Real Time Data for Norway: Challenges for Monetary Policy Tom Bernhardsen, Øyvind Eitrheim, Anne Sofie Jore and Øistein Røisland Norges Bank Prepared for the workshop on ”Model evaluation in macroeconomics” University of Oslo, 6-7. May 2005 ØEi/7. May 2005 1 Motivation • Real time data: Challenges for monetary policy – Monetary policy desicions are taken in real time. – Macroeconomic data are plagued with inaccuracy, untimeliness and are subject to frequent revisions. – How should monetary policy take this into account? – Are revisions persistent? – Are revisions predictable? – Can monetary policy be robustified? ØEi/7. May 2005 2 Disposition • The real-time database for Norway (quarterly vintages 1993:1-2002:4). • Output gap estimation in real time. • Monetary policy in real time. • Optimal simple rules under output gap uncertainty. ØEi/7. May 2005 3 • Three main reasons for data revisions: • Changes due to incomplete information (ordinary revision cycle) • Base-year changes • Major data revisions ØEi/7. May 2005 4 • Data revisions in Norway 1993:1-2002:4 • Two major revisions of National Accounts data in less than 10 years • GDP revised upwards • Particularly strong upward revisions in the second half of the 1990’s ØEi/7. May 2005 5 Table 1: Annual growth rate of Mainland Norway GDP over five-year periods Vintage Vintage 1995q1 Vintage 2002q1 Vintage 2003q4 Period 1974 to 1979 3,6 3,9 3,8 2,0 2,3 2,2 1979 to 1984 1,5 1,8 1,9 1984 to 1989 1,7 2,3 2,3 1989 to 1994 3,1 4,1 1994 to 1999 Table 2: Annual growth rate of Mainland Norway GDP over five-year periods Period 1991 to 1996 1996 to 2001 ØEi/7. May 2005 Vintage 2002q1 Vintage 2003q4 3,1 2,3 3,5 3,2 6 • Data revisions: news or noise? (Mankiw and Shapiro, 1986) • Vintage data for mainland GDP 1993:1 to 2002:4 (labeled e.g., a, b, c, . . .) • Description (tables, graphs) log(Yta/Ytb) − log(Yta/Ytb) • Are revisions characterized by news or noise? • Let ytf = ytp + εt. Under the news view ytp ⊥ εt while under the noise view εt ⊥ ytf • Are revisions to GDP predictable? ØEi/7. May 2005 7 Final and Real-Time output growth over 4 quarters .10 Accumulated revisions over 4, 8 and 12 quarters .06 12 quarters .08 .04 .06 Final .04 .02 .02 .00 .00 4 quarters -.02 -.02 Realtime -.04 8 quarters -.04 93 94 95 96 97 98 99 00 01 02 03 D4YFINAL D4YREALTIME (a) Final and Real time output growth 93 94 95 96 97 98 99 00 01 02 03 D4YFL_RT D8YFL_RT D12YFL_RT (b) Accumulated revisions Figure 1: Final and Real time output growth, accumulated revisions over 4, 8 and 12 quarters 1993.1 – 2002.1 ØEi/7. May 2005 8 Log output ratio benchmark 1-benchmark 2 0,08 0,08 0,06 0,06 0,04 0,04 0,02 0,02 0 Benchmark 1: Vintage 1995q1 Benchmark 2: Vintage 2002q1 Benchmark 3: Vintage 2003q4 0 -0,02 -0,02 -0,04 -0,04 -0,06 -0,06 -0,08 1970 -0,08 1975 1980 1985 1990 1995 2000 benchmark 1-benchmark 3 benchmark 2-benchmark 3 0,08 0,08 0,08 0,08 0,06 0,06 0,06 0,06 0,04 0,04 0,04 0,04 0,02 0,02 0,02 0,02 0 0 0 0 -0,02 -0,02 -0,02 -0,02 -0,04 -0,04 -0,04 -0,04 -0,06 -0,06 -0,06 -0,06 -0,08 1970 -0,08 -0,08 1970 1975 1980 1985 1990 1995 2000 -0,08 1975 1980 1985 1990 1995 2000 Figure 2: Log output ratios for three different vintages of Real Time data ØEi/7. May 2005 9 • Are revisions to GDP predictable? • Real time growth data may deviate considerably from final growth data. • For Norwegian data the standard deviation of the regression is around 1 percentage point, though for some numbers the deviation is substantially higher • For Norwegian data the standard deviation of the regression is around 1 percentage point, though for some numbers the deviation is substantially higher ØEi/7. May 2005 10 • Evolution of growth data in the subsequent quarters after first time publication: • No substantial revisions after one quarter (the standard deviation unchanged) • In the second and the third quarter after first time publication the standard deviation is notably less, around 0.70 percentage point. • After a year measurement errors are considerably lower, though some errors remain ØEi/7. May 2005 11 Real time data: 2. observation versus final data 1992.Q4 – 2002.Q2 8 8 6 6 Final data Final data Real time data: 1. observation versus final data 1992.Q4 – 2002.Q2 4 Y=1.05x + 0.22 R2=0.69 S=1.11 2 4 Y=0.91x + 0.58 R2=0.71 S=1.08 2 0 0 0 2 4 6 8 Real time data (a) 1. observation versus final data 0 2 4 6 8 Real time data (b) 2. observation versus final data Figure 3: Real-time observations versus final data 1992.4–2002.2 ØEi/7. May 2005 12 Real time data: 4. observation versus final data 1992.Q4 – 2002.Q2 8 8 6 6 Final data Final data Real time data: 3. observation versus final data 1992.Q4 – 2002.Q2 4 Y=1.09x + 0.13 R2=0.88 S=0.70 2 4 Y=1.14x + 0.18 R2=0.89 S=0.69 2 0 0 0 2 4 6 8 Real time data (a) 3. observation versus final data 0 2 4 6 8 Real time data (b) 4. observation versus final data Figure 4: Real-time observations versus final data 1992.4–2002.2 ØEi/7. May 2005 13 • Are revisions to GDP predictable? (cont’d) • Test of H0 : α = β = 0 in ∆ytf − ∆ytr = α + β∆ytr + εt • Some information may exist, not embedded in the real time data, which would help us predicting the final outcome. • Test for additional information Ztp • Tests for Real time macroeconomic information (labour market variables, goods market variables, financial indicators) indicate that final growth data cannot be predicted beyond the information contained in the numbers published in real time. ØEi/7. May 2005 14 Table 3: Omitted variable tests (OVT) for additional effects on revisions from macroeconomic variables ØEi/7. May 2005 Labour market variables New Jobs FOV T (3 ,30)=0.1872[0.8303] Vacancies FOV T (3 ,30)=0.2616[0.7716] Employment and vacancies FOV T (3 ,30)=0.1814[0.8350] Unemployment FOV T (3 ,30)=0.2298[0.7961] Change in unemployment FOV T (3 ,30)=1.5212[0.2354] Hours worked FOV T (3 ,30)=0.2616[0.7716] Goods market variables Industrial production FOV T (3,30)=0.1144[0.8923] D(Industrial production) FOV T (3 ,30)=0.3211[0.7279] Retail sales FOV T (3 ,30)=0.069[0.9338] D(Retail sales) FOV T (3 ,30)=0.2422[0.7864] New orders FOV T (3 ,30)=0.0671[0.9352] D(New orders) FOV T (3 ,30)=0.3681[0.6952] Industrial investment FOV T (3 ,30)=0.2616[0.7716] D(Industrial investment) FOV T (3 ,30)=0.4538[0.6397] Bankruptcies FOV T (3 ,30)=0.3716[0.6928] 15 Table 4: Omitted variable tests (OVT) for additional effects on revisions from macroeconomic variables Financial market Credit growth, C1 D(Credit growth, C1) Credit growth, C2 D(Credit growth, C2) Credit growth, C3 D(Credit growth, C3) Nominal effective exchange rate D(Nominal effective exchange rate) Slope of the yield curve ØEi/7. May 2005 variables FOV T (3 ,30)=0.0700[0.9324] FOV T (3 ,30)=0.6087[0.5509] FOV T (3 ,30)=0.0698[0.9327] FOV T (3 ,30)=0.7627[0.4755] FOV T (3 ,30)=0.0682[0.9342] FOV T (3 ,30)=1.1621[0.3269] FOV T (3 ,30)=0.3213[0.7277] FOV T (3 ,30)=0.3036[0.7405] FOV T (3 ,30)=0.8261[0.1922] 16 • Economic developments in Norway • No authoritative determination of business cycles in Norway • The last decades characterized by the deep recession in the late 1980’s followed by a long expansion in the 1990’s • The period covered by the Real-time data-base starts close to the beginning of an expansion that peaks in the late 1990’s and ends with a trough in 2003 ØEi/7. May 2005 17 Detrending output in practice We assume that we can decompose the log of output, yt, in a trend component, µt, and a cyclical component, zt: yt = µt + zt The cyclical component, zt, may be used as a measure of the output gap, ygapt = yt − µt. There is considerable uncertainty with respect to the measurement of potential output, yt∗, and in this paper we will use estimates of the trend, µt, as our estimate of the potential output. We follow Orphanides and van Norden (1999) and consider a fairly wide range of univariate models of the output gap. In the table below we ØEi/7. May 2005 18 present seven univariate models of the output gap ranging from simple deterministic trend models, through filtering models (Hodrick-Prescott), frequency domain models (band pass) to univariate unobserved components (UC) models. Orphanides and van Norden (2001) have also considered bivariate unobserved components models which are estimated with Kalman filter algorithms.1 In addition to the seven univariate models, we estimate output gaps using a production function model. We follow the approach in Nymoen and Frøyland (2000), basing the calculations on a production function for the sectors manufacturing, construction, services and distributive trade, accounting for about 34 of production for mainland Norway.2 The 1 The estimation results for the UC-models in Orphanides and van Norden (2001) are based on Kalman filter algorithms in the TSM-module in GAUSS. We are grateful to Simon van Norden for providing access to his procedures written in RATS and Gauss for estimating the different models in Table 5. 2 In the univariate methods, we use GDP for mainland Norway. One reason for only taking account of selected sectors in the Production Function model, is an assumption that the production function model is ØEi/7. May 2005 19 aggregated production function is assumed to be of a Cobb-Douglas type with constant returns to scale. The elasticities are given by the income factor shares of the two production factors. The weights can according to the Ministry of Finance be estimated at 23 for person-hours and 13 for real capital for mainland enterprises. The equilibrium unemployment rate, trend factor productivity and potential person-hours-employment are constructed by using a Hodrick-Prescott filter on the actual series. less applicable to production in the public sector and agriculture. ØEi/7. May 2005 20 Table 5: Output gap models QT UC HP PF Trend/cycle decomposition Quadratic trend Unobserved component (local trend model) Harvey(1985),Clark(1987) Hodrick-Prescott (λ = 1600) Production Function yt = µt + zt µt = α + βt + γt2 + εt µt = δt + µt−1 + ηt δt = δt−1 + νt zt = ρ1zt−1 + ρ2zt−2 + εt T µt = argmin t=1 {(yt − µt )2 + λ[∆2µt+1 ]} µt = α̂ + 23 lt∗ + 13 kt∗ + tf p∗t , The Production Function based estimates of potential output relies on the parameters of the Cobb-Douglas function, but also on data for the long run trends in employment, lt∗, real capital, kt∗, and total factor productivity, tf p∗t . The equilibrium unemployment rate, trend factor productivity and potential person-hours-employment are constructed