Jointly Evaluating the Federal Reserve s Forecasts of GDP Growth and Inflation

Transcription

1 Jointly Evaluating te Federal Reserve s Forecasts o GDP Growt and Inlation Edward N. Gamber Department o Economics and Business Laayette College Easton, PA gambere@laayette.edu H.O. Stekler Department o Economics George Wasington University Wasington, DC stekler@gwu.edu Tara M. Sinclair Department o Economics George Wasington University Wasington, DC tsinc@gwu.edu Elizabet Reid Department o Economics George Wasington University Wasington, DC ereid@gwu.edu RPF Working Paper No ttp:// Original Version:* April 15, 2008 Tis Version: Marc 8, 2011 RESEARCH PROGRAM ON FORECASTING Center o Economic Researc Department o Economics Te George Wasington University Wasington, DC ttp:// *Tis paper was originally titled: Multivariate Forecast Evaluation Based on Linear Policy Rules Researc Program on Forecasting (RPF) Working Papers represent preliminary work circulated or comment and discussion. Please contact te autor(s) beore citing tis paper in any publications. Te views expressed in RPF Working Papers are solely tose o te autor(s) and do not necessarily represent te views o RPF or George Wasington University.

2 Jointly Evaluating te Federal Reserve s Forecasts o GDP Growt and Inlation Tara M. Sinclair 1 Department o Economics George Wasington University Wasington, DC tsinc@gwu.edu H.O. Stekler Department o Economics George Wasington University Wasington, DC stekler@gwu.edu Edward N. Gamber Department o Economics and Business Laayette College Easton, PA gambere@laayette.edu Elizabet Reid Department o Economics George Wasington University Wasington, DC ereid@gwu.edu JEL Codes: C53, E37, E52, E58 Keywords: Evaluating Forecasts, Macroeconomic Forecasts, Loss Function, Inlation Forecasting, GDP Growt Forecasting, Monetary Policy Tis Drat: Marc 8, 2011 Abstract In tis paper we jointly evaluate te Federal Reserve sta orecasts o U.S. real output growt and te inlation rate assuming te orecasts are to be used as inputs or te Taylor rule. Our simple metodology generates policy orecast errors wic ave a direct interpretation or te impact o orecast errors on te target interest rate given by te Taylor rule. Witout interest rate smooting, we ind tat, on average, te Taylor rule target interest rate would ave been approximately a ull percentage point away rom te intended target because o errors in orecasting output growt and inlation. Our results are robust to canges in te orecast orizon and to canges in te weigts on te variables in te policy rule. 1 Corresponding autor. Pone: Fax: Te autors wis to tank Mike Bradley, Dean Crousore, Fred Joutz, Kajal Lairi, George Monokroussos, Keit Ord, Micael Owyang, Roberto Sameniego, Peter Tinsley, and Harry Watson or elpul discussions. We also tank participants in te GWU Brown Bag Seminar Series on Forecasting, te 16 t Federal Forecasters Conerence, te 28 t Annual International Symposium on Forecasting, te 1st Macroeconomic Forecasting Conerence in Rome, West Virginia University Economics Department Seminar, Ricmond University Economics Department Seminar, University o Mempis Economics Department Seminar, University at Albany Economics Seminar, and Wesleyan University Economics Department Seminar or teir comments and suggestions. Warren Carnow, Alexander Kapinos, and Blake Saville provided excellent researc assistance. Generous support or tis project was provided by te Institute or International Economic Policy at te Elliott Scool o International Aairs and te GW Institute o Public Policy.

