Banque nationale de Belgique SA · EURONEXT:BNB
Banque nationale de Belgique SA — Technical Presentation (March 2005)
40 slides from Banque nationale de Belgique SA's technical presentation, Mar 24, 2005
Slide 1: On the fit and forecasting performance of New-Keynesian models
On the fit and forecasting performance of New-Keynesian models Marco Del Negro (FRBAtlanta), Frank Schorfheide (UPenn), Frank Smets (ECB) and Raf Wouters (NBB) Joint Seminar Series NBB-KUL-UCL-ULB 24 March 2005
Slide 2: Motivation
Motivation • DSGE models are not only attractive from a theoretical perspective, but are also emerging as useful tools for forecasting and quantitative analysis; • Smets and Wouters (SW, 2003, 2005) find that estimated DSGE models a la Christiano et al (JPE, 2005) can compete with Bayesian VARs in t
Slide 3: Motivation (1)
Motivation (1) • Sims (2003) criticizes the SW model comparison exercise (based on Bayes factors and posterior odds) on the basis of the argument that the models considered are too sparse. • In such cases, posterior odds may lead to extreme outcomes and may also be highly dependent on the prior dist
Slide 4: Motivation (1)
Motivation (1) • In this paper, the hyper parameter ( ) is interpreted as a factor that scales the inverse of a prior covariance matrix of parameters that capture deviations from the DSGE model restrictions. implies that the prior on the misspecification parameters concentrates the mass around zero;
Slide 5: Motivation (2)
Motivation (2) • An analysis of the estimated stochastic processes in the work of Smets and Wouters indicates the presence of some misspecification: – These processes are often estimated to be very persistent and sometime exhibit a clear trend (in spite of being assumed stationary); – This is likely
Slide 6: Benefits of the methodology
Benefits of the methodology • The hyperparameter that has the highest posterior probability ( ^) can be interpreted as providing an assessment of the degree of misspecification. • The DSGE-VAR( ^) provides a natural benchmark for comparing the empirical fit of the DSGE model (Schorfheide, 2001): – O
Slide 7: Main findings
Main findings • The stationary state-space representation of the DSGE model is well approximated by a VAR with 4 lags, provided the error-correction terms are included; • The posterior distribution of the hyperparameter has an inverse U-shape, indicating that the fit of the autoregressive system can
Slide 8: Main findings
Main findings • Regarding the nature of misspecification of the DSGE model, there are three main findings: – There is a shift towards less stickiness in prices and wages, less persistence in the shocks and less habit formation, when misspecification is allowed for; – Nevertheless, many the impulse r
Slide 9: Outline of presentation
Outline of presentation • Brief overview of the DSGE model: – variant of CEE (JPE, 2005) and SW (2003) • Brief overview of the methodology: – similar to Del Negro and Schorfheide (2003), but different interpretation. • Presentation of main findings; • Conclusions and future work
Slide 10: The New Keynesian DSGE model
The New Keynesian DSGE model • Model is a variant of Altig et al (2002), Christiano et al (JPE, 2005) and SW (2003); • Households consume, supply labor, enjoy money, set wages, accumulate capital subject to adjustment costs, choose capital utilisation and have access to a full set of state-contingen
Slide 11: DSGE Model
DSGE Model Households’ Problem Households maximize: Pt+s 14, Mi+3(9) te ae NOG Cete(5) — hCep-2()) — Pat Legal)" + A A (Ft subject to: Pop sCrps(i) + Pepoleps(9) + Bets(a) + Mets(9) < Reps Beps_1 09) + Mips—109) + Arts (9) + Migs + Wie4s(j)Le4s(9) + (RE, mete (J) Ke¢ 1) — Pr s0(te45(7)) Ki4s-1(3)),
Slide 12: Households' Problem — Sticky Wages
Households’ Problem — Sticky Wages 1 1+A, Labor packers “pack” L;(j) labor into L; = [le L(j) ail , and demand: ibe Wilj)\ Li(j) = (A) Li, 1 Nw W,= | Wi( iytail . oO Households that can reset wages (probability 1 —¢,,) solve: maxy,) Es Do (4w8)*bets [- Fler ye] s.t. budget constraint for s =0,...,00
Slide 13: The New Keynesian DSGE model
The New Keynesian DSGE model • Intermediate goods producers use capital services and labor; set prices in a monopolistic competitive market and face a random walk productivity process; • Final goods producers aggregate a continuum of intermediate goods.
Slide 14: Producers' Problem
Producers’ Problem max P.(4)¥i(i) — WL: (i) — REK; (A), subject to the technology: Y¥,(i) = max {K:(i)°(Z:Li(i))* — ZF, O}, where = log(Z;/Z;-1) follows the process: (2% —Y) = pz (“1 —y) + eee.
