«p> EUROPEAN CENTRAL BANK W O R K I N G PA P E R S E R I E S P K E E C B T K E B Z E B C E FO BA RC TH T KG EH R PO EC E E OU WORKING PAPER NO. 254 LI ...»
EUROPEAN CENTRAL BANK
W O R K I N G PA P E R S E R I E S
TH T KG
PO EC E E OU
WORKING PAPER NO. 254 LI B’ VA ND
CY S LU
ST O AT TU
R NE IO DY
FORECASTING REAL GDP: WHAT AT T NEG AR O
ROLE FOR NARROW MONEY? YYFBY C. BRAND, H.-E. REIMERS AND F. SEITZ September 2003
EUROPEAN CENTRAL BANK
W O R K I N G PA P E R S E R I E SFO BA RC TH T KG EH R
PO EC E E OULI B’ VA ND WORKING PAPER NO. 254 CY S LU S M
ST O AT TUN IO D R
FORECASTING REAL GDP: WHAT AT ET N YEG AR O
ROLE FOR NARROW MONEY? 1YYF BY C. BRAND2, H.-E. REIMERS3 AND F. SEITZ4 September 2003 1 We thank seminar participants at the ECB, at the Center for European Economic Research and at the Univ
All rights reserved.
Reproduction for educational and non-commercial purposes is permitted provided that the source is acknowledged.
The views expressed in this paper do not necessarily reflect those of the European Central Bank.
ISSN 1561-0810 (print) ISSN 1725-2806 (online) Contents
ECB • Working Paper No 254 • September 2003
This paper analyses the information content of M1 for euro area real GDP since the beginning of the 1980s and reviews theoretical arguments on why real narrow money should help predict real GDP. We find that, unlike in the U.S., in the euro area, M1 has better and more robust forecasting properties for real GDP than yield spreads. This property persists when one controls for a number of other influences. We also evaluate the out-of-sample forecasting performance of different classes of VAR models comprising real M1, GDP and other indicators, using as benchmark a simple univariate model. As a result, only VARs in first differences are able to outperform the benchmark.
JEL: E41, E52, E58 Key words: Money; business cycle; forecast comparison; VAR models.
4 ECB • Working Paper No 254 • September 2003 Non-technical summary This study reviews the role of narrow money - predominantly real M1 - in terms of leading cyclical conditions in the euro area over the sample period 1981Q1 - 2001Q4. It also looks at how important real narrow money growth has been in relation to information already captured by short-term interest rates and a number of alternative leading indicator variables, such as various yield-spread measures, oil prices, stock price indices and exchange rates. Evidence is obtained on the basis of single-equation test regressions, as well as a more rigorous comparison of the out-of-sample forecast performance of different VAR-specifications (VARs in levels, first-difference VARs, VECMs and BVARs) with a univariate benchmark model (where GDP is forecast only on the basis of past information in GDP itself). These VARs comprise real M1 growth, GDP growth and the other potential leading indicator variables for real GDP.
As a result, the information content of different interest rate measures (long-term/short-term, real and nominal) for future changes in GDP seems rather mixed, while the predictive content of M1 remains important and tends to be largely unaltered in terms of coefficient size and forecasting horizon.
Furthermore, in contrast to results obtained from studies for the U.S., in the euro area, M1 has better and more robust forecasting properties than the term spread (especially, when measured on the basis of historical euro-area interest rates). These properties of M1 are also maintained when comparing with a broader set of variables, like real and nominal effective exchange rates and oil prices. In all cases, narrow money seems crucial for cyclical developments. VARs in differences always outperform the benchmark irrespective of the forecast horizon considered, whereas the other classes of VAR models are not able to do so. Therefore, once the information content of money is taken into account, the choice of the general model seems to be more important than the selection of additional variables or the concrete forecast horizon. Furthermore, estimates of forecast models based on other variables, ignoring the information content of money, may be subject to a so-called "omitted-variable bias".
The paper also presents a number of theoretical arguments why money may affect cyclical conditions beyond the information already captured by interest rates. In this context it reconsiders conditions for the existence of a real balance effect. Furthermore it reviews the arguments recently re-stated by Meltzer (2001) and Nelson (2002a, 2002b) who focus on how money serves as a summary statistic reflecting a whole range of relative asset prices which are crucial for the monetary transmission process.
ECB • Working Paper No 254 • September 20031. Introduction
Is money a good indicator of future real economic developments? And if yes, what are the theoretical justifications and what is the concrete mechanism behind the connection? These questions are among the most hotly debated in monetary economics. The present paper tries to shed some light on these issues for the euro area. In particular, it concentrates on the forecast performance of M1 for real GDP.