by using a Hodrick-Prescott filter on the actual series. ØEi/7. May 2005 21 Quadratic trend (QT) Unobserved component (UC) 10 10 Quasi real-time gap CL_FLGAP CL_RTGAP CL_QRGAP Real-time gap 5 Final gap 5 Quasi real-time gap 0 0 Real-time gap Final gap -5 -5 QT_FLGAP QT_RTGAP QT_QRGAP -10 -10 93 94 95 96 97 98 99 00 01 93 94 95 96 (a) QT 97 98 99 00 01 (b) UC Hodrick Prescott (HP) Production function (PF) 10 10 PF_FLGAP PF_RTGAP PF_QRGAP 5 5 Quasi real-time gap Final gap Quasi real-time gap 0 0 Real-time gap Real-time gap -5 Final gap -5 HP1600_FLGAP HP1600_RTGAP HP1600_QRGAP -10 -10 93 94 95 96 97 (c) HP 98 99 00 01 93 94 95 96 97 98 99 00 01 (d) PF Figure 5: Final gaps, real-time gaps and quasi real-time gaps 1993.1–2002.1 ØEi/7. May 2005 22 Final Output Gap Upper and Lower bounds 10 10 8 8 6 6 4 4 2 2 0 0 -2 -2 -4 -4 -6 -6 -8 -8 -10 1971 -10 1976 1981 1986 1991 1996 2001 Figure 6: Final output gaps (upper and lower bounds 1971.1–2003.3) An example of thick modeling (Granger and Jeon, 2004) ØEi/7. May 2005 23 Final Output Gap Real-time Output Gap Upper and Lower bounds Upper and Lower bounds 10 10 10 10 8 8 8 8 6 6 6 6 4 4 4 4 2 2 2 2 0 0 0 0 -2 -2 -2 -2 -4 -4 -4 -4 -6 -6 -6 -6 -8 -8 -8 -8 -10 1993 -10 1994 1995 1996 1997 1998 (a) Final 1999 2000 2001 -10 1993 -10 1994 1995 1996 1997 1998 1999 2000 2001 (b) Real time Figure 7: Final and Real time output gaps 1993.1–2002.1 ØEi/7. May 2005 24 Some apparent features include: - Output gaps measured by the alternative models generally move in the same direction, both as Real Time and Final gaps. - Measured in Real Time, most of the gaps are positive for most of the period but turn negative the last few years. Calculated on Final data, on the other hand, the gaps are negative in the first part of the period and positive in the last part. - The size of the output gaps covers a wide range, particularly measured as Real Time gaps. - Measured as Final output gaps, the difference between most of the models is markedly reduced after the first years of the period. ØEi/7. May 2005 25 • Total Revisions are generally large and persistent • High degree of serial correlation. Coefficients ranging from 0.28 for PF model to 0.93 for QT model • In five of the eight models, the absolute value of the mean of total revisions is larger than the absolute value of the mean output gap ØEi/7. May 2005 26 • Other revisions is the the main contributor to Total revisions for all models except the LT model • Data revisions are smaller and less serially correlated • Standard deviations of Data revisions are smaller than standard deviations of Other revisions ØEi/7. May 2005 27 Total Revisions 8 8 6 6 4 4 PF 2 2 0 0 -2 -2 -4 HP -4 UC -6 -6 -8 -8 QT -10 -12 1993 -10 -12 1994 1995 1996 1997 1998 1999 2000 2001 (a) Total revisions Figure 8: Total revisions 1993.2–2002.1 ØEi/7. May 2005 28 Total Revisions Upper and Lower bounds 8 8 6 6 4 4 2 2 0 0 -2 -2 -4 -4 -6 -6 -8 -8 -10 -10 -12 1993 -12 1994 1995 1996 1997 1998 1999 2000 2001 (a) Total revisions Figure 9: Total revisions 1993.1–2001.4 ØEi/7. May 2005 29 Data Revisions Other Revisions 8 8 8 8 6 6 6 6 4 4 4 2 2 2 0 0 0 0 -2 -2 -2 -4 -4 -6 -6 -6 -6 -8 -8 -8 -8 -10 -10 -10 -10 -12 1993 1994 1995 1996 1997 1998 1999 2000 2001 -12 -12 1993 1994 1995 1996 1997 1998 1999 2000 2001 -12 4 -2 -4 QT PF UC HP (a) Data revisions PF 2 UC HP QT -4 (b) Other revisions Figure 10: Data revisions and Other revisions 1993.1–2002.1 ØEi/7. May 2005 30 Data Revisions Other Revisions