Journal Issue

The Information Content of Money in Forecasting Euro Area Inflation

Emil Stavrev, and Helge Berger
Published Date:
July 2008
  • ShareShare
Show Summary Details

I. Introduction

Many Central Banks look at money as a source of information in their decision-making. For example, the European Central Bank (ECB) has assigned a special role for monetary aggregates as a cross-check of the economic analysis, supporting its inflation forecasts and risk assessment. A number of other central banks, in particular, the central banks of Japan and Switzerland, also follow monetary developments closely and incorporate monetary analysis in their core policy framework in a style reminiscent of the ECB’s two pillar approach (Assenmacher-Wesche and others 2007). In addition, the Bank of England, as well as the central banks in New Zealand, Australia, and Sweden, regularly discuss monetary developments in one form or another (OECD 2007).

At the same time, there is an ongoing debate in the literature about the relevance of money for monetary policy making. At a theoretical level, the findings are not conclusive. For instance, standard models, such as the New Keynesian Model (NKM), attribute no role for money in the inflation dynamics and, hence, monetary policy (Woodford 2003). This outcome follows, in part, from the assumption of full (intra- and inter-temporal) separability between consumption and real money balances in the households’ utility function. However, in models in which the separability assumption is lifted, money does play a structural or causal role by affecting both aggregate demand and supply and, ultimately, inflation (McCallum 2001, Nelson 2002). Another reason why monetary aggregates may matter could be informational frictions. For instance, a central bank that incompletely observes determinants of inflation, such as output, may gather relevant information from monetary aggregates, in particular if money demand is forward-looking (Nelson 2003).

With the theoretical discussion ongoing, the issue is ultimately an empirical question, but the empirical literature sends mixed signals, too. A number of studies concludes that money may provide relevant information for inflation in the euro area. Others reject this finding, and recent cross-country studies also cast doubt on a reliable link between money and inflation. One explanation for the heterogeneity in results is that the literature so far employs any number of inflation models, empirical approaches, and sample periods.

We contribute to the debate by documenting the information content of money for inflation in the euro area for a wide range of theoretical and empirical models, using a consistent framework and Bayesian and other estimation techniques. Our approach is to compare the out-of-sample forecasting accuracy of models with and without money within a given model class. The model classes include dynamic stochastic general equilibrium (DSGE) models along the lines of the NKM, partial equilibrium models ranging from the so-called two-pillar Phillips curve to the P* model, and purely empirical model such as the general dynamic factor model (GDFM) approach. In addition, we also take a look at relative forecasting accuracy across these model classes. All models are initially estimated over the same training period, 1993 to 1999, and their inflation forecasting performance is assessed for the period 2000 to 2007.

There are several interesting results. First, we find that money contains relevant information for inflation in the sense that for a number of model classes the accuracy of out-of-sample inflation forecasts increases as monetary information is added. For instance, the generalized NKM approaches that, one way or the other, incorporate money balances, perform better than the cashless NKM baseline. Time-series models, too, seem to generally support the idea that money plays a role in forecasting inflation. However, within the class of GDFM and—somewhat surprisingly—partial equilibrium models including monetary information does little to improve inflation forecasting accuracy.

A second finding is that the relative performance of the money-based models suggests a “u-shaped” relationship between the degree of their theoretical underpinnings and their forecasting performance. The best empirical model and the best DSGE model are doing better than the best partial equilibrium model. As a corollary of this, it would seem that the information content of money may not be adequately captured by the change in monetary aggregates or other transformations of the money stock, as some of these partial models assume. Instead, a more explicit modeling of the dynamics and underlying theoretical structure of the money-inflation relationship may be called for.

Finally, there are indications that the contribution of money to the accuracy of inflation forecasting out-of-sample has its limits. The improvements in forecasting accuracy are often small, not all money-based approaches add to the precision of inflation predictions within their model class, and monetary models are dominated by non-monetary models overall.

The paper is organized as follows. Section II provides a brief summary of the related literature. Section III discusses theory. Section IV explains the empirical methods and discusses the data sources. Section V presents the estimation results, and Section VI concludes.

II. Related Literature

A number of studies suggest that the indicator properties of money for inflation may be limited—either because money-based models do not perform well in a cross-country framework or because money is severely outperformed by other indicators. For example, Roffia and Zaghini (2007) and De Grauwe and Polan (2005) argue that the relationship between money and inflation across industrialized countries and time my be weaker than what is commonly thought. For the euro area, Gerlach and Svensson (2003) find that the growth rate of nominal M3 adds little to the forecasting accuracy of an output-gap based model of euro area inflation. Similarly, Stavrev (2006) finds that inflation forecasts for the euro area based on some quantity-theory inspired inflation models are outperformed by non-monetary approaches. And the OECD (2007) reports results from an euro area inflation forecasting horserace between alternative time-series models suggesting that money played a prominent role only up to 2000, but is out-performed by real time measures of the output gap as a predictor of inflation thereafter.

