Monetary policy through the ‘credit-cost channel’: Italy and Germany pre- and post-EMU

Abstract

We present an empirical analysis of the ‘Credit-Cost Channel’ (CCC) of monetary policy transmission. This channel combines bank credit supply and interest rates on loans as a cost to firms. The thrust of the CCC is that it makes both aggregate demand and aggregate supply dependent on monetary policy. As a consequence (1) credit market conditions (e.g. risk spreads) are important sources and indicators of macroeconomic shocks, (2) the real effects of monetary policy are larger and persistent. We have applied the Cointegrated Vector Autoregression (CVAR) econometric methodology to Italy and Germany in the ‘hard’ EMS period and in the European Monetary Union (EMU) period. The short-run and long-run effects of the CCC are detectable for both countries in both periods. Simulation of the estimated model also confirms that inflation-targeting by way of inter-bank rate control stabilizes inflation through structural shifts of the stochastic equilibrium paths of both inflation and the output.

Keywords:

• macroeconomics and monetary economics
• monetary transmission mechanisms
• structural cointegration models
• Italian economy
• German economy

JEL Classification::

• E51
• C32

Acknowledgements

This article is part of a joint research project with the Department of Economics of the University of Oldenburg (Germany) financed by the ‘Vigoni Programme 2007’ of the CRUI-DAAD Italian-German Agreement. We wish to acknowledge valuable comments and suggestions by Katharina Juselius, Tobias Mayer, Ronny Mazzocchi, Emiliano Santoro, Michael Trautwein. We remain fully responsible for this article.

Notes

¹A few examples of various inspirations are Christiano et al. (2007), Leijonhufvud (2007), Cecchetti et al. (2009), Bernanke (2010), Woodford (2010).

² More recently, there has been growing interest in the cost channel of monetary policy per se, both in partial equilibrium (e.g. Barth and Ramey, 2001) and in general equilibrium (e.g. Chowdhury et al., 2006; Ravenna and Walsh, 2006). However, most of these recent models do not treat financial constraints and firms–banks relationships explicitly, and simply plug the policy rate into the production function or the Phillips curve. For the reasons explained below, this is not satisfactory.

³ As regards Italy, see Fiorentini and Tamborini (2002). Gaiotti and Secchi (2004) find evidence of a cost channel of monetary policy at industry-level data. Yet they follow the Barth and Ramey (2001) approach, that is, industry partial equilibrium, with no explicit modelling of the credit market. Moreover, they assume imperfect competition in such a way that the cost channel is identified by a positive pass-through of the interest rate on prices. A similar framework is adopted by Chowdhury et al. (2006) who estimate New-Keynesian Phillips curves augmented with short-term interest rates for a number of industrialized countries. They find evidence of a significant cost channel of interest rates for Italy and France, less so for Germany and Japan.

⁴ Recently, there has been a resurgence of the debate concerning the relative strengths and weaknesses of ‘data-driven’ versus ‘model-driven’ empirical methods (see e.g. Hendry and Mizon, 2000; Hoover et al., 2008; Spanos, 2009). This article is (mildly) inspired by the former methodology; mildly because our econometric model is driven by some theoretical priors about the model's structure.

⁵ A well-known shortcoming of the Old and New-Keynesian sticky price hypothesis is that it implies that the real wages are instead countercyclical after a monetary shock (Christiano et al., 1997).

⁶ A complete analytical treatment can also be found in Tamborini (2009).

⁷ For example, poor sale management technology.

⁸ See Tamborini (2009, sec. 4). The relevant conditions result from five structural parameters, the labour input coefficient of the production function, the real balance effect and the intertemporal substitution effect in the consumption function, the real wage elasticity and intertemporal substitution effect in the labour supply function. The author also shows that different configurations of these parameters may instead generate different macroeconomic outcomes, such as the well-known ‘price puzzle’ (monetary restrictions followed by higher prices) pointed out by Sims (1992).

⁹ The entire empirical analysis was performed using the CATS software, which needs the RATS package to be run. The results are available upon request.

¹⁰ This is common practice since the interbank rates are generally closely driven by the relevant policy rate. It should be noted, however, that in our theoretical model k is regarded as a risk-free rate, whereas the interbank rate may embody some counter-party risk. This component may be neglected under normal operation conditions of the interbank market, when it is unsystematic and relatively small. In our data we do not observe large and systematic deviations of the interbank rates from the policy rates.

¹¹ For technical reasons, in light of other empirical studies (e.g. Chiades and Gambacorta, 2004), we have opted for different reference rates for the two countries, namely, the medium-term yield rate of government bonds (IMF definition) for Italy, and the call money rate (IMF definition for the overnight interbank rate) for Germany. We have also tested the credit-risk premium computed with the same reference rate for the two countries, namely the call money rate. The econometric results are essentially invariant except that k no longer passes the weak-exogeneity test for Italy. This is not a major problem for our purposes; however, we have preferred the former specification of, which allows for more degrees of freedom.

¹² A time lead of 12 months has been chosen empirically by means of sensitivity tests. Recall that the time lead s should capture the theoretical gestation time of output and the related time-horizon for expected inflation. Consequent ly, gestation time has the same empirical effect as the sticky-price hypothesis: that is, the observed changes in output and prices occur with a time lag after the observed change in the policy rate. 12 months is in fact a common time lag associated with the effects of changes in policy rates.

