Which Dollar Debt Trigger the Balance Sheet Effect? Evidence for Peruvian Firms

Transcription

1 Which Dollar Debt Trigger the Balance Sheet Effect? Evidence for Peruvian Firms Nelson Ramírez-Rondán Central Bank of Peru July 2015 Abstract This paper analyzes the impact of the exchange rate depreciation on firm investment using financial information from 69 non-financial Peruvian firms from 2003 to 2013 We find evidence that, for firms holding dollar-denominated debt below 16 percent of total debt, competitiveness effect dominates the balance sheet effect, this is, investment decisions are positively affected by real exchange rate depreciation For firms holding dollar debt between 16 percent and 32 percent, competitiveness effect offsets the balance sheet effects, which means that investment decisions are not affected by real exchange rate depreciation Finally, firms holding dollar debt above 32 percent, balance sheet effect dominates the competitiveness effect, which implies that investment decisions are negatively affected by real exchange rate depreciation JEL Classification: C33, E22, F31, F34 Keywords: Balance Sheet, Dollar Debt, Peru 1 Introduction When firms make investment decisions also have to decide how to finance this investment, in emerging economies firms have to decide their debt structure between domestic and foreign currencies When firms liabilities are denominated in foreign currency and if as a I am grateful to Bryce Cruz and Yamily León for providing an excellent research assistance All remaining errors are mine Nelson Ramírez-Rondán (nelsonramirez@bcrpgobpe) is a researcher in the Research Division at the Central Bank of Peru, Jr Miró Quesada 441, Lima, Peru 1

2 consequence it creates currency mismatches between revenues and expenditures of firms, then the firm balance sheet is affected by variations in the exchange rate 1 Exchange rate depreciation increases the debt-asset ratio, makes more difficult to access to alternative sources of financing Thus, for firms in the private sector, these balance sheet negatively affect their plans on investment and production, which can lead a contractive effect an aggregate level by dominating the competitiveness effect Theoretically, a large body of literature is being developed upon the work of Bernanke y Gertler (1995), who include imperfection in the domestic financial market within in an open economy model In those models, if there exists a significant currency mismatch in the economy, a large devaluation will deteriorate the firm s net worth As the firm s risk increases, credit becomes more expensive and more restricted, which finally affects investment and therefore, aggregate demand Using this balance-sheet channel, Krugman (1999) and Aghion et al (2001) present models with multiple equilibrium Further literature on liability dollarization and currency mismatch has suggested that a balance-sheet effect induced by exchange rate depreciations could be an explanation for this negative impact (see Cespedes et al, 2004; Choi and Cook, 2004; Magud, 2010; Ize and Levy-Yeyati, 2005; Batini et al, 2007; Bleakley y Cowan, 2008; Carranza et al, 2009) Empirical analyses, however, have found only weak evidence for this effect (see Luengnaruemitchai, 2003, for a review), and usually only in the context of quite large depreciations (see, among others, Burstein et al, 2005; Galindo et al, 2003; Leiderman et al, 2006) These empirical findings suggest that the aggregate investment function may present a nonlinearity in its dependence on the (real) exchange rate Carranza et al (2011) show that the negative balance-sheet effect of an exchange rate depreciation may be observable only if the magnitude of the depreciation is large enough Azabache (2010) shows the effects depend on the firm s leverage level Thus, given the nonlinear effects that suggest the theory (multiple equilibrium) and empirical evidence, we argue that the balance sheet effect depend on the dollar debt as a percentage of total debt For this purpose, we estimate a dynamic panel threshold model proposed by Ramírez-Rondán (2015), where we estimate two threshold levels of dollar debt, which define three regimes In the first one (high dollar debt), the balance sheet effect dominate the competitiveness effect; in the second one (low dollar debt), one effect offsets the other; and in the third regime (low dollar debt), the competitiveness effect 1 Financial risk associated to debt in foreign currencies has played a central role in the Asian crisis at the end of 90 2

