Paper No. 98/90, International Monetary Fund (May 1998). Demirgfig-Kunt, Ash, and Harry Huizinga, "Determinants of Commercial Bank Interest Margins and Profitability: Some International Evidence," Policy Research Working Paper No. 1913, the World Bank (March 1998). Edwards, Sebastian, and Carlos A. V6gh, "Banks and Macroeconomic Distur- bances under Predetermined Exchange Rates," Journal of Monetary Eco- nomics, 40 (November 1997), 239-78. Eichengreen, Barry, and Ashoka, "What Explains Changing Spreads on Emerging- Market Debt: Fundamentals or Market Sentiment?," unpublished, Interna- tional Monetary Fund (December 1997). Greenwald, Bruce C., and Joseph E. Stiglitz, "Financial Market Imperfections and the Business Cycle," Quarterly Journal of Economics, 108 (February 1993), 77-114. Hanson, James, and Roberto de Resende Rocha, "High Interest Rates, Spreads and the Cost of Intermediation: Two Studies," Industry and Finance Series No. 18, the World Bank (February 1986). 24 Ho, Thomas S., and Anthony Sanders, "The Determinants of Bank Interest Margins: Theory and Empirical Evidence," Journal of Financial and Quan- titative Analysis, 16 (- 1981), 581-600. International Monetary Fund, International Capital Markets: Developments, Prospects, and Key Policy Issues, International Monetary Fund (Washington DC: November 1997). Isard, Peter, Donald J. Mathieson, and Liliana Rojas-Sudrez, "A Framework for the Analysis of Financial Reforms and the Cost of Official Safety Nets," Journal of Development Economics, 50 (June 1996), 25-79. Koop, Gary, M. Hashem Pesaran, and Simon N. Potter, "Impulse Response Analysis in Nonlinear Multivariate Models," Journal of Econometrics, 74 (September 1996), 119-47. Pesaran, M. Hashem, and Yongcheol Shin, "Generalized Impulse Response Analy- sis in Linear Multivariate Models," Economics Letters, 58 (- 1998), 17-29. Rojas-Su6rez, Liliana, and Steven Weisbrod, Financial Fragilities in Latin Amer- ica: The 1980s and 1990s, Occasional Paper No 132, International Monetary Fund (October 1995). Townsend, Robert M., "Optimal Contracts and Competitive Markets with Costly State Verification," Journal of Economic Theory, 21 (October 1979), 265-93. 25 Figure 1 Argentina: Output and Interest Rates 1/ 120 Industrial output 40 Bank lending rate (deseasonalized, Dec 1994= 100) (in percent) 110 30 In U.S. dollar terms 100-- 90 20 80 In peso terns 80 10 70 60 ""0 Jan90 Jan92 Jan94 Jan96 Jan93 Jan94 Jan95 Jan96 17 -Interest Rate Spreads 33 25 External spreads 16 15 30 20 14 Brady par bonds 13 27 15 In Peso terms 12 (right scale) I 11- - 24 1 jo ' II 10 - In US Dollarterms 9 (leftscale) 21 5 8 - . 7 _____________________"_______18 8SovBrdyparbond s 7 18 0 Jun93 Jun94 Jun95 Jun96 Jul92 Jul93 Jul94 Jul95 Jul96 Jul97 Sources: FIEL; International Monetary Fund, Bloomberg, Inc., and Merryll Lynch. 1/ The vertical line corresponds to the Mexican peso crisis (December 20, 1994). Figure 2. Generalized Impulse Responses, Model A. (Historical Shock to the External Spread) External Spread Response Domestic Spread Response 3.0 ---- -- - -- - -- - --- - - -3.0- - - -- - 2.5 2.5 2.0 2.0 1.51. 