Others suggests that money-enhanced models can help forecasting euro area inflation. The Bundesbank (2005) and Fischer and others (2008) report that P* models contain information on euro-area inflation. Gerlach (2004) and Assenmacher-Wesche and Gerlach (2006a, 2006b) show a significant contribution of longer-run movements in money growth (appropriately filtered) in two-pillar Phillips-curve type dynamic inflation. Gerlach and Svensson (2003) find that a real money gap representation of the P* model adds to the predictive power of a conventional Phillips curve approach. Nicoletti-Altimari (2001), Hofmann (2008), and Scharnagl and Schumacher (2007) all find that M3 growth (or its trend) is useful for inflation forecasts at medium-term horizons. Finally, in a recent contribution, Berger and Österholm (2008a) confirm that M3 growth can help improve forecasting models of euro-area inflation in a mean-adjusted Bayesian VAR framework, but they also caution that the size of this improvement tends to be small. Interestingly, the result seems not to be an artifact of the particularities of the euro area: a related paper reports similar findings for M2 growth in the U.S. (Berger and Österholm 2008b).

A number of factors may explain these widely differing results. An important factor may be differences in the empirical approach. The literature has employed a wide variety of models, ranging from the purely empirical to more theory-guided. While these models incorporate money in different ways, and, in principle, allow a comparison with non-monetary approaches, these comparisons are rarely nested. This makes it difficult to identify the marginal contribution of money to inflation forecasting accuracy. In addition, the literature does not focus on a unified sample period. This could be relevant because, as D’Agostino and others (2006) stress, the predictability of macroeconomic variables has been lowered as macroeconomic volatility declined during the so-called great moderation. Finally, Woodford (2008) and Galí and others (2004) warn that establishing a structural or causal relationship between money and inflation requires use of structural dynamic equilibrium models, an approach not very common in the inflation forecasting literature so far.

In what follows, we will attempt to evaluate the information content of money in forecasting inflation in a systematic approach taking into account these considerations.

III. Models of Inflation

Our main approach is to evaluate the information content of money for inflation within three distinct classes of models: DSGE models, partial equilibrium models, and empirical models. In this section, we will describe, in turn, each model class and the potential role money plays within it. Appendix I lays out the full details of the specifications used in the empirical application.

A. DSGE Models

The cashless New Keynesian Model

The standard cashless NKM is based on an aggregate supply and an aggregate demand equation that, in combination with a monetary policy rule, determine the equilibrium path of inflation. The aggregate supply equation often combines Calvo-style price setting with price indexing (see, inter alia, Woodford 2003 and Christiano and others 2005). Profit maximizing firms setting prices take into account the current output gap and expected future price developments. In a log-linearized form around the steady state, the aggregate supply equation (or Philips curve) thus takes the form

where π^t is the deviation of inflation from its (time-invariant) steady state, xt=ytytn is the output gap, defined as the log deviation of output from its flexible-price potential, επt is a cost-push shock, and the coefficients, fulfilling θ > 0, summarize the deep structural parameters of the underlying model.1 The expectations operator E is based on the information set at period t.

Dynamic aggregate demand is derived from intertemporal household optimization based on a period utility function that may include real balances, but in the standard model is assumed to be fully separable in all its arguments. In this case, the Euler equation implies

where i^t is the deviation of the nominal interest rate controlled by the central bank from its steady state and εx,t is an aggregate demand shock. The coefficients φ > 0 summarize parameters in the households’ utility function.

Finally, monetary policy determines inflation by setting the time path of nominal interest rates to anchor inflation expectations, which closes the model. For tractability, we opt for a Taylor-type rule

where the coefficients k > 0 are defined such that that (1NKM), (2NKM), and (3NKM) describe a well-determined equilibrium. The rule states, quite plausibly, that monetary reacts with some inertia to deviations of expected inflation from its steady state (assumed to be identical with the central bank’s inflation target) and the output gap.

The model determines inflation in equilibrium without any reference to money. Of course, if money is included in the utility function as another (and separable) argument, household optimization yields a standard money demand function. But in this case monetary developments merely mirror the behavior of the model’s core variables without influencing equilibrium outcomes.

A variant of the above model introduces habit persistence in consumption. As, for instance, Fuhrer (2000) and Christiano and others (2005) show, habit persistence adds additional leads and lags of output to aggregate demand and supply and is thought to help the model capture important features of the data.

Generalizations of the NKM Model Incorporating Money

Money plays a more prominent role in generalized versions of the NKM model. For instance, taking into account financial frictions (Bernanke and others 1996, Diamond and Rajan 2006), informational frictions (Nelson 2002), or assuming non-separability of money and consumption in household utility (McCallum 2001) all introduce a structural or at least informative role for money with regard to inflation. In what follows, we will focus on the latter two arguments, which lend themselves readily to an empirical implementation.

As with the cashless baseline model, in addition, habit persistence in consumption can be added to the generalized NKM approaches.

Adjustment costs and informational frictions

One approach to give money a more important role in the NKM focuses on its potential informational function. For instance, Andrés and others (2007) argue that households have a (separable) preference for holding money, which adds money demand to the otherwise unchanged NKM framework. Assuming, in addition, that adjustment of the money stock is costly for households, real money demand is

where γ¯>0, ŷt is the deviation of the level of output from its steady state, and where m^t and Em^t+1 are the current and the expected deviation of real balances from their steady state, respectively. Note that money demand is influenced by lagged as well as expected real balances, which implies that it will reflect the entire expected future time path of interest rates and output.

Nelson (2003) then argues that in the presence of informational frictions—for instance, if current private sector shocks are unknown to the central bank (see Aoki 2006)—incorporating money into the policy function will be optimal.2 While this will also hold in the absence of adjustment costs, the weight given to monetary aggregates in the policy rule is particularly large when money demand is forward-looking. Accordingly, we adjust the Taylor-type rule to

with all κ˜>0 and where M^t represents the deviation of nominal balances from their steady state value.