3 dominates the balance sheet effects The remainder of this paper is organized as follows Section 2 discusses the empirical methodology such as specification, data and estimation method Section 3 presents the empirical results: estimation of the threshold models Finally, section 4 concludes 2 Empirical Methodology 21 Specification In order to consider the threshold level of dollar debt that triggers balance sheet effect, we estimate the following variation of an investment model with a threshold variable, I it = α i + βi it 1 + θ 1 q t 1(D it γ) + θ 2 q t 1(D it > γ) + π z it + u it, (1) where I is the investment; 2 q is the bilateral (Pen/US) real exchange rate depreciation; D is the dollar denominated debt expressed in local currency as a fraction of total debt; α i is an unobserved country specific-effect; γ is a threshold level of dollar debt; 1() is an indicator variable; z is a set of other determinants of investment; i refers to non-financial firm; and t refers to time period (year) 22 Other Determinants of Firm Investment We consider two groups of explanatory variables of firm s investment: the firs group is related with firm s specific variables such as cash flow, sales, working capital, size firm and leverage The second group is related with macroeconomic conditions such as real exchange rate (PEN/US$), dollarization ratio, terms of trade, fiscal impulse and monetary conditions 23 Data The period of study spans from 2003 to 2011 for a sample of 69 countries Data are constructed manually from the firm financial information available from the Superintendencia de Mercado de Valores Table 1 shows constructions and definitions of the variables used in the estimation analysis 2 Investment is the expenditure in machinery and equipment net of fixed asset sales 3

4 Variable Investment Debt in US$ Leverage Sales Working capital Size firm Table 1: Definitions Definition Investment flow in the period over total asset at the end of the previous period Investment is the expenditure in machinery and equipment net of fixed asset sales Total liabilities in foreign currency expressed in terms of domestic currency as a percentage of total liabilities Total liabilities divided by equity Real sales growth divided by total assets Difference between current assets minus current liabilities divided by total assets Total sales in logarithm 24 Estimation Method Threshold regression models have had a great development in time series analysis, Hansen (1999) extends such models in a panel data context; and Ramírez-Rondán (2015) extends this work to allow dynamics in a panel data threshold model In equation (1), the observations are divided into two regimes depending on whether the threshold variable q it is smaller or larger than the threshold parameter γ The regimes are distinguished by differing regression slopes, θ 1 and θ 2 For the identification of these coefficients, it is required that the elements of q it are not time invariant The error u it is assumed to be normally independent and identically distributed (iid) with mean zero and finite variance σ 2 The regressors y it 1 and z it are assumed to be exogenous and the threshold variable x it is assumed to be strict exogenous and it must have a continuous distribution Taking the first difference transformation in order to eliminate the individual specific effect, we have y it = β y it 1 (γ) + θ 1 x it(γ) + θ 2 x + it (γ) + π z it + u it, (2) where, y it (y it y it 1 ) (y it 1 y it 2 ), y it 1 (γ) y it 1 1(q it γ) y it 2 1(q it 1 γ), x it(γ) x it 1(q it γ) x it 1 1(q it 1 γ), y + it 1 (γ) y it 11(q it > γ) y it 2 1(q it 1 > γ), x + it (γ) x it1(q it > γ) x it 1 1(q it 1 > γ), z it z it z it 1 and u it u it u it 1 For t = 2, 3,, T, (2) is well defined But not for y i1 because y i0 (γ) is missing, that is y i, 1 is not available Since the initial conditions are critical in the estimation procedure proposed by Ramírez- 4