3 1. Introduction Forecasts are an essential input into all decisions involving te uture. Te accuracy o tese orecasts can determine te adequacy, validity, and credibility o bot macro and micro economic decisions. It is or tis reason tat te accuracy o tese orecasts is evaluated. However, tere are dierent perspectives or approaces or conducting tese evaluations. In one approac, te orecaster s accuracy is evaluated wile anoter metod considers te value o te orecast to its users. Te traditional approac or evaluating accuracy as been to measure te magnitude o te error tat is associated wit te orecasts o eac particular variable. 1 I an individual or organization orecasts bot te inlation rate and te GDP growt rate or next year, te accuracy o te orecast o eac variable is calculated separately. Tis is te approac tat as been previously used to evaluate te orecasts o te Federal Reserve s sta (Clements et al., 2007; Joutz and Stekler, 2000; Romer and Romer, 2000; Sims, 2002; Stekler, 1994). 2 Tere is, owever, a problem tat is associated wit tis traditional orecast evaluation approac: te loss unction may contain more tan one argument. Tis paper addresses tis problem, ocusing on te orecasts o te Federal Reserve (Fed). Te Fed is concerned wit bot inlation and output growt. Tis suggests tat te accuracy o bot variables sould be evaluated simultaneously. Sinclair et al. (2010) provided te procedures or jointly evaluating te directional accuracy o te Fed s orecasts but did not consider ow to jointly evaluate te quantitative predictions. 1 Usually mean square error is te error metric because it is associated wit a quadratic loss unction. 2 Wile most evaluations ave ocused on te accuracy o a particular variable, tere ave also been discussions about te appropriate procedures or jointly evaluating multiple orecasts (See Hymans, 1968; Clements and Hendry and subsequent comments, 1993; Eisenbeis, et al., 2002; and Komunjer and Owyang, 2007) or or comparing orecasts o dierent variables (Lairi and Seng, 2010). However, tese evaluations involving multiple variables were not done in te context o policy loss unctions or decision rules. 1

4 I te intended use o te Fed s orecasts is known, it may be possible to jointly evaluate te accuracy o te quantitative growt rate and inlation orecasts. In te case o monetary policy, tere as been considerable discussion tat some orm o te Taylor Rule is a good approximation or te decision rule used by te Fed in setting teir target interest (i.e. ederal unds) rate (see, or example, te collection o articles in Taylor, 1999). By assuming tat tis is te decision rule, we can jointly evaluate te GDP and inlation orecasts o te Fed and estimate te size o te errors resulting rom te use o tese orecasts in te Taylor rule. We ask te ollowing question: Given tis speciic policy rule, wat would be te impact o orecast errors on te implied ederal unds rate target? 3 Our work is in te spirit o evaluating orecasts by considering te economic costs o prediction errors (Clements, 2004; Granger and Pesaran, 2000a and 2000b; and Pesaran and Skouras, 2002). It is also related to te work by Orpanides (2001, 2003) on te impact o realtime data on monetary policy. Our contribution is to draw tese two lines o researc togeter to speciically evaluate orecasts in te context or wic tey are prepared. We next present data used in tis analysis. We evaluate te Greenbook orecasts tat are made by te sta o te Board o Governors o te Fed. Tese orecasts are produced beore eac meeting o te Federal Open Market Committee and contain projections on te economy up to eigt quarters into te uture. We ten present a metodology or evaluating policy orecast errors assuming tat te Fed ollows te Taylor rule. Subsequent sections present te results, robustness cecks wit respect to te speciication o te policy rule, and our conclusions. 3 Tere is muc debate in te literature on te Taylor Rule as to te speciic orm o te rule. We sow in section 6 tat our results are robust to te weigts cosen or te variables in te policy rule. We urter sow tat interest rate smooting will reduce te impact o te current orecast errors. 2

5 2. Data We examine te Greenbook orecasts or eac quarter rom troug , but, or some analyses, also divide te data into two subperiods, and Tese periods correspond to te pre-volcker and Volcker-Greenspan eras at te Fed. Te projections used in tis analysis are te annualized growt rate o real output (GNP rom 1965 to 1991 and GDP rom 1992 on) and te annualized inlation rate (based on te implicit price delator troug te irst quarter o 1996, ten te cain-weigted price index rom on). 5 We only analyze te current quarter and one quarter aead projections. We ocus on sort orizons because te Greenbook orecasts ave sometimes been based on an assumed (possibly varying) pat or monetary policy. At oter times, owever, te orecasts assumed tat monetary policy would remain uncanged over its orecast orizon (Reiscneider and Tulip, 2007). Since te assumed pat or monetary policy associated wit eac Greenbook orecast is not known, a possible complication arises wen analyzing longerterm orecasts. Te current quarter and one-quarter aead orecasts are too sort o a time orizon, owever, to be aected by te Fed s uture pat or monetary policy. Tereore, regardless o weter te Fed assumes a constant pat or a varying pat or monetary policy, te current and one-quarter-aead orecasts will be unaected by tose assumptions. Tere are at least two orecasts per quarter, eac made in various monts o te quarter. Consequently te orecasts made or te current and next quarter ave leads o zero to ive monts to te end o te relevant quarter. Te orizons o te orecasts vary between zero and ive monts: orizons o (a) zero to two monts and (b) tree to ive monts or te current and next quarters, respectively. Because orecasts were not made at all orizons in every quarter, te 4 Te Greenbook data are only available wit at least a 5-year lag. 5 We use te compounded annual rate o cange to construct te growt rates. 3

