The effects of monetary policy in the Czech Republic: an empirical study

Abstract

In this paper, we examine the effects of Czech monetary policy on the economy within the vector auto regression (VAR), structural VAR, Bayesian VAR with sign restrictions, and factor-augmented VAR, frameworks. We document a well-functioning transmission mechanism similar to the euro area countries, especially in terms of persistence of monetary policy shocks. Subject to various sensitivity tests, we find that a contractionary monetary policy shock has a negative effect on the degree of economic activity and the price level, both with a peak response after one year or so. Regarding prices at the sectoral level, tradables adjust faster than non-tradables, which is in line with microeconomic evidence on price stickiness. There is no price puzzle, as our data come from a single monetary policy regime. There is a rationale in using the real-time output gap instead of current GDP growth, as using the former results in much more precise estimates. The results indicate a rather persistent appreciation of the domestic currency after a monetary tightening, with a gradual depreciation afterwards.

Appendices

Appendix 1

See Fig. 5.

Fig. 5

Time series

Appendix 2—additional results

2.1 Impulse responses to a monetary shock

See Figs. 6, 7, 8, 9, 10, 11.

Fig. 6

GDP instead of output gap, VAR

Fig. 7

GDP instead of output gap, sectoral prices, VAR

Fig. 8

No exogenous variables, VAR

Fig. 9

No exogenous variables, sectoral prices, VAR

Fig. 10

SVAR

Fig. 11

SVAR, sectoral prices

Appendix 3—factor-augmented VAR

In this Appendix, we briefly document our attempt to study monetary policy effects within FAVAR. However, as documented below, we find that the results based on FAVAR are very sensitive and the confidence intervals for the impulse responses are rather large.

We follow an approach developed by Bernanke et al. (2005).¹⁸ FAVAR can be represented in the following form:

$$ \left[ {\begin{array}{*{20}c} {F_{t} } \\ {Y_{t} } \\ \end{array} } \right] = \Upphi (L)\left[ {\begin{array}{*{20}c} {F_{t - p} } \\ {Y_{t - p} } \\ \end{array} } \right] + v_{t} $$

(3)

Vector Y _t contains observable economic variables, whereas F _t represents unobserved factors which provide additional economic information not fully captured by the Y _t . We estimate these unobservable factors using a principal components approach, which exploits the assumption that information about the unobservable economic factors can be inferred from a large number of economic time series X _t . Specifically, we can think of the unobservable factors in terms of concepts such as “economic activity” or “investment climate.” They can be represented not by a single economic variable, but rather by several time series of economic indicators.

The FAVAR methodology allows us not only to use richer information set in the model specification, but also to analyze the effects of a monetary policy shock on a greater number of economic variables. There are two main approaches to estimating FAVAR: a two-step principal components approach and a one-step approach that estimates (3) and a dynamic factor model jointly. As Bernanke et al. (2005) do not find any particular differences between these two estimators in terms of inference, we opt for the computationally simpler two-step approach.¹⁹

In our FAVAR specification, X _t consists of a balanced panel of 40 series that have been transformed in order to ensure their stationarity. The description of these series and their transformations is included in Table 3. The data is at monthly frequency and spans the period from February 1998 to May 2006. Following Bernanke et al. (2005), we assume that the monetary policy instrument (the 3-month interest rate) is the only observable factor, hence the only variable included in Y _t . For identification purposes the monetary policy instrument is ordered last, which implies that latent factors do not respond contemporaneously (within a month) to innovations in monetary policy. As in Bernanke et al. (2005), we distinguish between “slow-moving” and “fast-moving” variables. A “slow-moving” variable is assumed not to react contemporaneously to shocks, while the “fast-moving” variables react instantaneously to changes in monetary policy or economic conditions. The classification of the variables into these two categories is included in Table 2.

Table 3 Data description

In the first step of the two-step estimation, we can distinguish three stages. First, we use principal component analysis to estimate the common factors C _t from all the variables in X _t . Second, after dividing the series in X _t into slow- and fast-moving ones, we estimate the “slow-moving” factors $\overset{\lower0.5em\hbox{$\smash{\scriptscriptstyle\frown}$}}{F}_{t}^{s}$ as the principal components of the “slow-moving” variables. Finally, we estimate the following regression:

$$ \overset{\lower0.5em\hbox{$\smash{\scriptscriptstyle\frown}$}}{C}_{t} = b_{Fs} \overset{\lower0.5em\hbox{$\smash{\scriptscriptstyle\frown}$}}{F}_{t}^{s} + b_{Y} Y_{{t + e_{t} }} $$

Based on these estimates, $\overset{\lower0.5em\hbox{$\smash{\scriptscriptstyle\frown}$}}{F}_{t}$is constructed as $\overset{\lower0.5em\hbox{$\smash{\scriptscriptstyle\frown}$}}{C}_{t} - \overset{\lower0.5em\hbox{$\smash{\scriptscriptstyle\frown}$}}{b}_{Y} Y_{t}$. In the second step, we estimate the VAR in $\overset{\lower0.5em\hbox{$\smash{\scriptscriptstyle\frown}$}}{F}_{t}$ and Y _t , using a recursive assumption.

One caveat in our analysis is the fact that we have only 40 series available for principal component analysis, as compared to the 120 used in Bernanke et al. (2005). While this may be viewed as a weakness at first sight, it has been argued by Boivin and Ng (2006) that more series do not necessarily ensure better data quality, due to cross-correlation of idiosyncratic errors. In one of their tests, they show that the factors extracted from 40 pre-screened series may in some cases yield better results as compared to using 147 series. Therefore, for the sake of size, 40 series, at least in general, should not pose a problem.

Our main results are given in Fig. 9. Each panel shows the impulse responses of selected macroeconomic variables to a monetary policy shock with 90% confidence intervals. The FAVAR model in Fig. 9 includes three principal factors, but the results were no different when the number of factors was changed. In the benchmark specification we use one lag. The results are highly sensitive to the numbers of lags used, with more lags resulting in highly improbable results.

As a result, the FAVAR model does not appear to properly capture the developments in the Czech economy. Most importantly, the confidence intervals are too large to infer the direction of the impact of the monetary policy change on the macroeconomic variables. The exception is actual GDP growth, lgdp _t , which declines after monetary policy shock, as predicted by the theory.

There may be several reasons for the lack of significant results in the FAVAR estimation for the Czech economy. One reason is likely to do with the relatively short span of the available data; another may be data quality, as discussed by Boivin and Ng (2006). As it is at monthly frequency, our dataset lacks variables related to consumption, housing starts and sales as well as real inventories and therefore some important economic information may be missing (Fig. 12).

Fig. 12

FAVAR results. Note: Impulse responses with 90% confidence intervals are presented