News Shocks: Different Effects in Boom and Recession?

Transcription

1 News Shocks: Different Effects in Boom and Recession? Maria Bolboaca, Sarah Fischer University of Bern Study Center Gerzensee June 7, 5 /

2 Introduction News are defined in the literature as exogenous changes in the information sets that economic agents use to form their perceptions regarding future economic activity. The news-driven business cycle hypothesis assumes that business cycles can arise because of changes in expectations of future fundamentals. /

3 Research question Are the effects of news on future productivity state-dependent? We take an empirical approach to test whether the reactions to news shock are state-dependent and/or asymmetric. 3 /

4 In this paper we estimate a logistic smooth transition VAR model.... identify the news shock with a medium-run identification scheme.... use generalized impulse responses to analyze the effect of a news shock in different states of the economy.... use generalized forecast error variance decomposition to estimate the contribution of the news shock to the variance of the variables. 4 /

5 Related literature Short-run restrictions: Beaudry and Portier (AER, 6) Medium-run restrictions: Barsky and Sims (JME, ), Beaudry and Portier (JEL, 4) Non-fundamentalness: Sims (AE, ), Forni and Gambetti (JME, 4) 5 /

6 Contributions Main Contribution: Analysis of the state-dependent and asymmetric effects of news shocks. Technical Contributions: Estimation of different smooth transition processes - in the mean and the variance equation, respectively. Application of the maximum forecast error variance identification method in a nonlinear model. Comparison of short-run and medium-run identification methods in a nonlinear setting. 6 /

7 Main results The effect of the news shock is... qualitatively independent of the state of the economy. quantitatively different in expansions and recessions. not significantly asymmetric (sign/magnitude). affecting the probability of regime transition (e.g. escaping a recession). 7 /

8 Data In our model we include Total Factor Productivity: adjusted for factor utilization (Basu, Fernald and Kimball (6)) (logged) Measure of consumer confidence: Index of Consumer Sentiment from the Michigan Survey of Consumers Output: GDP nonfarm (logged, real, per capita) Inflation Rate: annualized log-difference of the GDP price deflator nonfarm Stock Prices: S&P 5 (logged, real, per capita) We estimate the model using quarterly data for the US for the sample period 955Q-Q4 with four lags. * 8 /

9 Logistic Smooth Transition VAR Y t = ( F M (s t ))Π E X t + F M (s t )Π R X t + ɛ t ɛ t N(, Σ t ) Σ t = Σ E ( F V (s t )) + Σ R F V (s t ) Transition Functions: F i (s t ) = exp( γ i(s t c i )) + exp( γ i (s t c i )), γ i >, for i = M, V Estimation results: ˆγ M = 3., ĉ M =.6; ˆγ V = 6.3, ĉ V =.5 9 /

10 Switching Variable NBER identified recessions NBER identified recessions.9 Probability of a recession given by the logistic function F.9 Probability of a recession given by the logistic function F * /

11 Generalized Impulse Responses The generalized impulse response function (GIRF) allows for the response to depend not only on the history Θ t but also on the magnitude and sign of the shock ξ it. GIRF (h, ξ it, Θ t ) = E {Y t+h ξ it = δ, Θ t } E {Y t+h ξ it =, Θ t } The responses are grouped according to the state indicated by F M (s t ), (s t Θ t ). REC: F M (s t ).5 and EXP: F M (s t ) <.5 The economy spends 5% of the time in recession. /

12 Medium-run Identification Scheme News shock: The shock with no impact effect on TFP, that has the largest contribution to the GFEVD of TFP in the medium-run (in years). Generalized Forecast Error Variance Decomposition: λ ij,θt (h) = h l= GIRF (l, ξ it, Θ t ) j K h i= l= GIRF (l, ξ it, Θ t ) j, i, j =,..., m /

13 Generalized Impulse Responses to a News Shock Total Factor Productivity 6 Index of Consumer Sentiment Output Stock Prices Inflation /