6 number o observations diers between orizons. Te actual values were te data available approximately 90 days ater te end o te quarter to wic tey reer Te Fed s Policy Rule "Even i te actual policy process is ar more sopisticated tan any simple rule could completely describe, te [Taylor] rule oten provides a reasonably good approximation o wat policymakers decide and a ramework or tinking about policy actions." --Federal Reserve Board Vice Cairman Donald L. Kon (2007) In order to analyze te eects o te Fed s orecasts errors on its selection o a target interest rate, it is necessary to irst establis te relevant policy rule. We assume te Fed eiter ollows te original Taylor rule (Taylor, 1993) or uses it as a guide in setting monetary policy. Altoug Taylor originally proposed is rule as an empirical description o past Fed policy actions, Woodord (2001a, 2001b) as sown tat te Taylor rule can also be teoretically justiied as a prescriptive rule. Clarida et al (2000) suggested tat it sould be orward-looking, employing expected uture values o bot output and inlation. As pointed out by Orpanides (2001), even i te Fed is using current values o bot output and inlation, te data te Fed as at te time tey are setting policy is not te actual values, but predictions o tose values. Rater tan ocusing on te various issues associated wit tis rule, we start wit te ixed weigts proposed by Taylor (1993). We sow in Section 6 tat our results are robust to te range o weigts considered in te literature. 6 Te use o te real time data avoids deinitional and classiication canges. Our results are robust to wen te data publised days ater te end o te quarter are used. Our real-time data were obtained rom te ArcivaL Federal Reserve Economic Data (ALFRED ), maintained by te Federal Reserve Bank o St. Louis at ttp://alred.stlouised.org/. 4

7 According to te orward-looking Taylor rule, te Fed sets a target ederal unds rate, i T t, based on equation (1), were, te superscript denotes tat te target is based on orecasted variables. 7 Te Fed s policy decision P + ) is written as: P ( t, t T t, = it π Y = r * + π + 0.5( π *) + 0.5(lnY ln *), (1) were r* is te equilibrium real interest rate, π* is te Fed s implicit inlation rate target, and Y* is potential output. Wile te output gap (i.e. lny lny*) is typically used in te Taylor rule, te growt rate o real GDP is typically used in orecast evaluation. For our application we replace te output gap by te GDP growt rates, yielding te ollowing policy rule: P T t, = it π y = r * + π + 0.5( π *) + 0.5( y *), (1 ) were y* is te potential output growt rate. 8 Te Fed orecasts bot inlation, π, and output growt, y t +, periods aead. Te actual outcomes, owever, will in general dier rom te Fed s orecasts. I te actual values or π and t were instead included in te Taylor rule (i.e. i te Fed ad t + y + perect orecasts or perect oresigt), a potentially dierent target ederal unds rate would ave been suggested by te policy rule. Consequently, under perect oresigt te implied interest rate target would ave been: T A A A Pt, t + = it = r * +πt ( πt + π*) + 0.5( yt + y*), (2) 7 Following Orpanides (2001), we assume tat te Fed uses te Greenbook orecasts in teir decision rule. 8 I we assume tat potential output, Y*, is eiter known or a constant, ten we can dierence out potential output wic permits us to use te growt rate in order to construct te policy rule. A similar approac as been used by Hamilton et al. (2009). For a discussion o te role o real time output gap estimates and te Taylor rule, see Orpanides (2001). 5