Another approach relaxes the within-period separability assumption of money and consumption. This causes both the present and the expected level of real balances, m^t and Em^t+1, to enter as additional terms in the aggregate demand equation (2NKM) (e.g., Andrés and others 2006). In addition to this indirect channel, non-separability will also introduce a direct effect of mt on inflation into price setting equation (1NKM).3 Summarizing, aggregate inflation and aggregate demand become

with θ˜, ϕ˜>0. 4 Household optimization will also yield a standard money demand function

with εm,t representing liquidity preference shocks and where the coefficients γ> 0 are based on parameters in the households utility function. In addition, monetary policy is assumed to incorporate money balances along the lines of (4NKMA).

B. Partial Equilibrium Models

A popular alternative to the general equilibrium approach are the reduced form or partial equilibrium models. While less complete in their description of the economic forces driving inflation, they are not completely without theoretical background and comparatively easy to implement.

The Phillips curve with and without money

A simple Phillips curve model of inflation may take the form

where all β > 0. The model assumes that the level of inflation, in addition to showing some inertia, is a positive function of the output gap. Its empirical relevance can be contrasted with generalized versions of the model taking into account monetary factors.

One recent extension is the so-called two-pillar Phillips curve, which adds long-run excess movements in money growth equation to (1PC). The idea, broadly following the quantity theory of money, is that long-term movements of money unrelated to changes in output and interest rates help forecasting inflation (e.g., Gerlach 2004, Assenmacher-Wesche and Gerlach 2006a, 2006b). The long-term component of these variables is extracted using various filtering techniques and then added linearly to lagged inflation and the output gap,

with βx > 0, and ΔmtL, ΔytL, and ΔitL representing the growth rate of the long-term component of real money balances, real GDP growth, and the nominal interest rate change, correspondingly. In addition, the assumptions hold that αr > 0 and αm = -αy = 1.

Another specification of the two-pillar Phillips curve uses the real money gap instead of long-run excess money growth. More formally, following Gerlach and Svensson (2003), the two-pillar Phillips curve takes the form:

where, mt and mt* are the real level of money balances and their equilibrium value, respectively, while ∆mt and Δmt* are the corresponding growth rates.

The P* Model with and without money

The P* model, in its original representation, rests on the assumption that deviations of the log actual price level, p, from the log equilibrium price level, p*, the so-called price gap, is a good predictor of future price adjustments and, thus, inflation. Broadly following Svensson (2000), the price gap can be related to inflation as follows:

where α, β > 0.5

An expanded version of the P* model includes money balances. As shown in Svensson (2000), the price gap can be related to the monetary gap or excess liquidity, m˜t=mtmt*, defined as the difference between the (log) real money stock and the (log) of its equilibrium level. This implies an amended inflation model of the form

Alternatively, Reynard (2007) has interpreted p* as the equilibrium level of prices supported by the current quantity of money in circulation, given potential output, yt* , and equilibrium velocity, vt* , which might include a trend. This yields

with c > 0.

C. Empirical Models

In addition to theory-guided approaches, practitioners often rely on purely empirical models of inflation (Fischer and others 2008). Two broad classes of empirical models, time series and cross-section, seem to be particularly interesting to assess the information content of money for inflation.

Time series models

One of the simplest time series model for forecasting inflation is the univariate autoregressive model (AR), which relates current inflation to a constant and its own lags:

where ρ(L) are lag polynomial autoregressive and θ (L) lag moving average coefficients. The question is again, whether adding monetary information can improve this model in its ability to forecast inflation.

One straightforward extension of equation (1TS) is a bivariate autoregressive distributed lag model (ARDL), which adds lags of money growth,

where α(L) are lag polynomial coefficients capturing the effects of nominal money growth. An important difference between (1TS) and (2TS) is that the AR model provides an unconditional dynamic forecast of inflation, while forecasts based on the ARDL model will be conditional on money growth—which will have to be forecasted outside the model, for instance using an AR approach.

To ensure against bias in the comparison of the conditional and unconditional forecasts of inflation, it is useful to also consider a bivariate vector-autoregressive (VAR) model of inflation and money,

where A is a matrix of reduced form coefficients, j is the number of lags, and επt and εμ,t are independently identically normally distributed (iid) residuals. Berger and Österholm (2008a, 2008b) demonstrate that this framework can be used for comparing the accuracy of inflation models with and without money.

Generalized dynamic factor models

An alternative empirical framework are GDFM, which have been successfully employed to forecast inflation with large cross-section of data (Stock and Watson 1999, Forni and others 2003). The main idea is that the information content of a large set of variables can be represented by a smaller set of indicators, so-called common factors in a dynamic setting. As a result, a composite indicator can be constructed to forecast inflation. Specifically

where, xi,t, with i = 1,.., n, is a set of variables that, in addition to inflation, includes, for instance, other price variables, indicators of real activity, or financial variables. In addition, xi,t may or may not include monetary information. Furthermore, the fj,t, with j = 1,..,q, are common factors affecting all variables, λk,i, with k = 1,..,p are variable-specific loading factors, xi,tc is the common component of xi,t, which is driven by the common shocks to all data, while xi,ti is the idiosyncratic component, which is driven by pervasive shocks. The two components are identified by assuming that they are orthogonal.