0 10 DO 02 04 008 10 12 14 16 18 20 22 24 28 28 30 32 34 36 00 02 04 06 08 10 12 14 16 18 20 22 24 26 26 30 32 34 36 Output Response Real Lending Rate Response 3.0 ---- ------- - ---- - --- - ----_-__ __-_.. 0 20 0.15 2.0 0.10 1.0 0.05 0.0 0 - - ------- - -.0 -1.0 - ' -0.10 -2.0 - S - - -0.15 -3.0 -0.20 00 02 04 6 0 10 12 14 16 18 20 22 24 26 28 30 32 34 36 00 02 04 06 08 10 12 14 16 18 20 22 24 26 28 30 32 34 36 Note: The impulse responses were obtained from a VAR model with four variables: the external spread, the domestic spred, output (deviation from trend), and the real lending rate; all variables are measured in percentage points except output which is measured as the percentage deviation from trend output. The shock to the external spread equals the standard deviation of its VAR innovation, 120 basis points. The VAR model is estimated with three lags using monthly data from 1993:MI through 1998:M6. One standard error band in each direction are based on 1,000 Monte Carlo replications. See appendix for details. Figure 3. Generalized Impulse Responses, Model B. (Historical Shock to the External Spread) External Spread Response Domestic Spread Response 3.0 3.0 25 2.5 2.0 2.0 1.5 1.5 1.0 1.0 0.5 . 0.5 / 0.0 --- - 0.0 - - ---------------- -0.5 -0.5 -1.0 -1.0 00 02 04 06 08 10 12 14 16 18 20 22 24 26 28 30 32 34 36 00 02 04 08 08 10 12 14 16 18 20 22 24 2a 28 30 32 34 36 Output Response Real Lending Response 3.0 0.20 0.15 2.0 0.10 , 1,0 ** -- - --- ------ - - - - - - 0.0 - -. - - - *.-. . .-0.05 -0.10 -2.0 -0.15 -3.0 - -0.20 00 02 04 06 08 10 12 14 16 18 20 22 24 26 28 30 32 34 36 00 02 04 06 08 10 12 14 16 18 20 22 24 26 28 30 32 34 36 Note: The impulse responses were obtained from a VAR model with four variables: the external spread, the domestic spread, output (log(yt/yt-12)), and the real lending rate; all variables are measured in percentage points. The shock to the external spread equals the standard deviation of its VAR innovation, 120 basis points. The VAR model is estimated with three lags using monthly data from 1993:MI through 1998:M6. One standard error bands in each direction are based on 1,000 Monte Carlo replications. See appendix for details. Table I. Generalized Variance Decompositions Model A Model B External Spread (ES) External Spread (ES) Months Percentage of the variance associated with Percentage of the variance associated with historical shocks to historical shocks to ES DS ln(ylYT) RL ES DS In(y/y.12) RL 1 100.0 0.4 2.9 0.0 100.0 0.0 0.8 0.4 2 99.5 - 0.4 5.2 0.1 98.6 0.3 0.5 0.2 3 95.9 1.5 3.8 1.8 92.8 2.0 4.5 3.1 6 92.8 5.0 6.1 1.3 89.2 6.7 3.6 2.2 9 87.4 8.2 8.4 1.1 83.3 11.9 2.9 2.1 12 81.4 10.3 10.3 1.6 76.9 16.0 2.5 2.6 24 70.5 11.2 12.2 4.7 61.8 20.8 1.9 6.7 Cyclical Component of Output (In(y/yt)) Output Growth (In(y/y,12)) Percentage of the variance associated with Percentage of the variance associated with historical shocks to historical shocksto ES DS In(y/yr) RL ES DS ln(yVy1) RL 1 2.9 0.4 100.0 5.7 0.8 0.0 100.0 9.2 2 2.8 3.0 94.1 6.1 1.4 0.0 96.1 9.9 3 2.6 5.1 87.6 10.0 2.7 2.9 92.4 9.4 6 9.7 5.9 78.2 9.5 15.5 