8 were A π t + and A y t + represent te actual realizations o t + y + π and t. Te dierence between i T and i T measures te dierence in te Fed unds rate suggested by te Taylor rule tat occurs because o inaccurate orecasts and tus represents te implied policy orecast error, PFE t : T T A A t t t ( ) + 0. ( yt + yt ) A A Te dierences, ( π ) and ( y ) PFE = i i = 1.5 π π 5 +. (3) t t + + π y t t + +, are te Fed s orecast errors or te inlation rate and real output growt respectively. Given te PFEs, te evaluation procedures are ten similar to tose used in judging individual orecast errors Evaluation Procedure: Standards o Comparison Tere are two sets o orecasts tat we use as standards o comparison: te PFEs tat would ave occurred i (1) naïve orecasts or (2) te median orecasts rom te Survey o Proessional Forecasters 10 ad been used to set policy. In eac case, te comparison uses te same weigts and as te same number o observations as were used in constructing te PFEs o te Fed orecasts tat we are evaluating. We calculate te mean absolute policy orecast error (MAPFE) and root mean squared policy orecast error (RMSPFE) or te ull sample as well as or a sub-sample were we ave te median orecasts rom te Survey o Proessional Forecasters (SPF) available or comparison. Te naïve orecast assumes tat te same growt rate o output and te same inlation rate as were observed in te previous quarter will occur in te uture period(s). 11 We also use te 9 For a discussion o te impact o using projections on te estimates o te Taylor rule, see Orpanides and Wieland (2008). We do not use te interest rate, te equilibrium real interest rate, te Fed s implicit inlation rate target, or potential output in our analysis. Considerable researc as gone into estimating tese variables (e.g. Clark and Kozicki, 2005, or te equilibrium real interest rate and Leig, 2008, or te Fed s implicit inlation rate target). Wile tese variables may be time-varying, we treat tem ere as known to te policymaker and tey tus drop out o our expression. Tis assumption allows us to isolate te impact o orecast errors. 10 Tese data come rom te Federal Reserve Bank o Piladelpia (or a description o tis survey, see Crousore, 1993). 11 We construct tese naïve orecasts rom te real-time data, publised in te Greenbook, o GDP growt and te inlation rate or te quarter beore te Greenbook orecasts were made. Tis assures tat te same inormation as 6

9 median orecasts rom te SPF to construct orecasts comparable to our Greenbook data. Since te SPF orecasts are only reported once per quarter (in te middle o te second mont o eac quarter), tey are comparable to te Fed orecasts made wit one and our mont orizons. Te SPF orecasts are available beginning wit te ourt quarter o Results For te entire sample 1965:4-2002:4, te mean absolute policy orecast error (MAPFE) based on Fed orecasts at te zero mont orizon is 101 basis points (Table 1). Tis can be interpreted as te amount by wic te implied target interest rate based on te Taylor rule would ave diered by using orecasts rater tan te actual data in te policy rule. 12 According to tis analysis, te implied target interest rate based on te Taylor rule was at least a ull percentage point o because o errors in orecasting growt and inlation at te zero orizon. 13 We also tested or bias, but te null o no bias was usually not rejected. Table 1 also presents te root mean squared policy orecast errors (RMSPFE) or te samples and orizons we analyze in our subsequent tests. As expected, bot te MAPFE and RMSPFE in general increase wit te lengt o te orecast orizon and te MAPFE and RMSPFE are bot smaller tan tose o te naïve and SPF orecasts in all cases. Tis result is consistent wit previous indings rom evaluations o a single variable. A comparison o te results or te Pre-Volcker sample (Panel B) and te Volcker- Greenspan sample (Panel C) suggests tat tere as been a small reduction o bot MAPFE and was available to te Fed orecasters at te time te predictions were made was used in constructing tis alternative orecast. 12 Bernanke (2010) pointed out tat using orecasts in te Taylor rule instead o actual values may explain te dierence in te interpretation o te stance o monetary policy rom 2002 to Based on orecasts it appears tat te Fed was close to te recommended policy in te Taylor rule. Tereore te intended policy o te Fed was to approximately ollow te Taylor rule. I, as Taylor (2007) points out, owever, we use te actual values or output and inlation in te Taylor rule, it appears tat monetary policy was muc looser tan prescribed by te Taylor rule. 13 Tis result olds even i we look at te 4-quarter average o te orecasts instead o just te current quarter. Te volatility o te orecast errors is reduced wit te 4-quarter average, but te size o te errors increases and osets te volatility reduction. 7