Slide 15: Producers' Problem — Sticky Prices
Producers’ Problem — Sticky Prices Final goods producers “pack" ¥;(z) into a composite good: 1 41 1+Ay ,= / Om ; (In Ape —In As) = py, (In Ape1 — In Af) + OD Fer 4: 0 and demand: where Producers that can reset prices (probability 1 — ¢,) solve: MAX py Ee Lo GOH. (PO (Marta) — MCrs) Yer) s.t. demand
Slide 16: Government
Government PGt + Rei Bei + M1 = Ti + Me + B where G; = (1 — 1/9) and: IN ge = (1 — pg) ING + pg IN ge_a + Og€ gt: The central bank follows a nominal interest rate rule: pR wa 7 \ Ye 1px Re Re-1 Tt Yt eps zs =| _ vd oRE Re Re Ts y,;* where Y," can be either the stochastic steady state level of output
Slide 17: Detrending and model solution
Detrending and model solution _R Ww, _M ~_B ~_W, = Fh w= Eh m= Bh = Oe = Wh, & ==.Zf, & = =tZ,, Log-linearize + Sims’ gensys — transition equations.
Slide 18: Measurement equations
Measurement equations Output growth (log differences, quarter-to-quarter, in %): 100 x ( InYi—InY%i1) =100x(%—-%1+% )+100y Consumption growth (...): 100 x ( Inc;-—InC,_4) =100x(&-G@1+%)+100y. Investment growth (...): 100 ( In}K—Inh-1) =100x(%—u%1+% )+100y. Real wage growth (...):100 x (InW — InP, —
Slide 19: The DSGE-VAR methodology
The DSGE-VAR methodology • VAR/VECM representation of the DSGE model • Prior distribution of the VAR parameters given the DSGE parameters • Posterior distribution of VAR parameters and DSGE parameters • Identification
Slide 20: VAR/VECM representation
VAR/VECM representation e We approximate DSGE model by VAR Ay = Pot PiAy it... + O,Ayptu, Elwui) == (1) e or alternatively by a VECM Au = Po + O6(8'y-1) + Pr Ayr +... + PpAyep + ut. (2) e Y Is Txn matrix with rows Ay}; X is Txk matrix with rows a, = [1, Ay! _,,..., Ay, U is T x n matrix with rows uj
Slide 21: VAR/VECM representation
VAR/VECM representation e Log-lin DSGE model has In general no exact finite-order VAR representation, but can often be well approximated with few lags, provided y Is covariance stationary. e Define autocovariance matrices Cxx(@) = EP [x], Cxy(0) = EP [ew] E?[-] expectation under DSGE model. e Let *(
Slide 22: Prior distribution of VAR parameters
Prior distribution of VAR parameters
Slide 23: Posterior distribution of VAR parameters
Posterior distribution of VAR parameters
Slide 24: Implementation
Implementation e Place prior over @. e MCMC methods to generate draws from the joint posterior of 6, 6, and =. e Use marginal data density to form a posterior for A: p(Y) = [ 00, E,®)p.(0,E, 9)(0,E, 0). (10)
Slide 25: Identification
Identification To assess the DSGE model based on impulse response comparisons between DSGE and VAR/VECM we need a mapping from VAR Innovations into structural shocks. An (exactly) Identified VAR Is a triplet: (&,£,&), where Q Is orthonormal: (*) =F, Oe) var The DSGE model Is Identified: there is a m
Slide 26: Findings
Findings Can DSGE-VAR reproduce the dynamics of the DSGE model when AX — co? — Dealing with Cointegration e Y, C, I, and W all grow along a balanced growth path — according to the DSGE model. e A standard VAR in growth rates (even for \ — oo) does not impose these coin- tegrating restrictions. e If
Slide 27: How Well Do DSGE-VAR/VECM Approximate the DSGE Model?
How Well Do DSGE-VAR/VECM Approximate the DSGE Model? Y Cc | H Ww inflation R y 1i-—— 4 D5 1 A p.5 g b.5|/ Lo in Vaan a ol ee 0 0 5.5 0 A bs5 0 5 0 1 4 1 J 4 FZ >. Nee N_1 0 Los aN mM _ps an 1 -5 1 0 0 0 ™ D.5 20 1 a 5 05 p.5 < =~. 4 J eee a. | Noe oh o a 0 —_ 0 Een 4 aie D5 —=20 -1 0) los —=5.5 2 0
Slide 28: Preliminary: Direct evaluation of the DSGE model
Preliminary: Direct evaluation of the DSGE model • Choosing the DSGE specification; • In-sample fit of the DSGE model; • Parameters of the DSGE model; • Exogenous processes.