5 Rondán (2015), we assume that the process has started from a finite period in the past, namely for given values of y i, 1 the marginal distribution of y i1 conditional on the observables can be written as y i1 = δ 1 1(q i1 γ) + δ 2 1(q i1 > γ) + δ 3 x i (γ) + δ 4 x + i (γ) + δ 5 z i + υ i1, (3) where δ 1, δ 2, δ 3, δ 4 and δ 5 are external parameters, x i (γ) = ( x i1(γ),, x it (γ)), x + i (γ) = ( x+ i1 (γ),, x+ it (γ)) and z i = ( z i1,, z it ) Notice that we take the conditional expectation also on all the observables since x it and z it are assumed to be exogenous for all t Maximum Likelihood Estimation Under the strict exogeneity of x it, z it and q it and by construction E(υ i1 x i, z i, q i ) = 0, Eυi1 2 = συ, 2 where x i = (x i1,, x it ), z i = (z i1,, z it ) and q i = (q i1,, q it ) We assume that Cov(υ i1, u i2 ) = σu 2 and Cov(υ i1, u it ) = 0 for t = 3,, T, i = 1,, n Let y i = ( y i1, y i2,, y it ) and u i = (υ i1, u i2,, u it ) The Jacobian of the transformation from u i to y i is unity and the joint probability distribution function of y i and u i are therefore the same The covariance matrix of u i has the form Ω = σu 2 where ω = συ/σ 2 u 2 ω = σ uω 2, (4) Under the assumption that u it is independent normal, the joint probability distribution function of y i conditional on x i, z i and q i is given by ln L(φ, γ, σu, 2 ω) = nt 2 ln(2π) nt 2 ln(σ2 u) n ln[1 + T (ω 1)] 2 1 n ( y i y i, 1 (γ)φ) Ω 1 ( y i y i, 1 (γ)φ), (5) 2 where φ = (δ 1, δ 2, δ 3, δ 4, δ 5, β, θ 1, θ 2, π ) and the matrix y i, 1 (γ) is defined as follows 5

6 y i, 1 (γ) = 1(q i1 γ) 1(q i1 > γ) x i (γ) x + i (γ) z i T 0 1 T 0 1 T T 0 1 T 0 1 T ; T 0 1 T 0 1 T k y i1 (γ) x i2(γ) x + i2 (γ) z i2 y i2 (γ) x i3(γ) x + i3 (γ) z i3 y it 1 (γ) x it (γ) x+ it (γ) z it, where k is the number of variables in z it The likelihood function (5) is well defined, depends on a fixed number of parameters The MLE ( φ, γ, σ u, 2 ω) are the values which maximize ln L(φ, γ, σu, 2 ω) The algorithm has the following procedure: first, form a grid on the empirical support, [γ; γ], of the threshold variable q it Second, for each value of γ on this grid, calculate the MLE φ( γ), σ u(γ) 2 and ω(γ) by maximizing the criterion (5), for that we can use either an iterative technique such as the Newton-Raphson procedure by using initial consistent estimates or a grid search procedure on ω(γ) at a given γ, and then choosing that value of ω(γ) which globally maximizes the function (5) Third, with the ML estimators computed in the second part for each γ, find the MLE γ as the value of γ on the grid on the empirical support [γ; γ] which yields the highest value of the following concentrated criterion (6) ln L(γ) = nt 2 ln(2π) nt 2 ln( σ2 u(γ)) n ln[1 + T ( ω(γ) 1)] 2 1 n ( y i y i, 1 (γ) φ(γ)) Ω(γ) 1 ( y i y i, 1 (γ) φ(γ)) (6) 2 Finally, set φ = φ( γ), σ 2 u = σ 2 u( γ), ω = ω( γ) and û i = û i ( γ) Threshold estimate In the context of the Conditional Least Square Estimation (CLSE) of a threshold autoregression, Chan (1993) develops the strong consistency of the threshold parameter while Hansen (2000) shows its consistency by using the small threshold effect assumption 6