There are plenty of empirical papers dealing with this question for the US (see e.g. Hamilton and Kim (2002), Amato and Swanson (2001), Vilasuso (2000), Swanson (1998), Estrella and Mishkin (1997), Feldstein and Stock (1997), Friedman and Kuttner (1992)). The general conclusion is that M1 is not very useful in predicting future GDP growth1 Compared with these studies, evidence on this with respect to euro area countries is scarce. For Germany, Kirchgässner and Savioz (2001) show that for four-quarter ahead forecasts of real GDP growth real M1 clearly outperforms forecasts based on interest rate spreads. This indicator role for M1 is also apparent in Sauer and Scheide (1995), who present evidence that there is a causal relationship from M1 to real economic activity measured by real domestic spending. Moreover, Fritsche and Kouzine (2002), find that M1 is one of the best leading indicators for business cycle turning points, measured by the index of industrial production, within a Markov switching model.2 On the other hand, in the paper by Estrella and Mishkin (1997), the one-quarter growth of M1 is not significant in an equation forecasting the annualised growth rate of GDP four to eight quarters ahead. In Plosser and Rouwenhorst (1994), past and future monetary growth only helps to predict future growth in industrial production for relatively long forecasting horizons (five years). Furthermore, they show that the term spread is a significant predictor of future M1 growth. And finally, Seitz (1998), who looks at the best leading indicators for the growth of GDP Exceptions are Swanson (1998), Vilasuso (2000) and Nelson (2002a). The first works with a special model selection procedure while the second uses detrended M1 growth that incorporates trend breaks. Nelson’s paper is discussed in detail in section 2. Moreover, Leeper and Zha (2001), using a VAR analysis, conclude that the exclusion of money from this class of models is not empirically innocuous as the interpretation of the historical policy behaviour changes substantially once money is reintroduced. This result is confirmed by Favara and Giordani (2002), who examine the role played by shocks to the LM equation in shaping the dynamic behaviour of output, inflation and interest rates. A distinctive feature of their VAR analysis is that both the variables included in the system and the identifying restrictions used to isolate shocks to the LM equation are suggested by the class of models that assign a marginal role to monetary aggregates. They also find that all the variables they considered are not block-exogenous with respect to money. While Hamilton and Kim (2002) find that M1 is significant in explaining output growth in the U.S., this finding is not robust to all extensions of the information set.
This is not true, however, when they use a probit model.
6 ECB • Working Paper No 254 • September 2003 from the 1960s until the 1990s, shows that monetary aggregates do not play a significant role within a wide range of variables.
For France, Sauer and Scheide (1995), reveal a causal relationship between real M1 and real domestic spending within a cointegration framework whereas, in Estrella and Mishkin 1997, monetary aggregates are not helpful in predicting GDP irrespective of the chosen forecasting horizon. For Italy, Sauer and Scheide (1995), interpret evidence of a common trend in M1 and real economic activity as a special case of a causal role between the two variables. Furthermore, the interest rate spread does not contain any additional information on future output developments. Comparing the information content of the term spread and M1 for real GDP in Italy, Estrella and Mishkin (1997), reveal a slight puzzle in that the spread only becomes significant once M1 is added to the relation, although the monetary aggregate itself is mostly insignificant and has the wrong sign. Altissimo et al. (2002), use two approaches to analyse the relation between surprises in GDP and innovations in monetary variables. The first requires filtering the new information contained in monetary variables by mapping surprises into estimates of the structural disturbances impinging on the variables of interest and then starting a new forecasting round of the model; the second looks directly at the correlations among surprises. The monetary variables taken into account are M2 and the currency component of M1.
Within the first approach neither M2 nor currency contribute to reducing the forecast uncertainty on GDP. In contrast, the second approach reveals that there is information in the two monetary aggregates for forecasting real GDP.
Finally, Canova and de Nicoló (2002), assess the importance of different monetary disturbances as sources of cyclical movements for the G-7 from 1973 to 1995. For that purpose they use a VAR model with industrial production as a proxy for real activity, real M1, the term spread and inflation. The major result of their paper is that the combined contribution of these monetary disturbances for real economic fluctuations is large in Germany and Italy. In Germany, there is a single monetary shock which explains the major part of output variability. In contrast, in France, monetary disturbances hardly contribute to output fluctuations. These conclusions are robust to the choice of sample period and to the inclusion of further variables, especially stock returns and short- and long-term nominal interest rates. The peculiarity with Canova and de Nicoló’s approach is that monetary disturbances are an amalgam of many different factors, not just M1.
ECB • Working Paper No 254 • September 2003 Overall, the results concerning the information content of M1 for real activity in general and real GDP in particular in euro area countries are not conclusive. Furthermore, up to now, there were only a few euro area countries under investigation.3 This study differs from the aforementioned ones in several respects. First, the role of narrow money for output has so far not been studied for the whole euro area. There are several papers dealing with the situation in individual euro area countries (see the discussion above) but the results may differ for the euro area as a whole. Second, we distinguish between different forecasting horizons ranging from one quarter to two years. Usually, only one such horizon is evaluated.4 Third, in assessing the role of M1 for output we perform an ex-post and an ex-ante analysis. And fourth, we compare the forecasting ability of different optimal VARs because the time series properties of the data and the results from preliminary model analysis are ambiguous.