14 Robustness checks: GIRFs in expansion Total Factor Productivity 6 Index of Consumer Sentiment Output Stock Prices Inflation /

15 Robustness checks: GIRFs in recession Total Factor Productivity 6 Index of Consumer Sentiment Output Stock Prices Inflation * 5 /

16 Generalized Forecast Error Variance Decomposition Expansion News Shock Recession h= h=4 h=8 h=6 h=4 h= h=4 h=8 h=6 h=4 TFP TFP ICS ICS Output Output Inflation Inflation SP SP /

17 Switching Probability: Expansion 5 positive small news shock 5 positive big news shock negative small news shock 3 negative big news shock /

18 Switching Probability: Recession positive small news shock positive big news shock negative small news shock 5 negative big news shock /

19 Conclusion The news shock leads to business cycle movements, independent of the state of the economy. Stronger effects of news shocks in expansion than in recession. No evidence in favour of asymmetries. The probability of regime transition is strongly influenced by the news shock. 9 /

20 /

21 Appendix: Linear Appendix: Nonlinear Technical Appendix Linear model Y t = Π + p Π jy t j + ɛ t = Π X t + ɛ t, ɛ t N(, Σ) j= where Y t = (Y,t,...Y m,t ) is an m vector of endogenous variables, Π is an m intercept vector, Π j is a m m parameter matrix. X t = (, Y t,..., Y t p ). MA representation: Y t = B(L)ɛ t Structural MA representation: Y t = C(L)u t /

22 Appendix: Linear Appendix: Nonlinear Technical Appendix Identification Identification I Linear mapping between innovations and structural shocks: ɛ t = Au t AA = Σ TS: The only shock that affects TFP on impact. NS: The shock with no impact effect on TFP, that has an impact effect on consumer confidence. Identification II TS: The only shock that affects TFP on impact. NS: The shock with no impact effect on TFP, that has the largest effect on TFP in the medium-run (in years). /

23 Appendix: Linear Appendix: Nonlinear Technical Appendix Identification II h-step ahead forecast error: Y t+h Y t+h t = Σ h τ= B τ ÃDu t+h τ, DD = I Contribution of shock j to the variance of variable i at horizon h: Ξ i,j (h) = e i (Σh τ= B τ ÃDe je j D Ã B τ )e i e i (Σh τ= B τ ÃÃ = (Σh τ= B i,τ Ãθθ Ã B i,τ ) B τ )e i (Σ h τ= B i,τ ÃÃ B i,τ ) Choose θ such that Ξ,NS (h) is maximized at horizon h = 4 quarters. 3 /

24 Appendix: Linear Appendix: Nonlinear Technical Appendix Impulse Responses to a News Shock.6 Total Factor Productivity 6 Index of Consumer Sentiment Output Stock Prices Inflation /

25 Appendix: Linear Appendix: Nonlinear Technical Appendix Variance Decomposition Table: Share of forecast error variance attributable to the news shock obtained with two different identification schemes Short-run identification scheme h= h=4 h=8 h=6 h=4 TFP ICS Output Inflation SP Medium-run identification scheme h= h=4 h=8 h=6 h=4 TFP ICS Output Inflation SP /

26 Appendix: Linear Appendix: Nonlinear Technical Appendix News shocks in the linear model 4 3 news shock (indentification II) news shock (identification I) Figure: News shock 6 /

27 Appendix: Linear Appendix: Nonlinear Technical Appendix News shocks in a linear 7-variable model.5 Total Factor Productivity 5 Index of Consumer Sentiment Output Inflation 3 4 Stock Prices Consumption Hours worked Figure: Impulse responses to a news shock. 7 /

28 Appendix: Linear Appendix: Nonlinear Technical Appendix Stability check Figure: Counterfactuals * 8 /

29 Appendix: Linear Appendix: Nonlinear Technical Appendix Generalized Impulse Responses to a News Shock (on ICS) Total Factor Productivity 6 Index of Consumer Sentiment Output Stock Prices Inflation /