¹³ The literature on monetary policy in the EMS (e.g. De Grauwe, 1992) would predict that the domestic interest rate in a country like Italy could not deviate systematically from uncovered parity with Germany, as implied by

where E _t () is the expected depreciation rate. However, temporary nonzero interest differentials would still be possible as long as the implied expected change in the exchange-rate remained within the band of the parity. On this view, a monetary policy shock can be identified by a deviation from uncovered interest parity, i.e. a nonzero interest-rate differential. Suppose rises in Germany while k_t remains constant in Italy: the interest rate differential in Italy falls. Given the commitment to exchange-rate parity, this is perceived as a positive monetary shock. We consequently introduced the two inter-bank rates as two independent variables with opposite expected sign, and we let the data to explain to what extent they actually exerted independent effects. It is worth noting that the introduction of the German inter-bank rate substantially improved the overall quality of the estimates.

¹⁴ Given linear trends in the data, this choice is generally the best specification with which to begin, unless we have a strong prior hypothesis that the trends cancel in the cointegration relations.

¹⁵ We introduced one transitory intervention dummy to account for the exit of the Italian Lira from the EMS in September 1992. The results of misspecification tests for the unrestricted VAR(2) model with the dummy have been the following: the LM(1) test for first-order autocorrelation and the LM(2) test for second-order autocorrelation, asymptotically distributed as variables, are equal to 36.635 with a p-value of 0.063 and to 36.893 with a p-value of 0.059, respectively.

¹⁶ As shown in Chapter 10 of Juselius (2006), the long-run structure can be identified in the so called reduced form (Equation 1) of the cointegrated VAR model, so that we can test structural hypotheses on the long-run structure without having jointly identifying the short-run structure. Information provided by the theoretical structure is that, of the three endogenous variables, one is forward-looking (inflation), one is contemporaneous (wage) and one has a gestation period (output). Therefore, we have imposed a ‘cascade’ structure such that the first variable affects the others but not the other way round, the second affects the third but not the other way round, and the third is affected by the other two.

¹⁷ The degrees of freedom of the LR test correspond to the weak exogeneity restrictions for the variable k_t .

¹⁸ Allowing for a certain number of significant lags in the autoregressive polynomial, the series can be considered to be near I(2) when the modulus of the largest root, ρ₁, is significantly equal to 1.0 and the next root, ρ₂, is significantly less than 1.0, but not too far from it. In our dataset, the largest roots of Lib are 0.98 and 0.91, the largest roots of k are 1.00 and 0.88 and the largest roots of are 0.99 and 0.83. Therefore, our univariate series can be considered as generated either by an I(1) process, or by a near I(2) process (see Juselius, 2010, p. 9).

¹⁹ The statistical significance of the breaks has resulted from the variable exclusion testing provided as an automated test procedure in RATS.

²⁰ The misspecification tests for the unrestricted VAR(1) model have taken the following values: the LM(1) test for first-order autocorrelation and the LM(2) test for second-order autocorrelation, asymptotically distributed as variables, are equal to 32.513 with a p-value of 0.144 and to 34.900 with a p-value of 0.090, respectively.

²¹ The values of the misspecification tests for the unrestricted VAR(2) model are the following: the LM(1) test for first-order autocorrelation and the LM(2) test for second-order autocorrelation, asymptotically distributed as variables, are equal to 32.257 with a p-value of 0.151 and to 24.622 with a p-value of 0.484, respectively.

²² The degrees of freedom of the LR test correspond to the weak exogeneity restrictions for the variable k_t .

²³ Given that Germany had no explicit non-EMS exchange-rate target, Lib_t , unlike k_t ^GER for Italy, should take the same sign as the domestic rate.

²⁴ For Germany, we have that the largest roots of k are 1.00 and 0.90, while the largest roots of Lib, being the series the same as for Italy, are 0.98 and 0.91.

²⁵ In January 2003, the BB's earlier survey of lending and deposit rates was discontinued and replaced with the new harmonized MFI interest rate statistics. Since the two sets of statistics differ in their methodology, we have introduced a dummy for the change from the lending rate series SU0004 to lending rate series SUD123.

²⁶ The values of the misspecification tests for the unrestricted VAR(1) model are following: the LM(1) test for first-order autocorrelation and the LM(2) test for second-order autocorrelation, asymptotically distributed as variables, are equal to 26.320 with a p-value of 0.391 and to 20.095 with a p-value of 0.742, respectively.

²⁷ In this decomposition, , where and are (p × (p-r)) matrices orthogonal to and , respectively, and .

²⁸ Hence, we have simulated a pure inflation targeting regime, rather than a common Taylor rule where output is also a target.

²⁹ Note that the 2006–2007 inflationary phase took off much earlier in Germany than in Italy both in the real data and in the simulated ones. The timing of the rise in the observed inter-bank rates follows that of Germany's inflation. Accordingly, the simulated control rule yields no interventions on the inter-bank rate for Germany but downward impulses for Italy.

³⁰ This is derived from the model by Tamborini (2009).

³¹ If this value were positive, the credit-risk premium would be lower. This, however, has no significant impact on the core results (qualitative and quantitative) of the model.

³² Claims on D_bt could be randomized with no alteration of the key results.