Slide 29: Choosing the DSGE Model Specification
Choosing the DSGE Model Specification (i) Dynamic Indexation (tp = w = 1) (ii) Flexible-price Output Target (iii) F sets steady-state profits = 0 Log-Differences in Bayes Factors Marginal Likelihood wrt Baseline Model wrt Baseline Model (in %) -31.49 10-** % -2.08 12.49 % -28.87 10-13 %
Slide 30: In-Sample Fit of the DSGE Model
In-Sample Fit of the DSGE Model Y Cc | 4 2 10 2 is) ie} 0 -2 -10 -2 —4 —4 -—20 1974.2980.259861992.751999 1974.2980.269861992.751999 1974.2980.259861992.751999 H Ww inflation 4.2 3 4 41 2 3 4 1 2 3.9 0 1 3.8 —1 ie] 4 1974.2980.259861992.751999 1974.2980.269861992.751999 1974.2980.259861992.751999 R
Slide 31: DSGE Model's Parameter Estimates
DSGE Model's Parameter Estimates Parameter Prior MeanPrior Stdd Post Mean 90% Lower Band 90% Upper Band a 0.250 0.100 0.172 0.149 0.195 Cp 0.750 0.100 0.848 0.814 0.883 8 4.000 1.500 5.827 3.238 8.115 h 0.800 0.100 0.793 0.725 0.857 a! 0.200 0.075 0.167 0.067 0.273 Y 2.000 0.750 2.204 1.050 3.271 Cw
Slide 32: Exogenous Processes
Exogenous Processes Tech 1980.25 T 1986.5 ® 1992.75 T T 1980.25 T 1986.5 1992.75 T T n 1980.25 T n 1986.5 1992.75 7 1980.25 T 1986.5 g 1992.75 T 1980.25 1986.5 1992.75
Slide 33: Evaluation of the DSGE model using the DSGE-VAR toolkit
Evaluation of the DSGE model using the DSGE-VAR toolkit • In-sample fit: how large a weight should we put on the DSGE model restrictions? • Out-of-sample forecasting performance (rolling sample 1985QI-2000Q4); • Comparing parameter estimates; • Comparing impulse responses;
Slide 34: In-sample fit: How large a weight should we put on the DSGE model prior?
In-sample fit: How large a weight should we put on the DSGE model prior? Weight of the DSGE-VECM DSGE-VAR DSGE model prior (1) (2) (3) (4) (5) (6) \ Artificial Log-Marginal ( Difference yLog-Marginal ( Difference ) ‘Total Likelihood wrt Best Model Likelihood wrt Best Model 0.5 0.33 -456.41 ( -13.03
Slide 35: Out-of-Sample Root Mean Squared Errors: VECM Specification
Out-of-Sample Root Mean Squared Errors: VECM Specification forecast horizon Vananits ane) Taney ID QO QHLOLA BAS SHaVrocnn Lacon a toe a pee wes LO 5 0 te O°. AODNNTOH -DTOND NY og PTONAD “gO GOD MYA QOT COWL tT NIL OSD TIOBLT ANA HASSLE SOM SVAN Os AON tad MMi COg AN GOTH AHOHNO Herat Nev CHM MT HI
Slide 36: Table 2: DSGE Model's Parameter Estimates (Part I)
Table 2: DSGE Model’s Parameter Estimates (Part I) Prior DSGE Post. DSGE-VECM(\) Post. Distr. Mean Stdd Mean Interval Mean Interval a B 0.250 0.100 0.172 0.149 0.195 0.153 0.125 0.181 & B 0.750 0.100 0.848 0.814 0.883 0.667 0.512 0.834 s! N 4.000 1.500 5.827 3.238 8.115 4.204 3.999 4.405 h B 0.800 0
Slide 37: Table 3: DSGE Model's Parameter Estimates (Part II)
Table 3: DSGE Model’s Parameter Estimates (Part II) Prior DSGE Post. DSGE-VECM(A) Post. Distr. Mean Stdd Mean Interval Mean Interval Pz B 0.200 0.050 0.218 0.146 0.294 0.210 0.134 0.283 B 0.800 0.050 0.705 0.625 0.791 0.856 0.766 0.951 B 0.800 0.050 0.518 0.449 0.589 0.690 0.471 0.895 Pu B 0.800 0.0
Slide 38: Comparing the Propagation of Shocks
2 Ss @ 1 0 0 se -1 -2 1 = 0 -1 1 2 0 -1 0.5 a is) -0.5 1 = 0 -1 1 > a 5 0 = Comparing the Propagation of Shocks Y - cumulativeC — cumulative | — cumulative H W - cumulative _ inflation R 2 5 1 1 —p.2 1 sere] 4 ee) OP obese ssps| 4 of erere4 Obey aan a 7 . ie) -5 -1 ie) 2 -1 (0) 5 0 3 1 5 1 Sa SY =e
Slide 39: Conclusions
Conclusions • Application of the DSGE-VAR tool kit to a relatively rich New Keynesian DSGE model suggests that some model features are still at odds with the data: – Relaxing the DSGE model restrictions improves the fit in terms of Bayesian marginal likelihood and out-of-sample forecasting performan
Slide 40: Work ahead
Work ahead • Work harder on building models that can be successfully taken to non-detrended data to fulfil Kydland and Prescott (1982)’s original promise of integrating growth and business cycle theory; • Need to develop approaches that use the DSGE model restrictions, but down-weight those frequenc
Banque Nationale De Belgique — Technical Presentation (March 2005) on 6ix