THE BRAZILIAN QUARTERLY REAL GDP:TEMPORAL DISAGGREGATION AND NOWCASTING.

Transcription

1 THE BRAZILIAN QUARTERLY REAL GDP:TEMPORAL DISAGGREGATION AND NOWCASTING. André Nunes Maranhão - University of Brasília and Bank of Brazil André Minella - Central Bank of Brazil Cleomar Gomes da Silva - Federal University of Uberlandia. FOURTH ANNUAL CONFERENCE OF THE INTERNATIONAL ASSOCIATION FOR APPLIED ECONOMETRICS, SAPPORO, JAPAN (IAAE-2017) 26 - June andrenmaranhao@gmail.com

2 Presentation 1 Introduction Monitoring of Economic Activity 2 Temporal Disaggregation Methods International Review Aggregation and Disaggregation Methods 3 Model Selection Criteria and Data Description Model Selection Criteria Data Description and Variable Selection 4 Results Results of Model Selection Nowcasting Exercises 5 Conclusion Conclusion and Research Agenda

3 Monitoring of Economic Activity Nowcasting Problem Economic variables and diferent frequencies: industrial production (PIM), retail survey (PMC) Monthly, 2 months lag (t +2); Average Electric Energy Consumption Monthly, 1 month lag (t +1); GDP frequency: Quarterly with Lag; IBC-Br: The Central Bank of Brazil Economic Activity Index; IBC-Br frequency: Monthly, 2 months lag (t +2); IBC-Br : Weaknesses and questions; Is there any room for advancement and improvement?

4 International Review Literature Article Boot, Feibes & Lisma (1965) Denton (1971) Chow & Lin (1971) Fernandez (1981) Gregoir (1995) Salazar et al. (1997, 1998) Santos, Silva & Cardoso (2001) Monch & Uhlig (2005) Cardoso (1981) Notini et alli. (2012) Theme Interpolation methods Interpolation methods Parametric model Unit root case Dynamic model Dynamic model Dynamic model State-Space (SS) model Fisrt brazilian case SS Coincident index for brazilian GDP Table : Time evolution of the disaggregation models

5 Aggregation and Disaggregation Methods Concepts and Definitions s y l,t = c i y h,i i=1 Y l = CY h Y h = AY l C t 3t =

6 Aggregation and Disaggregation Methods First Model Chow & Lin (1971): Y l = X l β + u l u l N(0, V l ) Y h = X h β + u h u h N(0, V h ) u h,t = ρu h,t 1 + ɛ t ɛ N(0, σ 2 ɛ ) Ŷ h = A(X l β + u l) Ŷh = AY l min β,ρ COV (Ŷh Y l )

7 Aggregation and Disaggregation Methods A Disaggregation Model using Kalman Filter Monch & Uhlig (2005): y + = (0 0 y y y 9...) y + l,t = i=0 y h,t i, t=3,6,9,... 0, otherwise ξ t = y h,t y h,t 1 y h,t 2 u h.t H t = = y + l,t = H t ξ t { [1/3 1/3 1/3 0], para, t = 3, 6, 9,... [ ], otherwise φ 0 0 ρ ρ y h,t 1 y h,t 2 y h,t 3 u h,t 1 + X h,t β ɛ t 0 0 ɛ t

8 Aggregation and Disaggregation Methods State-Space Models for Disaggregation

9 Model Selection Criteria Methods of Evaluation In-Sample Criteria Monch & Uhlig (2005): R 2 diff = Var( y + t T ) Var( y + t T ) + Var( u t T ) Out-of-Sample Criteria Proietti (2006): MAPE = 1 T RMSE = 1 T T y ˆ + t y + t t=1 y + t T ( y ˆ+ t y + t=1 t ) 2

10 Data Description and Variable Selection Variable Selection 1st quarter of 2003 to the 4rd quarter of 2012 (Real GDP); 54 variables were tested; Unit root tests: ADF, MADF gls, with and without structural breaks - Perron (1989), determinist trend with structural breaks; Cross correlation in first difference: Monch & Uhlig (2005), Notini & Issler (2008) and Notini et alli. (2012); Causality tests; Principal Component Analysis in time series context;

11 Data Description and Variable Selection Selected Variables

12 Results of Model Selection Models Selected

13 Results of Model Selection Predictive Ability

14 Nowcasting Exercises Model t+1

15 Nowcasting Exercises Model t+2 without IBC-Br

16 Nowcasting Exercises Model t+2

17 Conclusion and Research Agenda Final Considerations Main Contribuitions 1 Model (t+1) and GDP disaggregation; Same predictive capacity of the IBC-Br, however with variables t + 1; The quarterly average coincides with the values of real GDP, in this sense it is literally a coincident indicator. 2 Model (t+2) without IBC-Br and Model (t+2); Both models have a better predictive performance than the IBC-Br; The quarterly average coincides with the values of real GDP.

18 Conclusion and Research Agenda Final Considerations Research Agenda 1 Possibility of testing variables from the big data context; 2 Use of high-dimensional time series models; 3 Extended temporal sample for the purpose of using predictability tests as presented by Giacomini and White (2006); 4 Study of turning points and economic cycles.

19 Conclusion and Research Agenda Article Presented: THE BRAZILIAN QUARTERLY REAL GDP:TEMPORAL DISAGGREGATION AND NOWCASTING. André Nunes Maranhão. THANK YOU ALL.