11.2 72.6 7.2 9 19.1 8.0 68.3 8.5 23.3 19.8 55.3 5.8 12 24.8 10.3 61.4 7.5 2S.6 25.4 43.6 5.8 24 25.2 12.5 53.0 8.6 20.9 30.1 28.0 10.7 Domestic Spread (DS) Domestic Spread (DS) Percentage of the variance associated with Percentage of the variance associated with historical shocks to historical shocks to ES DS fl(y/yT) RL ES DS In(y/y,-i2) RL 1 0.4 100.0 0.4 13.0 0.0 100.0 0.0 10.5 2 4.1 93.9 3.6 11.3 3.5 92.3 1.8 8.6 3 6.5 76.2 4.0 14.0 6.3 75.9 2.6 13.6 6 9.0 60.9 8.6 12.4 5.9 67.6 2.2 13.3 9 8.4 51.2 1l.0 14.5 5.2 61.2 1.7 15.6 12 7.4 45.9 11.9 16.2 4.5 56.9 1.5 17.8 24 11.6 40.6 11.3 18.7 5.0 50.9 1.4 21.0 Real Lending Rate (R ) Real Lending Rate (RI) Percentage of the variance associated with Percentage of the variance associated with historical shocks to historical shocksto ES DS In(yIyT) RL ES DS ln(y.12) RI 1 0.0 13.0 5.7 100.0 0.4 10.5 9.2 100.0 2 0.1 12.6 8.7 98.6 0.5 10.6 10.5 99.2 3 0.7 14.0 9.0 93.2 1.3 12.6 13.9 91.2 6 1.1 16.0 9.7 90.4 1.4 15.1 14.6 88.3 9 1.4 17.0 10.2 87.8 1.5 16.7 14.2 85.9 12 1.4 17.2 10.4 86.1 1.5 17.4 13 84.2 24 2.0 16.9 10.4 84.2 1.6 17.7 13.3 82.1 Note: These decompositions are based on the generalized VAR analysis following Koop Pesaran and Potter (1996) who propose to consider non-orthogonah historical shocks. Consequently the variance decompositions do not add up to 1O0 percent T'he variance decompositions are obtained fromn VAR models comprised by the following variables: ES, DS, In(y/yT) in Model A and ln(yttyt-12) in ModelB, and RL. The model is estimated with three lags using monthly data from 1993:M I through 1998:M6; see Appendix 11 for details. Table 2. Generalized Historical Decompositions Model A Model B Quarter Cyclical Component of Output (ln(y/yT)) Output Growth (ln(yt/yt- 12)) Actual Model Associated with historical shocks to: Actual Model Associated with historical shocks to: projection ES DS ln(Y/YT) RL projection ES DS lft(YVYt-12) RL 1995:Ql 0.064 0.015 0.016 0.003 0.036 0.011 0.043 0.028 -0.007 -0.004 0.011 0.004 1995:Q2 -0.031 -0.004 -0.035 -0.004 0.016 -0.009 -0.063 0.017 -0.048 -0.028 0.003 0.000 1995:Q3 -0.089 -0.010 -0.030 -0.013 -0.032 -0.010 -0.109 0.020 -0.042 -0.053 -0.025 -0.005 1995:Q4 -0.104 -0.009 -0.024 -0.017 -0.053 -0.004 -0.128 0.030 -0.034 -0.065 -0.055 0.004 1996:Q1 -0.029 -0.004 -0.026 -0.018 0.021 0.004 -0.075 0.043 -0.031 -0.067 -0.028 0.019 1996:Q2 0.003 0.003 -0.010 -0.025 0.035 -0.007 0.058 0.056 -0.005 -0.068 0.059 0.011 1996:Q3 -0.020 0.010 -0.007 -0.031 -0.012 0.003 0.101 0.068 0.002 -0.064 0.065 0.016 1996:Q4 -0.031 0.016 -0.001 -0.020 -0.041 0.015 0.114 0.078 0.000 -0.041 0.047 0.033 Domestic Spread (DS) Domestic Spread (DS) Actual Model Associated with historical shocks to: Actual Model Associated with historical shocks to: projection ES DS ln(y/yT) RL projection ES DS ln(yVyt-12) RL 1995:Ql 0.190 0.167 0.006 0.014 0.001 0.004 0.190 0.163 0.006 0.016 0.002 0.004 1995:Q2 0.186 0.165 0.008 0.012 0.005 0.001 0.186 0.159 0.005 0.023 -0.001 0.000 1995:Q3 0.172 0.161 0.003 0.010 0.005 -0.005 0.172 0.155 0.002 0.018 -0.001 -0.005 1995:Q4 0.162 0.156 0.002 0.013 -0.001 -0.006 0.162 