The approach can be applied to the task at hand by contrasting the ability of the GDFM approach to model inflation when xi,t is restricted to exclude money with the results when it includes money.

IV. Empirical Methods and Data

The models discussed in Section III, with some modifications, can be taken to the data and used to forecast inflation. In what follows, we describe the adjustments made in the actual empirical implementation (also see Appendix I).

A. Estimation Techniques

We apply several estimation techniques, depending on the complexity of the relevant class of models. For the DSGE and the P* models, which have a large number of parameters, we use a Bayesian approach. The key advantage of this technique is that it allows a more precise inference, as the information content is broadened beyond the estimation sample by using prior distributions. The more traditional reduced-form Phillips curve models and the time series models, which are less parameter intensive, are estimated by OLS. Lastly, for the large cross-section data set, we apply modern generalized dynamic factor techniques, which improve statistical inference compared to traditional panel methods.

Wherever feasible, we estimated unobservable variables such as potential output in a model-consistent way using a Kalman filter approach. One exception is the real money gap in the DSGE models, which, for various technical reasons, was approximated using an off-model Hodrick-Prescott filter. The ex-post real interest rate is calculated by subtracting the actual next period inflation rate from the nominal interest rate.

In general, the empirical equations of the DSGE models follow the theoretical specifications discussed in the theoretical part.6 However, we introduce the lagged output gap as an additional right-hand-side variable in the aggregate demand equations in the standard NKM model as well as in the generalized NKM model with intra-period non-separability to ensure that the dynamic response of the models is better aligned with the data.7

B. Prior Distribution of Parameters for the Bayesian Estimates

To set the parameters for the prior distribution of the coefficients for the DSGE models, we use the estimates provided in Smets and Wouters (2003, 2004) and our own single equation OLS estimates. The priors for the P* models are based on OLS estimates of the parameters. Tables A2.1 and A2.2 in Appendix II provide the specifics. In what follows, we highlight some of the more important assumptions regarding the priors.

An important parameter for the dynamic properties of the DSGE and P* models is the degree of inflation inertia. Focusing first on the DSGE model, we experimented with two values of the share of backward-looking economic agents in the aggregate supply equation—θπ = 0.75 (high inflation inertia), and θπ = 0.25 (low inflation inertia), with quite tight priors in both cases (see Table A2.1). We evaluated their fit based on the ratio of the exponentials of the marginal data density of the models, which represents the relative probabilities of the models being consistent with the historical data. For the euro area, this odds ratio suggests that a Phillips curve with less forward-looking agents fits the data better. For consistency reasons, the same prior is then used for the P* model, (i.e. α = 0.75).

For both classes of models, the priors for the coefficients of the money demand equations are based on OLS estimates by using relatively diffuse priors (see Table A2.2). Another important decision regarding DSGE models concerns the coefficients in the Taylor-type monetary policy rule. Here we follow Smets and Wouters (2003) without enforcing tight priors, except for the coefficient for money, which is based on OLS estimates.

C. Forecasting and the Information Content of Money

To assess the information content of money for inflation, we compare the out-of-sample forecasting performance of the models with and without money discussed above. This is akin to the Granger-causality test suggested by Ashley and others (1980). If money does contain information relevant for the level and/or dynamics of inflation in the euro area, we would expect money-based models to deliver better out-of-sample inflation forecasts. This comparison is particularly meaningful for nested models, that is, within the various model classes discussed in Section III. However, it can also be instructive to compare forecasting accuracy across model classes.

To compare forecasting accuracy, we use the root mean square error of forecasts (RMSE). RMSEs provide a direct measure of the size of forecast errors. If h is the relevant forecasting horizon, the RMSE is defined as

where the π˜t is the inflation forecast generated by a particular model and T is the number of observations.

The forecasts themselves are generated as follows. For the simulated out-of-sample forecasts, we estimate all models for an initial training period 1993Q1 to 1999Q4 and run forecasting exercises at horizons of one, four, eight, and twelve quarters ahead starting in 2000Q1. Then the forecasting window is moved by one quarter, and new forecasts at all horizons are computed. As a rule, we do not allow the model coefficients to update during after the training period. The exception are the two model classes that incorporate conditional forecasting models depending on off-model information. This includes the class of Phillips curve models (where all models are conditional) and the class of times series models (where the bi-variate single-equation times series model is conditional on M3 growth). Since we update the conditioning off-model information each period during the evaluation sample, we opted to update the forecasting equation as well for consistency reasons.8

D. Data

We use official aggregate euro area data taken from the ECB and Eurostat websites, complemented where necessary with the corresponding series from the Area Wide Model. The data frequency is quarterly and the full sample runs from the first quarter of 1993 to the second quarter of 2007.9

For the theoretical and time-series models, the following variables in levels are used: real GDP, the CPI index, nominal M3, and the 3-month T-bill interest rate. For the GDFM models, in addition, we use the following series for each euro area member state: nominal M1 and M2, unemployment, unit labor costs, wages, the HICP index, the 10-year government bond yield, and industrial production. With the exception of the interest rates, all series are in logs and seasonally adjusted. Real money stocks are obtained using the CPI index. Quarterly growth rates for real GDP, CPI, and monetary aggregates are defined as the annualized quarterly or annual percent change of these variables.

V. Results

We present the results along a number of dimensions. First, we look at the marginal contribution of money within each class of nested models. Second, we assess its role in inflation forecast performance across all models incorporating money. And, finally, we discuss the forecasting performance of all models.