7 In all these models the errors are independent of the threshold variable as well as in our model They are also independent (mean-independent) of the regressors, even though in our model we have the same environment in the structural model In the model in first differences we introduce a correlation between the first difference of the errors and the first difference of the lagged variable due to the dynamic nature of the model Then using a different technique Ramírez-Rondán (2015) proves the consistency of the threshold parameter In the context of threshold autoregression estimation, Chan (1993) establishes the limiting distribution of the threshold parameter estimator converges to a functional of a compound Poisson process at a rate n The distribution is too complicated to be used in practice due to the dependence on the nuisance parameters (including the marginal distribution of the threshold variable and all the regression coefficients) Hansen (2000) developed an asymptotic distribution for the threshold parameter estimate based on the small threshold effect assumption, in which the threshold model becomes the linear model asymptotically The limiting distribution converges to a functional of a two-sided Brownian motion process at a rate n 1 2α, where 0 < α < 1/2 The distribution does not depend on the nuisance parameters; thus, the distribution can be available in a simple closed form Ramírez-Rondán (2015) adopts Hansen s approach in his setting He established under which conditions the asymptotic distribution of the threshold parameter converges to a functional of a two sided Brownian motion distribution This asymptotic distribution yields a computationally attractive method for constructing confidence intervals, and is described in detail in Hansen (1999) and Ramírez-Rondán (2015) in the context of the non-dynamic and dynamic panel threshold models, respectively Basically, Hansen (2000) argues the best ways to form confidence intervals for the threshold is to form the no-rejection region using the likelihood ratio statistic for testing on γ To test hypothesis H 0 : γ = γ 0, the likelihood ratio test is to reject large values of LR(γ 0 ) where LR(γ) = nt Sn(γ) Sn( γ), (7) Sn( γ) where S n (γ) = n û i(γ) Ω 1 û i (γ) is the minimum distance estimator Hansen (1996) shows the LR(γ) converges in distribution to ξ as n, where ξ is a random variable with distribution function P (ξ z) = (1 exp( z/2)) 2 Then, the asymptotic distribution of the likelihood ratio statistic is non-standard, yet free of nuisance parameters Since the asymptotic distribution is pivotal, it may be used to form valid asymptotic 7

8 confidence intervals Furthermore, the distribution function ξ has the inverse c(a) = 2 ln ( 1 1 a ), (8) where a is the significance level To form an asymptotic confidence interval for γ, the norejection region of confidence level 1 a is the set of values of γ, such that LR(γ) c(a), where LR(γ) is defined in (7) and c(a) is defined in (8) This is easiest to find by plotting LR(γ) against γ and drawing a flat line at c(a) Slope parameters estimates The likelihood function (5) is well defined; it depends on a fixed number of parameters, and satisfies the usual regularity conditions conditional on γ Therefore, the MLE of (5) is consistent and asymptotically normally distributed, when n tends to infinity when T is fixed The estimated covariance matrix for the ML slope estimators φ is Cov( φ) = ( n y i, 1 ( γ) Ω 1 y i, 1 ( γ) ) 1 ( n 1 = σu 2 y i, 1 ( γ) Ω 1 y i, 1 ( γ)) Or, under a suitable partition of the matrix y i, 1 ( γ), the estimated covariance matrix for the ML slope estimators β 1 and β 2 is Cov β θ 2 θ 2 π ( n 1 = σ2 u yi, 1( γ) Ω 1 yi, 1( γ)), where, yi, 1(γ) = y i1 (γ) x i2(γ) x + i2 (γ) z i2 y i2 (γ) x i3(γ) x + i3 (γ) z i3 y it 1 (γ) x it (γ) x+ it (γ) z it 8

9 While we need the assumption that the errors are iid, this assumption can be relaxed when constructing confidence intervals for the slope coefficients If the errors are allowed to be conditionally heteroskedastic, the covariance matrix estimator for the slope parameters is Cov β θ 2 θ 2 π = ( n ) 1 ( n ) yi, 1( γ) Ω 1 yi, 1( γ) yi, 1( γ) ũ i ũ i yi, 1( γ) ( n y i, 1( γ) Ω 1 y i, 1( γ)) 1, where ũ i = Λ 1/2 H(υ i1, u i2,, u it ), and since Ω > 0 then Ω 1 = σ 2 uω 1 = H Λ 1 H and then there is a unique root Λ 1/2 H which is also positive definite Λ 1/2 H > 0 25 A Two Threshold Dynamic Model Model (1) has a single threshold In some applications there may be multiple thresholds For example, in our case, the double threshold model can take the form I it = α i +βi it 1 +θ 1 q t 1(D it γ 1 )+θ 2 q t 1(γ 1 < D it γ 2 )+θ 3 q t 1(D it > γ 2 )+π z it +u it, (9) where the thresholds are ordered so that γ 1 < γ 2 Thus, we can identify three regimes: high dollar debt (the balance sheet effect should dominate the competitiveness effect); low dollar debt (one effect should offset the other); and low dollar debt (the competitiveness effect should dominate the balance sheet effects) For given (γ 1, γ 2 ), (6) is linear in the slopes, then the ML estimation is appropriate Thus for given (γ 1, γ 2 ) the concentrated log-likelihood function ln L(γ 1, γ 2 ) is straightforward to calculate (as in the single threshold model) The joint maximum likelihood estimates of (γ 1, γ 2 ) are by definition the values which jointly maximize ln L(γ 1, γ 2 ) While these estimates might seem desirable, Hansen (1999) argues that they may be quite cumbersome to implement in practice Ramírez-Rondán (2015) suggest to estimate the model (9) sequentially 9