0.152 0.002 0.015 0.000 -0.006 1996:Ql 0.150 0.152 -0.004 0.016 -0.005 -0.005 0.150 0.149 -0.005 0.014 0.001 -0.003 1996:Q2 0.145 0.150 -0.005 0.012 0.000 -0.010 0.145 0.147 -0.003 0.015 0.000 -0.009 1996:Q3 0.133 0.148 -0.005 -0.004 0.003 -0.011 0.133 0.146 -0.002 -0.002 -0.002 -0.010 1996:Q4 0.132 0.148 -0.005 -0.004 0.000 -0.006 0.132 0.145 -0.002 -0.005 -0.002 -0.006 Note: These historical decompositions are calculated by averaging the monthly decompositions using generalized VAR analysis. The VAR. model projections are obtained as dynamic forecasts of the models conditional on information up to December 1994. Since historical shocks are not orthogonal, the model projections and the effects associated with each historical shock do not add up to the actual series. Table Al. VAR Estimates, Monthly Observations from January 1993 to June 1998. Model A ES DS In(y/yT) RL Coefficient of Determination (R2) 0.883 0.788 0.524 0.326 Adjusted R2 0.852 0.731 0.397 0.146 Sum of Squared Errors 84.094 54.854 832.493 7.059 Standard Error of Estimate 1.367 1.104 4.301 0.396 Significance of Lagged Regressors: External Spread 64.582 * 0.810 1.494 0.111 Domestic Spread 1.474 30.049 * 1.316 1.325 Output 0.707 1.676 2.804 * 0.505 Real Lending Rate 2.148 3.596 * 1.214 3.105 * Correlation with the VAR innovations of: External Spread 1.450 0.062 0.171 0.011 Domestic Spread 0.946 0.059 0.361 Output 14.353 0.239 Real Lending Rate 0.122 Tests for Serial Correlation: Breusch-Godfrey 64.89 52.62 9.95 8.84 LJung-Box Q 91.93 97.12 54.63 56.71 Model B ES DS ln(y/y,.,2) RL Coefficient of Determination (R2) 0.902 0.776 0.714 0.339 Adjusted R2 0.876 0.717 0.637 0.163 Sum of Squared Errors 0.007 0.006 0.108 0.001 Standard Error of Estimate 0.013 0.011 0.049 0.004 Significance of Lagged Regressors: External Spread 61.986 * 1.001 1.843 0.133 Domestic Spread 1.693 33.823 * 1.708 1.422 Output 3.710 * 0.837 5.458 * 0.819 Real Lending Rate 1.487 3.116 * 0.721 2.839 * Correlation with the VAR innovations of: External Spread 1.217 0.009 -0.087 -0.064 Domestic Spread 0.996 -0.016 0.323 Output 18.555 0.304 Real Lending Rate 0.119 Tests for Serial Correlation: Breusch-Godfrey 35.59 38.15 1.34 * 18.35 Ljung-Box Q 89.21 97.03 34.00 11.58 Note: The VAR models are estimated with three lags. The significance tests are F-tests for the joint significance of all of the lags of the corresponding variable; these tests have respectively three and 53 degrees of freedom in the numerator and the denominator. The tests for serial correlation test for serial correlations of up to order 12. An asterisk (*) denotes significant rejection of the respective null hypothesis at the five percent significance level. Policy Research Working Paper Series Contact Title Author Date for paper WPS2161 Will the Real "Natural Trading Maurice Schiff August 1999 M. Kasilag Partner" Please Stand Up? 39081 WPS2162 Quantifying the Fiscal Effects Shantayanan Devarajan August 1999 H. Sladovich of Trade Reform Delfin S. Go 37698 Hongyi Li WPS2163 Coverage under Old-Age Security Estelle James August 1999 M. Leenaerts Programs and Protection for the 