10 RMSPFE in te more recent period. Te similarity o tese results wit te ull sample results, owever, suggests tat our indings or te ull-sample are not dependent upon a speciic time period. 14 Furtermore, te improvement is also present or te naïve orecasts, wic suggests no relative improvement on te part o te Fed. 6. Alternative Speciications In our analysis we ave cosen to ocus on te policy rule as initially ormulated in Taylor (1993). Subsequently, a number o autors ave estimated te policy rule tat was actually used by te Fed (e.g. Clarida et al, 2000; Orpanides, 2001). Two key issues are te size o te policy rule weigts and te role o interest rate smooting. 15 6A. Alternative Policy Rule Weigts One ocus o tis literature, and te part relevant to our analysis, is te weigts associated wit inlation and output in te Taylor rule. We perormed a simple calculation to examine te eect on te PFE o using weigts oter tan tose o te original Taylor Rule. Te PFE is a weigted sum o te individual orecast errors, consequently we ask: Given te individual orecast errors, wat weigts would minimize te impact o te orecast errors on te implied target interest rate? We tus determined te weigts tat minimize te RMSPFE o te Fed s orecasts, assuming tat te weigts summed to two. 16 We calculated dierent weigts or eac orizon and allowed or bot te case were te weigts are constant or te ull sample and allowing te weigts to dier between te pre-volcker and Volcker-Greenspan periods. In all 14 Tese subsamples are representative o several alternative subsamples tat we explored. 15 Anoter concern, beyond te scope o tis paper, is tat te optimal Taylor rule may not be linear (Nobay and Peel, 2003). Tere as been a substantial amount o researc in evaluating orecasts in te case o nonlinear loss unctions (e.g. Elliott et al., 2005 and 2008; Patton and Timmermann, 20071), including evidence tat te Fed s loss unction may not be linear, wit respect to inlation (Capistran, 2008) and wit respect to output growt (Patton and Timermann, 2007b) wen te orecasts are evaluated separately. 16 Almost identical results were ound wen we instead minimized te MAPFE. Te restriction tat te coeicients sum to two is consistent wit te literature were every estimate or prescriptive rule we ound suggests te coeicients sum to someting greater tan or equal to two. We cose te lower bound so tat we are truly minimizing te RMSPFE. 8

11 cases te weigts are very similar to te Taylor weigts. Tese weigts do not necessarily imply anyting about actual Fed beavior, but tey allow us to set a lower bound on te impact o orecast errors on te implied target interest rate. 17 Te results are very similar to wat we ound using te Taylor weigts. Based on tis analysis we still conclude tat, witout smooting, te impact o orecast errors alone on te implied target interest rate based on te Taylor rule is approximately 100 basis points. 6B. Interest Rate Smooting According to Giannoni and Woodord (2002), te application o certainty equivalence requires interest rate smooting. Tereore, suppose te policy rule is modiied to allow or interest rate smooting. Te policy orecast error in tis case would be a simple proportion o te policy orecast error witout smooting: A A ( 1 )( 1.5( π π ) + 0. ( y y ) T T PFE t = it it = 5 ρ, (3 ) Estimates o te smooting parameter, ρ, ave varied widely in te literature (e.g. Clarida et al, 1998, ind ρ = 0.8, wereas Rudebusc, 2001, argues tat te smooting parameter may be negligible in practice). For example, i te smooting parameter is 0.5, ten te impact o te orecast errors on te policy decision is reduced by al. Tus, te larger te orecast errors are expected to be, te greater te beneit o smooting Conclusions In tis paper we developed a simple metodology in order to evaluate te impact o orecast errors on te Fed s monetary policy as caracterized by a policy rule tat is assumed to take te orm o a Taylor rule. We ind tat te Fed s policy orecast error is smaller tan te 17 In particular, te speciic weigts ound by Clarida et al (2000) were in te set o potential weigts we considered. Teir weigts resulted in a iger MAPFE tan tose tat minimized te MAPFE. Te weigts, RMSPFEs, and MAPFEs are all available rom te autors upon request. 18 Tis assumes, owever, tat te past interest rate was set in an appropriate way. As noted by Orpanides and Williams (2007), interest rate smooting can also cause past errors to persist. 9