A. The Marginal Contribution of Money

Figure 1 describes the performance of the various DSGE models discussed in Section III, depicting the out-of-sample RMSEs at forecasting horizons of one, four, eight, and twelve quarters ahead. A lower RMSE implies better forecasting accuracy at a particular horizon.10

Figure 1.Forecast Performance of DSGE Models

(RMSE of dynamic forecasts, in percentage points of year-on-year inflation)

Citation: 2008, 166; 10.5089/9781451870244.001.A001

The results suggest that the generalized NKM approaches that, one way or the other, incorporate money balances perform well compared to the cashless NKM baseline (Figure 1). The introduction of habit persistence, while improving the forecasting performance of the DSGE models overall, does not alter this result. The main difference between the right and the left panel is that the adjustment-cost model delivers the best forecasting performance in the case of NKM models with habit persistence, while the non-separability model shows the lowest RMSE of the NKM models without habit persistence.

The size of the improvements in forecasting accuracy from adding money varies with the time horizon but is limited on average. For example, at the two-year horizon, for the NKM without habit persistence, moving from the cashless baseline NKM model to a model with non-separability reduces the RMSE by as much as 0.25 percentage points, while for the NKM with habit persistence moving from the baseline to the model with adjustment costs lowers the RMSE by almost 0.55 percentage points of inflation.11 However, at 0.2 and 0.3 percentage points, respectively, the improvements across all forecasting horizons are smaller. Nevertheless, these results seem to contradict Woodford’s (2003, 2006) argument that there is no role for money in inflation determination in the NKM approach and are broadly in line with Andrés and others (2007), who claim that a NKM incorporating money provides a good empirical description the euro-area economy.

An additional result from the DSGE models with non-separability is that money affects inflation mainly through improving the empirical specification of the aggregate demand and aggregate supply equations, while the effect from monetary policy reaction to money balances is small. We checked the importance of the monetary policy channel by comparing the out-of-sample forecasting performance of the model with and without money and found very similar differences in the implied RMSEs.12

From general equilibrium models, we move to the other extreme, purely empirical models. Figure 2 presents the results in the now familiar setup. The results from the time-series models (left panel) indicate that money does improve inflation forecasting. The ARDL single equation approach containing money, while not beating the random walk model, delivers better inflation forecasts than the ARMA approach. The best time-series model in terms of RMSEs across all horizons is the bivariate VAR model. At about 0.5 to 1.0 percentage points depending on the horizon, the reduction in RMSE gained by the bivariate VAR model against the ARMA is considerable, but the improvement against the random walk model is modest at best. These results are consistent with other studies in the literature that have used time-series models. For example, Berger and Österholm (2008a), also using a Bayesian VAR approach, find that including money allows better predictions of inflation, while Fischer and others (2008) show that bivariate monetary forecasts bring new information to the broad macroeconomic projections.

Figure 2.Forecast Performance of Empirical Models

(RMSE of dynamic forecasts, in percentage points of year-on-year inflation)

Citation: 2008, 166; 10.5089/9781451870244.001.A001

The results from the GDFM models, however, do not suggest that incorporating monetary indicators appreciably improves inflation forecasting performance once a broader set of economic indicators is taken into account (Figure 2, right panel). While the GDFM model with two factors including various monetary indicators does better in forecasting inflation than the two-factor model without money, the reduction in RMSEs is miniscule. Turning to the one-factor GDFM models, we find that the opposite holds: the model without money outperforms the one-factor model with money, even though the differences in RMSE levels remain very small in this case as well. This result seems to run against Hofmann (2006), who reports that factor models with money do consistently better than factor models without money in forecasting inflation over the twelve-quarter horizon.

Finally, we consider partial equilibrium models. Other than the results from the DSGE and empirical models, the outcome from the partial equilibrium models imply no role for money (Figure 3). More specifically, the simplest P* model with no money has the best forecast performance in this class of models, with the two models with money lagging far behind (left panel). Similarly, we find that among the reduced-form Phillips curve models, the standard formulation excluding money produces the best forecast among this group of models at longerhorizons, with the two-pillar Phillips curves, and, in particular, the one with long-term money growth, faring much worse.

Figure 3.Forecast Performance of P* and Phillips Curve Models

(RMSE of dynamic forecasts, in percentage points of year-on-year inflation)

Citation: 2008, 166; 10.5089/9781451870244.001.A001

The relatively weak performance of the money-based P* and two-pillar Phillips curve models, two approaches that are used widely by forecasting practitioners, runs across some of the recent contributions in the literature (see Section II). The results for the P* models are surprising, given that we estimate the trend of money balances in a model-consistent way, which, in principle, delivers better forecasting performance than single equation filters, such as a Hodrick- Prescott filter. One possible explanation could be parameter uncertainty with regard to the money demand function. A factor relevant for the two-pillar Phillips approach may be the choice of the filter used to identify long-run movements in the data, where Assenmacher- Wesche and Gerlach (2006a, 2006b) apply fairly sophisticated frequency-based techniques.

B. Comparison of Money-Based Models

We now turn to comparing forecasting performance of money models (see Table 1). An obvious disadvantage of this perspective is that this involves contrasting non-nested models, but it still seems interesting to learn about the relative performance of the various money-based approaches in unified empirical setting. Table 1 presents the results in the columns dubbed “within money models”, which report the relative rank of all relevant models based on the RMSE for the twelve-quarter ahead forecasts and the average RMSE across all horizons, respectively.