30 Appendix: Linear Appendix: Nonlinear Technical Appendix GIRF: Recessions vs Linear - MR Identification Scheme Total Factor Productivity 6 Index of Consumer Sentiment Output Stock Prices Inflation * 3 /

31 Appendix: Linear Appendix: Nonlinear Technical Appendix GIRF: Expansions vs Linear - MR Identification Scheme Total Factor Productivity 6 Index of Consumer Sentiment Output Stock Prices Inflation /

32 Appendix: Linear Appendix: Nonlinear Technical Appendix GIRF: Positive small vs negative small news - MR Identification Scheme. TFP 5 ICS Output Inflation SP TFP 3.5 ICS Output Inflation 7 SP Top: Expansion, black = positive news, red=negative news Bottom: Recession, bue = positive news, purple=negative news 3 /

33 Appendix: Linear Appendix: Nonlinear Technical Appendix GIRF: Positive small vs positive big news - MR Identification Scheme. TFP 5 ICS Output Inflation SP TFP 3.5 ICS Output Inflation 7 SP Top: Expansion, black = small news, red=big news Bottom: Recession, bue = small news, purple=big news 33 /

34 Appendix: Linear Appendix: Nonlinear Technical Appendix Generalized Forecast Error Variance Decomposition - SR Identification Scheme Expansion News Shock Recession h= h=4 h=8 h=6 h=4 h= h=4 h=8 h=6 h=4 TFP TFP ICS ICS Output Output Inflation Inflation SP SP /

35 Appendix: Linear Appendix: Nonlinear Technical Appendix Orthogonality test * We take a large dataset Q t, which contains 87 quarterly macroeconomic series for the U.S. from 955Q to Q4. We set the maximum number of factors p = and compute the first p principal components of Q t. We use the principal components to obtain the unobserved factors. 3 We test whether the estimated shock is orthogonal to the past of the principal components, p (we use lags, 4, and 6), by regressing the critical structural shock (news shock) on the past of the principal components and performing an F-test of the null hypothesis that the coefficients are jointly zero. 35 /

36 Appendix: Linear Appendix: Nonlinear Technical Appendix Linearity test of Teräsvirta and Yang (4) H : Π = Π, H : Π,j Π,j, for at least one j {,..., p}. We approximate the logistic function by a third order Taylor expansion. We then perform an LM test: Estimate the model under the null hypothesis (the linear model). Compute the matrix residual sum of squares, SSR = Ẽ Ẽ. Estimate the auxiliary regression, by regressing Y (or Ẽ) on X and the interaction terms. Compute SSR = Ê Ê. 3 Compute the asymptotic χ test statistic: LM χ = T (m tr { SSR SSR } ) We reject the null hypothesis of linearity at all significance levels. 36 /

37 Appendix: Linear Appendix: Nonlinear Technical Appendix Estimation The parameters of the LSTVAR model are estimated using NLS. The error terms are normally distributed, thus the NLS estimator is equivalent to the maximum likelihood estimator of the parameters Ψ = {γ V, c V, γ M, c M, Σ E, Σ R, Π E, Π R }: ˆΨ = arg min Ψ T t= ɛ tσ t ɛ t For given γ V, c V, γ M, c M, Σ E, and Σ R, estimates of Π can be obtained by weighted least squares (WLS), with weights given by Σ t. The procedure iterates on {γ F, c F, γ M, c M, Σ E, Σ R }, yielding Π and the likelihood, until an optimum is reached. We perform the estimation using a MCMC method - the MH algorithm. 37 /

38 Appendix: Linear Appendix: Nonlinear Technical Appendix Constancy of the error covariance matrix We use the test of Yang (4). First, we estimate the model under the null hypothesis assuming the error covariance matrix to be constant over time. Similar to the linearity test for the dynamic parameters, the alternative hypothesis is approximated by a third-order Taylor approximation given the transition variable. The LM statistic is then computed as follows: LM = p i= T SSG i RSS i SSG i, where SSG i is the sum of squared g it, and the RSS i the corresponding residual sum of squares in the auxiliary regression. 38 /