12 errors tat would ave resulted rom naïve orecasts or SPF predictions. Neverteless, te mean absolute policy orecast error o te Fed orecasts is approximately 100 basis points. Te results or selected subsamples were similar to tose or te entire period and a simple calculation sowed tat te weigts placed on te two variables tat minimize te PFE were similar to tose contained in te original Taylor Rule. Tus we ind te robust result tat, witout interest-rate smooting, orecast errors contribute a mean absolute error o approximately 100 basis points to target interest rates set by using a Taylor Rule. Reerences Bernanke, Ben S. (2010). Monetary Policy and te Housing Bubble speec at te Annual Meeting o te American Economic Association. Atlanta, Georgia, January 3, External Link ttp:// Capistran, C. (2008). Bias in Federal Reserve inlation orecasts: Is te Federal Reserve irrational or just cautious? Journal o Monetary Economics, 55, Clarida, R., Gali, J. and Gertler, M. (1998). Monetary Policy Rules in Practice: Some International Evidence. European Economic Review 42, Clarida, R., Gali, J. and Gertler, M. (2000). Monetary Policy Rules and Macroeconomic Stability: Evidence and Some Teory. Te Quarterly Journal o Economics 115(1), Clark, T. E. and Kozicki, S. (2005). Estimating Equilibrium Real Interest Rates in Real Time. Nort American Journal o Economics and Finance 16(3), Clements, M. P. (2004). Evaluating Te Bank O England Density Forecasts O Inlation. Te Economic Journal, 114 (October), Clements, M. P. and Hendry, D. F. (1993). On te Limitations o Comparing Mean Square Forecast Errors, and subsequent comments. Journal o Forecasting, 12(8), Clements, M. P., Joutz, F. and Stekler, H. O. (2007). An Evaluation o te Forecasts o te Federal Reserve: A Pooled Approac. Journal o Applied Econometrics 22(1), Crousore, D. (1993). Introducing: Te Survey o Proessional Forecasters. Federal Reserve Bank o Piladelpia Business Review, November/December, Eisenbeis, R., Waggoner, D. and Za, T. (2002). Evaluating Wall Street Journal Survey Forecasters: A Multivariate Approac. Business Economics, 37(3),

13 Elliott, G., Komunjer, I. and Timmermann, A. (2005). Biases in Macroeconomic Forecasts: Irrationality or Asymmetric Loss? Journal o te European Economic Association, 6(1), Elliott, G., Komunjer, I. and Timmermann, A. (2008). Estimation and Testing o Forecast Rationality under Flexible Loss. Review o Economic Studies, 72, Giannoni, M. P. and Woodord M. (2002). Optimal Interest Rate Rules: II. Applications. NBER Working Paper No Granger, C. W. J. and Pesaran, M. H. (2000a). A decision-based approac to orecast evaluation. In (W.S. Can, W. K. Li, and H. Tong, eds). Statistics and Finance: An Interace. London: Imperial College Press. Granger, C. W. J. and Pesaran, M. H. (2000b). Economic and statistical measures o orecast accuracy. Journal o Forecasting, 19, Hamilton, J. D., Pruitt, S. and Borger, S. C. (2009). Te Market-Perceived Monetary Policy Rule. Board o Governors o te Federal Reserve System International Finance Discussion Papers Number 982. Hymans, Saul H. (1968). Simultaneous Conidence Intervals in Econometric Forecasting. Econometrica, 36(1), Joutz, F. and Stekler, H.O. (2000). An Evaluation o te Predictions o te Federal Reserve. International Journal o Forecasting, 16, Kon, D. L. (2007). Jon Taylor Rules a speec at te Conerence on Jon Taylor's Contributions to Monetary Teory and Policy. Federal Reserve Bank o Dallas, Dallas, Texas, October 12, Komunjer, I. and Owyang M. T. (2007). Multivariate Forecast Evaluation and Rationality Testing. Federal Reserve Bank o St. Louis Working Paper No Lairi, K. and Seng, X. (2010). Learning and Heterogeneity in GDP and Inlation Forecasts. International Journal o Forecasting (special issue on Bayesian Forecasting in Economics), 26, Leig, D. (2008). Estimating te Federal Reserve's implicit inlation target: A state space approac. Journal o Economic Dynamics and Control, 32(6), Nobay, A. Robert and Peel, David A. (2003). Optimal Discretionary Monetary Policy in a Model o Asymmetric Central Bank Preerences. Te Economic Journal, 113 (489), Orpanides, A. (2001). Monetary Policy Rules Based on Real-Time Data. Te American Economic Review, 91(4),