10 3 Estimation and Inference Results First, we estimate the model in equation (1), where we have one threshold that determines two regimes The dollar debt (as a percent of total debt) threshold estimate is 32 percent; thus, the two classes of regimes indicated by the point estimate are those with low dollar debt for dollar debt-total debt ratios lower than 32 percent, and high dollar debt for dollar debt-total debt ratios higher than 32 percent Table 2: Estimation results Dependent variable: Investment Explanatory variables: Estimates ML SE White SE Real exchange rate depreciation (Dollar debt 32%) Real exchange rate depreciation (Dollar debt > 32%) Investment in the previous period Cash flow Sales growth over total assets Working capital Size firm Leverage over equity Dollarization Number of firms (n) 69 Number of periods, years (T) 11 Observations used in the estimation (n*(t-1)) 690 Negative log-likelihood Table 2 shows the estimation results of equation (1) The coefficients of primary interest are those on the real exchange rate depreciation The point estimates suggest that nonfinancial firms under the low dollar debt, real exchange rate depreciation has positive and significant effects (the parameter estimate is different from zero at the 5 percent significance level) on firm s investment; while, firms under the high dollar debt, real exchange rate depreciation has negative, but not significant effects on firm s investment The last result suggest that for some firms both effects offset one to each other; thus, it s 10

11 more appropriate to consider a model with two threshold parameters which produce three regimes Table 3: Estimation results Dependent variable: Investment Explanatory variables: Estimates ML SE White SE Real exchange rate depreciation (Dollar debt 17%) Real exchange rate depreciation (17% < Dollar debt 32%) Real exchange rate depreciation (Dollar debt > 32%) Investment in the previous period Cash flow Sales growth over total assets Working capital Size firm Leverage over equity Dollarization Number of firms (n) 69 Number of periods, years (T) 11 Observations used in the estimation (n*(t-1)) 690 Negative log-likelihood Table 3 shows the estimation results of equation (9), where we have two thresholds that determines three regimes The dollar debt (as a percent of total debt) thresholds estimate are 17 percent and 32 percent; thus, the three classes of regimes indicated by the points estimates are those with low dollar debt for dollar debt-total debt ratios lower than 17 percent, moderate dollar debt for dollar debt-total debt ratios between 17 and 32 percent, and high dollar debt for dollar debt-total debt ratios higher than 32 percent The point estimates suggest that non-financial firms under the low dollar debt, real exchange rate depreciation has positive and significant effects (the parameter estimate is different from zero at the 5 percent significance level) on firm s investment; while, firms 11

12 under the moderate dollar debt, real exchange rate depreciation has no effects on firm s investment, finally, under the high dollar debt regime, real exchange rate depreciation has a negative and significant effect Regarding the other firm s determinants of investment, working capital has a negative effect, size firm has a positive effect on firm s investment, and the other determinants have no effects It is important to mention to include macroeconomic determinants of investment such as terms of trade, fiscal impulse, monetary conditions, etc 4 Conclusion In this paper, we investigate the exchange rate depreciation effects on firm s investment for a sample of 69 non-financial firms during the period We basically estimate a set of dynamic panel threshold models, where we find that such effect depend on the specific regime We find that for firms holding dollar-denominated debt below 16 percent of total debt, competitiveness effect dominates the balance sheet effect, this is, investment decisions are positively affected by real exchange rate depreciation For firms holding dollar debt between 16 percent and 32 percent of total debt, competitiveness effect offsets the balance sheet effects, which means that investment decisions are not affected by real exchange rate depreciation Finally, firms holding dollar debt above 32 percent of total debt, balance sheet effect dominates the competitiveness effect, which implies that investment decisions are negatively affected by real exchange rate depreciation This work is still in progress, we are working in consider the leverage (total liabilities divided by equity) and the ratio liabilities-assent in US$ dollars as a threshold variable instead of dollar debt (percent of total debt) Also, we are working in specifications that include other macroeconomic determinants of firm s investment such as terms of trade, fiscal impulse, monetary conditions, etc 12