Table 1.Out-of-sample Forecasting Performance of Models

(RMSE of year-on-year inflation forecasts, in percentage points)

Ranked by 12QRanked by av. 1-12Q









Theory-based models
Dynamic stochastic general equilibrium models
New Keynesian model (NKM)
NKM with non-separability0.160.681.000.730.6411361237
NKM with adjustment costs0.210.941.160.860.7914491559
NKM with habit persistence0.1770.811.110.870.74166144
NKM with habit persistence and non-separability0.130.520.740.640.51925925
NKM with habit persistence and adjustment costs0.140.420.570.640.44814613
P-star models
P* without money0.140.440.560.510.417121
P* with money0.170.851.912.971.472031219311
P* with estimated trends of money and velocity0.130.411.001.700.811821016210
Reduced form Phillips curve models
Traditional Phillips curve0.450.310.250.230.311111
Two-pillar Phillips curve with real money gap0.310.590.790.860.6413281126
Two-pillar Phillips curve with long-run money growth1.911.992.012.072.001931120312
Empirical models
Time series models
Bivariate VAR with inflation and M3 growth0.310.410.460.490.42613311
Single equation with inflation and M3 growth0.330.880.810.760.7012271328
Generalized dynamic factor models (GDFM)
One-factor GDFM with money0.390.550.560.380.47542734
One-factor GDFM without money0.390.550.570.380.474384
Two-factor GDFM with money0.360.560.470.290.4221/a141/a2
Two-factor GDFM without money0.360.560.470.290.423252
Memo item
Random walk0.310.520.680.680.551010

The two factor GDFM model with money shows an improvement over GDFM model without money at the third decimal point.

The two factor GDFM model with money shows an improvement over GDFM model without money at the third decimal point.

There are some surprising results. For example, DSGE money models, which are not necessarily known for their good overall forecasting performance, fare fairly well among money models. Out of twelve money based models overall, the best generalized NKM approach ranks third. It is also interesting to note that the two-pillar Phillips curve with real money gap also scores relatively high (rank six), supporting results in Gerlach and Svensson (2003). The best relative forecasting performance, however, comes from the bivariate VAR model, followed by the factor models containing monetary information (ranks one and two). Partial equilibrium approaches perform worst.

One implication of the relative performance of the money based models is the seemingly “u-shaped” relationship between the degree of their theoretical underpinnings and their forecasting accuracy. The best empirical model and the best DSGE model are doing better than best partial equilibrium model. A corollary of this is that, from a forecasting standpoint, the information content of money is not necessarily adequately captured by partial equilibrium models—instead, a more explicit modeling of the dynamics and underlying theoretical structure of the money-inflation relationship seem to be called for.

C. Comparison Across All Models

Finally, we compare forecast performance across all models. While this again involves comparing non-nested models, the practice is common in the literature. The relevant rankings are displayed in the columns named “overall” in Table 1. Strikingly, the best performing model is an approach without money—the traditional Phillips curve. This holds both at the horizon of twelve quarters and for the average performance over all forecasting horizons. At about 0.1 percentage points of inflation, the difference to the runner-up monetary models is not overwhelming, however. For most models the overall ranking does not change substantially whether the assessment is done over all horizons or based on the twelve quarter forecast. But note that the GDFM approaches with and without money do best at longer horizons, while the P* model class, which does not perform well overall, does better than most at very short horizons.

Overall, these results warn against overestimating the contribution of money to the accuracy of inflation forecasting out-of-sample. While some money-based approaches do very well within their particular class of theoretical or empirical models, all are dominated by models excluding money in all-out the horserace.13

VI. Conclusions

This paper contributes to the debate on the role of money in monetary policy by analyzing the information content of money in forecasting inflation in the euro area. Our approach is to compare the predictive performance within a range of structural and empirical model classes. All models are estimated over the same initial training period, 1993 to 1999, and their inflation forecasting performance is assessed in a simulated out-of-sample exercise for the period 2000 to 2007.

We find that money contains relevant information for inflation. For instance, New Keynesian DSGE approaches incorporating money balances and money-based VAR models perform better in forecasting inflation than their cashless counterparts.

In addition, the comparison of the forecasting performance of money-based models points at a “u-shaped” relation between theoretical underpinnings and the prediction accuracy of these models: the best money-based VAR and DSGE models dominate the best partial equilibrium P* or two-pillar Phillips curve models. This seems to suggest that the information content of money is best captured by explicitly modeling the dynamics and underlying theoretical structure of the money-inflation relationship.

But there are also signs that the contribution of money to the accuracy of inflation forecasting out-of-sample has its limits. First, we do not find that including monetary information aids forecasting accuracy within the class of modern dynamic factor models (here the contribution is miniscule) and, surprisingly perhaps, partial equilibrium models (where it seems to be even negative). Second, we find that, across models overall, non-monetary approaches provide better inflation forecasts than money-based models. While the differences in RMSEs are not too large, this results cautions against overemphasizing the role of monetary information in inflation forecasting

Appendix I: Empirical Specifications

In this appendix we discuss the empirical specifications used to estimate the DSGE (or NKM) and P* models. All variables in the appendix are defined in levels.14

The cashless New Keynesian Model

Phillips curve

where, πt is quarterly seasonally adjusted inflation at an annual rate, π4,t-1 is year on year inflation, E, is the expectation operator based on the information available as of time t, xt is output gap (defined below), and επ,t is an error term that captures supply shocks. Note that the following restriction on lagged and expected inflation is imposed, θπ = 1 − θEπ and annual inflation is defined as the backward moving average of quarterly inflation, π4,t = (πt–3 + πt–2 + πt–1 + πt)/4.