14 Orpanides, A. (2003). Monetary Policy Evaluation wit Noisy Inormation. Journal o Monetary Economics, 50, Orpanides, A., and Weiland, V. (2008). Economic Projections and Rules o Tumb or Monetary Policy. Federal Reserve Bank o St. Louis Review, 90(4), Orpanides, A. and Williams, J. C. (2007). Robust Monetary Policy wit Imperect Knowledge. Journal o Monetary Economics, 54, Patton, Andrew J. and Timmermann, Allan (2007a). Properties o optimal orecasts under asymmetric loss and nonlinearity. Journal o Econometrics, 140, Patton, Andrew J. and Timmermann, Allan (2007b). Testing Forecast Optimality Under Unknown Loss. Journal o te American Statistical Association, 102 (480), Applications and Case Studies, Pesaran, M. H. and Skouras, S. (2002). Decision-based metods or orecast evaluation. In (M.P. Clements, and D.F. Hendry, eds) A Companion to Economic Forecasting, Oxord: Blackwell, pp Reiscneider, D. and Tulip, P. (2007). Gauging te Uncertainty o te Economic Outlook rom Historical Forecasting Errors. Federal Reserve Board Finance and Economics Discussion Series, Romer, C. D. and Romer, D. H. (2000). Federal Reserve Inormation and te Beavior o Interest Rates. American Economic Review, 90, Rudebusc, G. D. (2001). Term Structure Evidence on Interest Rate Smooting and Monetary Policy Inertia. Journal o Monetary Economics 49(6), Sims, C. A. (2002). Te Role o Models and Probabilities in te Monetary Policy Process. Brookings Papers on Economic Activity, 2, Sinclair, T. M., Stekler, H.O. and Kitzinger, L. (2010). Directional Forecasts o GDP and Inlation: A Joint Evaluation wit an Application to Federal Reserve Predictions. Applied Economics 42(18), Stekler, H.O. (1994). Are economic orecasts valuable? Journal o Forecasting 13, Taylor, J.B. (1993). Discretion versus Policy Rules in Practice. Carnegie-Rocester Conerence Series on Public Policy, 39, Taylor, J. B. (1999). Monetary Policy Rules. University o Cicago Press. Jon B. Taylor, (2007). Housing and monetary policy. Proceedings, Federal Reserve Bank o Kansas City,

15 Woodord, M. (2001a). Te Taylor Rule and Optimal Monetary Policy. American Economic Review, 91 (2), Woodord, M. (2001b). Inlation Stabilization and Welare. National Bureau o Economic Researc Working Paper Table 1: Mean Absolute Policy Forecast Error (MAPFE) and Root Mean Squared Policy Forecast Error (RMSPFE) Panel A: Full Sample ( ) 19 Horizon (monts) MAPFE Fed MAPFE naïve MAPFE SPF RMSPFE Fed RMSPFE naïve RMSPFE SPF N Panel B: Pre-Volcker Sample ( ) 19 Horizon (monts) MAPFE Fed MAPFE naïve MAPFE SPF RMSPFE Fed RMSPFE naïve RMSPFE SPF N Panel C: Volcker-Greenspan Sample ( ) Horizon (monts) MAPFE Fed MAPFE naïve MAPFE SPF RMSPFE Fed RMSPFE naïve RMSPFE SPF N SPF sample starts wit orecasts made in N = 98 or SPF. 21 N = 97 or SPF. 22 N = 42 or SPF. 23 N = 40 or SPF. 13