13 References Aghion, P, P Bacchetta and A Banerjee (2001) Currency Crises and Monetary Policy in an Economy with Credit Constraints European Economic Review 45(7), Azabache, P (2010) Decisiones de Inversión en Empresas con Dolarización Financiera: Un modelo Umbral del Efecto Hoja de Balance Consorcio de Investigación Económica y Social Batini, N, P Levine and J Pearlman (2007) Optimal Exchange Rate Stabilization in a Dollarized Economy with Inflation Targets Mimeo, London Metropolitan University Bernanke, B and M Gertler (1995) Inside the Black Box: The Credit Channel of Monetary PolicyTransmission Journal of Economics Perspectives 9(4), Bleakley, H and K Cowan (2008) Corporate Dollar Debt and Depreciations: Much Ado about Nothing? Review of Economics and Statistics 90(4), Bleakley, H and K Cowan (2010) Maturity Mismatch and Financial Crises: Evidence from Emerging Market Corporations Journal of Development Economics 93(2), Burstein, A, M Eichenbaum y S Rebelo (2005) Large Devaluations and the Real Exchange Rate Journal of Political Economy 113(4), Carranza, L, JE Galdon-Sanchez and J Gomez-Biscarri (2009) Exchange Rate and Inflation Dynamics in Dollarized Economies Journal of Development Economics 89(1), Carranza, L, JE Galdon-Sanchez y J Gomez-Biscarri (2011) The Relationship between Investment and Large Exchange Rate Depreciations in Dollarized Economies Journal of International Money and Finance 30(7), Chan, KS (1993): Consistency and Limiting Distribution of the Least Squares Estimator of a Threshold Autoregressive Model Annals of Statistics 21(1), Cespedes, L, R Chang y A Velasco (2004) Balance-sheet and exchange rate Policy American Economic Review 94(4), Choi, WG and Cook, D (2004) Liability Dollarization and the Bank Balance Sheet Channel Journal of International Economics 64(2),

14 Galindo, A, U Panizza F Schiantarelli (2003) Debt Composition and Balance Sheet Effects of Currency Depreciation: A Summary of the Micro Evidence Emerging Markets Review 4, Hansen, BE (1999): Threshold Effects in Non-Dynamic Panels: Estimation, Testing and Inference Journal of Econometrics 93(2), Hansen, BE (1996): Inference When a Nuisance Parameter is not Identified Under the Null Hypothesis Econometrica 64(2), Hansen, BE (2000): Sample Splitting and Threshold Estimation Econometrica 68(3), Hsiao, C, M Pesaran and K Tahmiscioglu (2002): Maximum Likelihood Estimation of Fixed Effects Dynamic Panel Data Models Covering Short Time Periods Journal of Econometrics 109(1), Ize, A and E Levy-Yeyati (2005) Financial De-Dollarization: Is It for Real? IMF Working Paper 05/187 Krugman, P (1999) Balance Sheet, the Transfer Problem, and Financial Crises International Tax and Public Finance November 6(4), Leiderman, L, R Maino and E Parrado (2006) Inflation Targeting in Dollarized Economies IMF Working Paper 06/157 Luengnaruemitchai, P (2003) The Asian Crisis and the Mystery of the Missing Balance Sheet Effect UC Berkeley working paper Magud, N (2010) Currency Mismatch, Openness and Exchange Rate Regime Choice Journal of Macroeconomics 32(1), Ramírez-Rondán, N (2015): Maximum Likelihood Estimation of Dynamic Panel Threshold Models Working Paper 32, Peruvian Economic Association 14