Aggregate demand

where, i is the nominal interest rate, rtn is the real natural interest rate, and εx,t is an error term that captures demand shocks. Note that we deviate here from the theoretical aggregate demand curve described in Section III.A by adding the lagged output gap.

Monetary policy reaction function

Where xt is defined as the difference between the log of real GDP yt and its equilibrium level ӯt (defined below), it is the nominal interest rate, π* is the central bank’s inflation target, and εi,t is an error term that captures monetary policy shocks.

In addition to the above main equations, we also add several equations to estimate the model- consistent equilibrium output, long-term real GDP growth, and real interest rate.

Equilibrium output

where, ӯt is log of the level of equilibrium real GDP, gt is the quarterly real GDP growth at annual rate, and εӯ,t is a shock to equilibrium output.

GDP growth

where g* is steady state real GDP growth and εg is a shock to steady state growth.

Real natural interest rate

where, r* is steady state real interest rate and εrtn,t is a shock to equilibrium real interest rates.

The above system of equations, (A1)-(A6), is used to estimate the cashless NKM.

Adjustment costs

In the case of the NKM with adjustment costs, the aggregate supply and the aggregate demand equations remain as specified in (A1) and (A2) above. The money demand equation takes the form,

where mt is real money gap, defined as the percent difference of real money balances from their Hodrick-Prescott (HP) filtered trend. The monetary policy reaction function is defined as,

where, M¯t is the deviation of nominal balances from their steady state, computed using mt, πt, and π*.

To estimate the NKM with adjustment costs, we replace equation (A3) with equation (A8) and add equation (A7) to the rest of the system to obtain the following system of equations (A1), (A2), (A8), (A4)-(A7).


In the case of the NKM with non-separability, the monetary policy reaction function is as in the NKM model with adjustment costs, equation (A8). The Phillips curve and the aggregate demand equations are described below.

Phillips curve

Again, as in the standard NKM, the sum of the coefficients on lagged and expected inflation is one.

Aggregate demand

The money demand function is defined as,

The NKM model with non-separability is estimated by replacing equations (A1) and (A2) with equations (A9) and (A10), correspondingly, and equation (A7) with equation (A11) in the system of equations used to estimate the NKM with adjustment costs, (A1), (A2), (A8), (A4)- (A7). This results in the following system of equations: (A9), (A10), (A8), (A4)-(A6), and (A11).

NKM with habit persistence

In the case of NKM with habit persistence the aggregate supply, aggregate demand, and money demand equations are defined as follows:

Phillips curve

Aggregate demand

Money demand

The NKM with habit persistence is estimated by replacing in the system of equations for the NKM model with non-separability—(A9), (A10), (A11)—with equations (A12), (A13), and (A14), which results in the system (A12), (A13), (A8), (A4)-(A6), and (A14).

NKM with habit persistence and non-separability

In the case of NKM with habit persistence and non-separability the aggregate supply, aggregate demand, and money demand equations are defined as follows:

Phillips curve

Aggregate demand

Money demand

The NKM with habit persistence and non-separability is estimated by replacing Phillips curve (A12), aggregate demand (A13), and money demand (A14) equations in the system of equations for the NKM model with habit persistence—(A12), (A13), (A8), (A4)-(A6), and (A14)—with equations (A15), (A16), and (A17), correspondingly, which results in the system (A15), (A16), (A8), (A4)-(A6), and (A17).

P* model without money

Inflation equation

where, π is quarterly inflation, π4,t year-on-year inflation, Eπ4,t+4 is expected inflation (both year-on-year and expected inflation are defined as in the NKM models), and p˜t is the gap between the equilibrium and actual price levels.

Price gap equation

where, pt* is the equilibrium price level and pt is the actual price level.

Equilibrium inflation

where, π* is the steady state inflation, while π¯ is the medium-term equilibrium inflation.

Equilibrium price level

where, εp*,t is a shock to the equilibrium price level.

Actual price level

p* model with money

Inflation equation

where,m˜t1 is the real money balances gap, defined as the difference between the logs of actual ( m ) and equilibrium (m¯) money balances.

Real money gap

where, εm˜,t is shock to real money gap.

Equilibrium real money balances

where, μ* is steady state real money growth and εm¯,t is a shock to equilibrium real money balances.

Equilibrium money growth

where, μt is real money growth and εμ,t is a shock to medium-term money growth.

Actual real money stock

p* model with money and trend velocity

Inflation equation

Equilibrium output

Output growth

All variables in equation (A30) are defined as in equation (A5).

Actual real GDP

Actual interest rate

where, it is nominal interest rate and i* is equilibrium nominal interest rate.

Trend in equilibrium velocity

We generalize more conventional representations of the P* model by modeling a trend in velocity. In particular,

where, ∆vt is the quarterly change in velocity, ∆v* is the equilibrium change in velocity, and ε∆v is a shock to the change in velocity.

Equilibrium velocity

where, vt* is equilibrium velocity.

Equilibrium money supply

Real money gap

Real money

Money growth

Phillips curve models

See main text.