84264 Uninsured - What Are the Issues? WPS2164 Challenging El Salvador's Rural Maureen Lewis August 1999 M. Lewis Health Care Strategy Gunnar S. Eskeland 39080 Ximena Traa-Valerezo WPS2165 The Russian City in Transition: Martha de Melo August 1999 H. Sladovich The First Six Years in 10 Volga Gur Ofer 37698 Capitals WPS2166 Seeking Votes: The Political Norbert R. Schady August 1999 N. Schady Economy of Expenditures by the 88247 Peruvian Social Fund (FONCODES), 1991-95 WPS2167 Bonds and Bridges: Social Capital Deepa Narayan August 1999 B. Jones and Poverty 39475 WPS2168 Wage and Productivity Gaps: Dorte Verner August 1999 H. Vargas Evidence from Ghana 37871 WPS2169 Corruption, Public Finances, Simon Johnson August 1999 D. Bouvet and the Unofficial Economy Daniel Kaufman 35818 Pablo Zoido-Lobat6n WPS2170 The Distributional Consequences Ilker Domag August 1999 A. Carcani of Monetary Policy: Evidence 30241 From Malaysia WPS2171 Productivity Growth and Convergence Will Martin August 1999 L. Tabada In Agriculture and Manufacturing Devashish Mitra 36896 WPS2172 The East Asian Crisis: Investigating Warwick McKibbin August 1999 L. Tabada Causes and Policy Responses Will Martin 36896 WPS2173 The Intriguing Relation between Adult Kaushik Basu August 1999 M. Mason Minimum Wage and Child Labor 38811 Policy Research Working Paper Series Contact Title Author Date for paper WPS2174 Prospective Deficits and the Asian Craig Burnside September 1999 E. Khine Currency Crisis Martin Eichenbaum 37471 Sergio Rebelo WPS2175 Sector Growth and the Dual Economy Niels-Hugh Blunch September 1999 H. Vargas Model: Evidence from C6te d'lvoire, Dorte Verner 37871 Ghana, and Zimbabwe WPS2176 Fiscal Risks and the Quality of Hana Polackova Brixi September 1999 A. Panton Fiscal Adjustment in Hungary Anita Papp 85433 Allen Schick WPS2177 Fiscal Adjustment and Contingent Hana Polackova Brixi September 1999 A. Panton Government Liabilities: Case Studies Hafez Ghanem 85433 Of the Czech Republic and Macedonia Roumeen Islam WPS2178 Nonfarm Income, Inequality, and Land Richard H. Adams Jr. September 1999 M. Coleridge-Taylor In Rural Egypt 33704 WPS2179 How Child Labor and Child Schooling Ranjan Ray September 1999 M. Mason Interact with Adult Labor 30809 WPS2180 Regulating Privatized Infrastructures Ofelia Betancor September 1999 G. Chenet-Smith and Airport Services Robert Rendeiro 36370 WPS2181 Privatization and Regulation of the Lourdes Trujillo September 1999 G. Chenet-Smith Seaport Industry Gustavo Nombela 36370 WPS2182 The Integration of Transition Constantine Michalopoulos September 1999 L. Tabada Economies into the World Trading 36896 System WPS2183 Market Discipline and Financial Asli Demirg0-Kunt September 1999 K. Labrie Safety Net Design Harry Huizinga 31001 WPS2184 Financial Services and the World Aaditya Mattoo September 1999 L. Tabada Trade Organization: Liberalization 36896 Commitments of the Developing and Transition Economies WPS2185 Financial Sector Inefficiencies and Pierre-Richard Ag6nor September 1999 T. Shiel Coordination Failures: Implications Joshua Aizenman 36317 for Crisis Management