Upper and Lower bounds Upper and Lower bounds 8 8 8 8 6 6 6 6 4 4 4 4 2 2 2 2 0 0 0 0 -2 -2 -2 -2 -4 -4 -4 -4 -6 -6 -6 -6 -8 -8 -8 -8 -10 -10 -10 -10 -12 1993 -12 -12 1993 -12 1994 1995 1996 1997 1998 1999 (a) Data revisions 2000 2001 1994 1995 1996 1997 1998 1999 2000 2001 (b) Other revisions Figure 11: Data revisions and Other revisions 1993.1–2002.1 ØEi/7. May 2005 31 Quadratic trend (QT) Unobserved component (UC) 10 10 5 5 Data revision 0 0 Total revision Other revisions -5 -10 Data revision Other revisions -5 Total revision -10 93 94 95 96 97 98 99 00 01 93 94 95 (a) QT 96 97 98 99 00 01 00 01 (b) UC Hodrick Prescott (HP) Production function (PF) 10 10 5 Data revision 5 Data revision 0 0 Other revisions Other revisions Total revision -5 Total revision -5 -10 -10 93 94 95 96 97 (c) HP 98 99 00 01 93 94 95 96 97 98 99 (d) PF Figure 12: Total revisions, data revisions and other revisions 1993.1–2002.1 ØEi/7. May 2005 32 Table 6: Summary Statistics: Output gaps and Total Revisions 1993q1 2002q1 METHOD MEAN SD MIN MAX CORR AR Quadratic Trend RTGAP 4.23 2.00 -0.67 7.30 0.33 QRGAP 5.22 2.26 0.22 11.10 0.64 FLGAP -0.22 3.70 -7.43 4.97 1.00 Revisions -4.39 3.57 -10.35 1.51 0.93 UC model (Clark) RTGAP -0.25 0.43 -1.06 1.02 0.22 QRGAP -0.17 0.72 -1.41 2.81 0.30 FLGAP 0.28 3.14 -5.98 4.95 1.00 Revisions 0.58 3.10 -5.28 4.91 0.82 ØEi/7. May 2005 33 Table 7: Summary Statistics: Output gaps and Total Revisions 1993q1 2002q1 METHOD MEAN SD MIN MAX CORR AR HP1600 RTGAP 0.16 1.56 -2.25 3.09 -0.01 QRGAP 0.41 1.70 -2.35 4.83 0.39 FLGAP 0.20 1.39 -3.59 4.23 1.00 Revisions 0.02 2.13 -4.92 3.06 0.73 Production Function RTGAP -0.43 1.96 -5.69 2.68 0.87 QRGAP 0.06 2.20 -5.10 5.05 0.95 FLGAP -1.08 2.83 -7.21 3.12 1.00 Revisions -0.65 1.51 -4.34 2.55 0.28 ØEi/7. May 2005 34 • Output gap estimates • Correlations between Real time and Final estimates of the output gap range from 0 for the HP1600 model to 0.87 for the PF model. • The ratio of the standard deviation of the Total revision to the standard deviation of the Final estimate of the output gap (proxy for the Noice-to-Signal ratio) is close to, or higher than 1 for all models except the PF model, which exhibits a ratio of 0.53. ØEi/7. May 2005 35 • Output gap estimates (cont’d) • For most models, Real time and Final estimates produce opposite signs in the output gap with a frequency of around 0.50. The PF model is an exception, with a frequency of only 0.06. • The absolute value of the revision exceeds the absolute value of the final gap with a frequency of 0.50 - 0.70 for most of the models. Again, the PF model is an exception with a frequency of 0.17. ØEi/7. May 2005 36 Table 8: Summary statistics for Output METHOD CORR Quadratic trend (QT) 0.33 0.22 Unobserved component (UC) -0.01 Hodrick-Prescott (HP) 0.87 Production Function (PF) ØEi/7. May 2005 gaps: 1993q1 - 2002q1 NS OPSIGN XSIZE 0.97 0.44 0.64 0.99 0.53 0.53 1.53 0.53 0.75 0.53 0.06 0.17 37 • Output gap estimates (cont’d) • Real time data produce in general unreliable estimates of the output gap for Norwegian data. • Using comparable models, the results are in line with findings for US, Australian and Canadian data. • Results from the Production Function model are somewhat more encouraging than results from univariate models. • Further work, using more sophisticated models, will be useful. ØEi/7. May 2005 38 • The models fall into three categories: • Inflexible trends (QT) • Moderately flexible trends (UC) • Strongly flexible trends (HP, PF) ØEi/7. May 2005 39 Growth rates of potential output 6 6 5 5 HP 4 3 4 UC 3 QT 2 2 PF 1 0 1971 1 0 1976 1981 1986 1991 1996 2001 Figure 13: Growth rates of potential output in the QT, UC, HP and PF models. ØEi/7. May 2005 40 HP-weights (1600) Centered HP-weights (1600) .24 .24 .20 t t .20 .16 .16 .12 t-4 .12 .08 .04 .00 -.04 1970 1975 1980 t t-4 t-8 1985 1990 t-12 t-16 t-20 1995 t-8 .08 t-12 t-16 .04 t-4 t-8 .00 2000 t-12 -.04 -60 t-24 t-28 Two-sided (a) HP-weights (1600) -50 -40 -30 -20 -10 0 20 30 (b) Centered HP-weights (1600) HP-weights (50000) Centered HP-weights (50000) .10 t .08 t-4 .10 t .08 t-4 .06 t-8 .06 .04 t-12 .02 t-16 t-20 t-24 t-28 .00 -.02 1970 10 t-8 t-12 .04 t-16 t-20 .02 .00 1975 1980 t t-4 t-8 ØEi/7. May 2005 1985 t-12 t-16 t-20 1990 1995 t-24 t-28 Two-sided (c) HP-weights (50000) 2000 -.02 -60 -50 -40 -30 -20 -10 0 10 (d) Centered HP-weights (50000) 20 30 41 Table 9: Revision Statistics: Details 1993q1 METHOD MEAN SD MIN MAX Quadratic Trend (QT) Total Revision -4.39 3.57 -10.35 1.51 0.99 1.14 -1.87 3.79 Data Revision -5.38 2.87 -8.58 -0.33 Other Revision Unobserved component (UC) Total Revision 0.58 3.10 -5.28 4.91 0.08 0.55 -1.49 1.79 Data Revision 0.50 3.03 -5.45 4.75 Other Revision Hodrick-Prescott (HP) Total Revision 0.02 2.13 -4.92 3.06 0.25 0.91 -2.37 3.25 Data Revision -0.23 1.75 -2.77 2.21 Other Revision Production Function (PF) Total Revision -0.65 1.51 -4.34 2.55 0.49 1.20 -1.50 4.53 Data ØEi/7. May 2005 Revision -1.14 1.00 -2.84 0.70 Other Revision 2002q1 AR N/S 0.93 0.19 1.03 1.53 0.41 1.65 0.82 -0.43 0.89 1.01 0.18 0.98 0.73 -0.02 0.97 1.53 0.68 1.09 0.28 0.09 0.97 0.58 0.46 0.54 42 Table 10: Summary statistics for Total revisions: 1993q1 - 2002q1 METHOD Quadratic Trend (QT) Unobserved component (UC) Hodrick-Prescott (HP) Production Function (PF) ØEi/7. May 2005 MEAN -4.39 0.58 0.02 -0.65 SD 3.57 3.10 2.13 1.51 AR 0.93 0.82 0.73 0.28 N/S(FL) 1.53 1.01 1.53 0.58 N/S(RT) 2.83 7.34 1.37 0.84 MEAN(DTrev) 0.99 0.08 0.25 0.49 43 • Optimal reaction under output gap uncertainty – We use a calibrated version of a fairly standard aggregated New Keynesian macromodel for illustration – The observed output gap in real time, yto, enters the interest rate rule. – The measurement error process εot is assumed to follow an AR(1) process. – Output gap estimates based on final data and real-time data are used to estimate ρ. – We optimize coefficients in the interest rate rule on the basis of a grid search using the simple loss function L = πt2 + λyt2 . ØEi/7. May 2005 44 The model πt = 0.8πt−1 + 0.2Etπt+1 + 0.2yt−1 + 0.1zt + επt (1) yt = 0.85yt−1 + 0.1Etyt+1 − 0.1(it−1 − Et−1πt) + 0.05zt−1 + εyt (2) zt = 0.4zt−1 + 0.6Etzt+1 − 0.2{(it − Etπt+1) − (ift − Etπtf )} + εzt(3) it = αiit−1 + απ πt + αy yto + α∆y ∆yto (4) yto = yt + εot (5) εot = 0.7εot−1 + ηto, ØEi/7. May 2005 σ̂ηto = 1.3 (6) 45 Equation (1) is a “hybrid” open-economy New Keynesian Phillips curve, where πt is the rate of (CPI) inflation, yt is the “true” (but unobservable) output gap, zt is the (log of) the real exchange rate, measured as deviation from the equilibrium real exchange rate, and επt is a cost-push shock. Equation (2) represents aggregate demand, where it is the nominal short-term interest rate. it − Etπt+1 is then the real interest rate, and the neutral real interest rate is for simplicity normalised to zero. Equation (3) is the UIP condition, where a lag is introduced to better capture observed dynamics in the real exchange rate (short-run