Appendix II: Bayesian Priors

Here we provide the priors for the DSGE and P* models used for the Bayesian estimation.

Table A2.1.Prior Distribution of the Parameters for the DSGE Models 1/
ParameterMean DistributionSt. dev.ParameterMean DistributionSt. dev.
Common parametersModel-specific parameters
New Keynesian model
θπ0.25 beta0.05NKM with adjustment costsγ¯yt10.15 gamma0.10
θxt0.25 gamma0.05γ¯yt0.75 gamma0.10
ϕxt10.75 gamma0.10γ¯Ey0.20 gamma0.05
ϕExt+10.15 beta0.05γ¯i1.00 gamma0.10
φi0.20 gamma0.05γ¯mt10.15 gamma0.05
ρi0.80 beta0.05γ¯Em0.15 gamma0.05
κEπt+11.50 gamma0.20κ˜Mt0.15 gamma0.05
κEπt+10.50 gamma0.05Non-separabilityθ˜m0.15 gamma0.05
τ0.10 beta0.05ϕ˜m0.15 gamma0.05
g*2.20 normal0.50ϕ˜Emt+10.14 betta0.05
ρ0.20 beta0.07Habit persistanceθ^Em0.15 gamma0.05
θ^xt10.05 gamma0.05
r*2.00 normal0.50θ^Ex0.05 gamma0.05

As the models are nested, the versions with money contain the parameters of the cashless NKM. The model- specific coefficients are for the lags/leads of the aggregate demand/supply equations and money in the policy rule.

As the models are nested, the versions with money contain the parameters of the cashless NKM. The model- specific coefficients are for the lags/leads of the aggregate demand/supply equations and money in the policy rule.

Table A2.2.Prior Distribution of the Parameters for the P* Models
ParameterMean Distribution

Common parameters
St. dev.ParameterMean Distribution

Model-specific parameters
St. dev.
No moneyα0.25 beta0.05With estimated trends of money and velocityψ2.10 normal0.30
β0.25 gamma0.05σ108.40 normal50.00
τ0.10 beta0.05τ0.30 normal0.10
π *2.00 normal0.30τ10.03 normal0.10
With moneyα0.25 beta0.05π*2.00 normal0.30
g *2.50 normal1.00
β0.25 gamma0.05μ *4.00 normal1.50
τ0.30 normal0.10i*4.50 normal1.00
π *2.00 normal0.30Δ v*-0.50 normal1.50
g *2.50 normal1.00ρ0.30 normal0.10

We would like to thank Henning Weber for helpful comments and suggestions.

All variables of the NKM models discussed in Section III.A are expressed as deviations from their steady state levels, while Section III.B as well as Appendix I use levels.

An alternative argument for incorporating money in the interest rate equation could be that money, for reasons not captured in the structural model, helps to forecast inflation or the central bank shares the households’ preference for a stable money stock. See Andres and others (2007).

The reason is that forward-looking price setters apply the households’ stochastic discount factor in their dynamic optimization problem, which, in turn, is influenced by money holdings.

Note that the direct and indirect impact of contemporaneous money on inflation move into opposite directions, with the overall impact being a question of the underlying parameters and, thus, ultimately an empirical matter.

Equation (1P*) combines a number of approaches. Hallman and others (1991), for instance, assume α= 1, which seems to be in line with Reynard’s (2007) empirical observation that the price gap influences inflation with considerable lags and some persistence. As a rule, the P* model does not restrict the expectations term, which, in principle, could take any form.

Note that the coefficients estimated for the DSGE models are semi-structural, as we do not enforce some of the restrictions implied by the deep parameters of the underlying model in the empirical implementation.

More specifically, the assumption of no inertia in aggregate demand is rejected by the data despite tight Bayesian priors.

Note that the within-model class comparison is not affected by this convention. To forecast the off-model conditional variables—in particular, the output and real money gaps and long-term money growth for the Phillips curve models, and money growth for the ARDL model—we proceed in two steps. For a given forecasting window, we first calculate these variables using a standard Hodrick-Prescott filter and then forecast them up to twelve quarters ahead using an ARMA process. This process is repeated each time the forecasting window is moved and a new observation is added.

The fact that our simulated out-of-sample exercise is based on revised instead of real time data should have no bearing on the results. A bias (if any) in forecasting accuracy that the use of revised data may create would influence all models simultaneously, leaving their relative performance unchanged. And indeed, in a recent paper, Faust and Wright (2007) show that the use of real time or revised data does not influence the relative performance of various model- and expert-based forecasts for inflation and output in the U.S. economy.

Table 1 further below, in addition, shows the relative rank of the models within each nested model class (see columns named “within model class”) based on the relative RMSE at the twelve quarter horizon and the average RMSE across all horizons, respectively.

Note that this is not simply an artifact of the NKM model lacking the ability to capture the persistence in the data, which may influence forecasting accuracy. In the empirical implementation, all estimated equations include lagged endogenous variables (see Section III and Appendix I).

Additional results available on request.

In part, this might be explained by the fact that our approach emphasizes within-class comparisons, and that we have made no effort to fine-tune the performance of a particular class of model vis-à-vis another. On the other hand, it is not immediately obvious how such an effort would alter the results.

In general, we follow the notation of the main text. For ease of exposition, we use Roman letters to abbreviate growth rates where required.

Other Resources Citing This Publication