deviations from UIP). ØEi/7. May 2005 46 Unobserved increase in the true output gap (productivity slowdown). Number of periods on the horizontal axis. 0.8 1 it 0.6 yt 0.8 0.6 0.4 0.4 0.2 0.2 yto 0 0 -0.2 0 4 8 12 16 20 24 28 32 36 40-0.2 0 0.4 4 8 12 16 0.2 πt 0.3 20 24 28 32 36 40 zt -0 0.2 -0.2 0.1 -0.4 0 -0.6 0 -0.1 0 4 ØEi/7. May 2005 8 12 16 20 24 28 32 36 40 4 8 12 16 20 24 28 32 36 40 -0.8 Figure 15: 47 Table 11: Optimal coefficients in the interest rate rule for varying degree of persistence in output gap mismeasurement (benchmark is no uncertainty). εot = ρεot−1 + ηto, period loss function Lt = πt2 + λyt2 λ 0 0.5 1.0 1.5 2.0 ØEi/7. May 2005 ρ 0.70 0.28 0.70 0.28 0.70 0.28 0.70 0.28 0.70 0.28 - απ 4.6 4.6 9.9 3.5 3.7 9.9 2.7 2.8 7.9 2.3 2.5 7.1 2.1 2.3 6.5 αy 0.4 0.6 3.0 0.8 1.2 6.8 0.7 1.2 7.6 0.6 1.2 8.2 0.6 0.7 8.6 α∆y 0.0 0.0 2.0 0.9 0.6 3.2 0.9 0.6 2.6 0.9 0.6 2.2 1.0 0.6 2.0 S.E.(π ) S.E.(y ) E[L] 0.71 0.70 0.61 0.81 0.79 0.75 0.91 0.90 0.86 0.98 0.96 0.92 1.03 1.01 0.98 1.53 1.44 1.28 1.25 1.15 0.90 1.15 1.03 0.75 1.10 0.98 0.69 1.07 0.96 0.64 0.50 0.49 0.37 1.44 1.28 0.96 2.15 1.87 1.30 2.78 2.37 1.56 3.37 3.38 1.77 48 • Consequences for optimal monetary policy in the model exercise – Output gap mismeasurements give rise to excessive variability in output and prices. – Numerically, output gap mismeasurements reduce the optimal coefficients considerably compared the full information case. – In particular, the weight on the output gap should be smaller. – Output gap mismeasurements tend to be persistent. A case for responding to the change in the gap. (Orphanides et al, 2000). ØEi/7. May 2005 49 • Main welfare costs in the model exercise – Monetary policy does not respond quickly enough to changes in the true output gap. – Inflation moves too far away from the target and closes the output gap too slowly. – Caveat: The results on the optimal monetary policy response may be model dependent and a robustness check is called for. ØEi/7. May 2005 50 • Wrapping up: – Data revisions may change our perception of economic growth both in the near but also more distant past. – Output gap revisions are often large. Estimates are subject to data uncertainty, parameter uncertainty and (more generally) model uncertainty, and may turn out as unreliable indicators of the current state of the economy. – Real time output gap uncertainty may seriously misguide interest rate setting. – Output gap uncertainty seems to cause a considerable change in the optimal monetary policy response to macroeconomic disturbances (coefficient reduction). ØEi/7. May 2005 51 References *References Granger, C. W. J. and Y. Jeon (2004). Thick modeling. Economic Modelling , 21 , 323–343. Mankiw, N. G. and M. D. Shapiro (1986). News or noise: An analysis of GNP revisions. Survey of Current Business, (May), 20–25. Nymoen, R. and E. Frøyland (2000). Output gap in the Norwegian economy - different methodologies, same result? Economic Bulletin 2000/2 , 71 , 46–52. Orphanides, A. and S. van Norden (1999). The reliability of output gap ØEi/7. May 2005 52 estimates in real time. Upublished Paper, Board of Governors of the Federal Reserve System and Ecole des Hautes Etudes Commerciales, Montreal. Orphanides, A. and S. van Norden (2001). The reliability of inflation forecasts based on output gap estimates in real time. Upublished Paper, Board of Governors of the Federal Reserve System and Ecole des Hautes Etudes Commerciales, Montreal and CIRANO. ØEi/7. May 2005 53