Nominal GDP now-casting
Michele Modugno, ECARES
Lucrezia Reichlin, London Business School and CEPR Frontiers of Macroeconometrics
Bank of England, UCL and cemmap workshop 25 and 26 April 2013
Motivation
• Since 2008 great recession, prolonged weakness of real economic activity
• Central Banks have been increasingly focusing on real economic activity ...
• Different proposals: dual target, forward guidance, nominal GDP targeting generally within flexible inflation targeting framework
Some problems:
• Real time monitoring of nominal GDP: nominal GDP published late and subject to large revisions
• Not clear exploitable Phillips curve trade-off in the data
Real time monitoring of economic variables
• In central banks (and markets) there is a long tradition of monitoring real GDP in real time
• Since Giannone et al (2008) it has been shown by a large literature that this can be done via formal models in an effective way →→ nowcasting:
contraction of the terms Now and Forecasting Meteorology Now-casting
forecasting up to 6-12 hours ahead long tradition, since 1860
Economic Now-casting
forecasting the near future, the present and even recent past
• Most empirical work is on real GDP
• What is the case for now-casting nominal GDP?
The idea of now-casting: learning from markets
Market participants can be viewed as now-casters
• they routinely monitor all macroeconomic data to get a view on current and future fundamentals and their effects on
policy
• The relevant information on the state of the economy is
conveyed to markets through the release of macroeconomic reports
• Market expectation for the headlines of these reports are collected up to the day before the actual release of the
indicator and distributed by data providers (i.e. Bloomberg)
The Real-Time Data-Flow: Last Month...
The Real-Time Data-Flow: Last Month...
The Real-Time Data-Flow: Last Month...
The Real-Time Data-Flow: Last Month...
The Real-Time Data-Flow: This Week...
The Real-Time Data-Flow: This Month...
The Real-Time Data-Flow: This Month...
The Real-Time Data-Flow: This Month...
The Real-Time Data-Flow: This Month...
Now-casting real GDP
Motivation:
• It arrives late and it is subject to large revisions
• The present is the only horizon of predictability - unpredictability beyond current quarter (cite)
• Accuracy of macroeconomic projections heavily rely on starting conditions
Empirical results (see Banbura et al 2012, Handbook of Forecasting):
1. Only real variables matter
2. Timeliness matters: in particular survey at the beginning of the quarter
3. Information matters
4. Daily model with daily financial variables does not outperform monthly real model
Now-casting - other key variables?
• Motivation less clear for CPI inflation: available monthly, less revised
although Modugno (2012) shows in a daily model of CPI inflation that high frequency information on oil price helps providing an accurate early signal on current inflation
• Clear motivation for nominal GDP: available quarterly, delay of publication, large revisions
However no now-casting studies or in general predictability studies are available for nominal GDP
Conjecture: daily model with financial variables, possibly monetary aggregates, may do well
Plan of the presentation
1. Introduce the idea of now-casting 2. Introduce the model
3. Apply to nominal GDP
-- data: daily and weekly and oil prices, daily financial variables, monthly price and real variables, quarterly deflator and real GDP -- evolution of the precision in relation to the flow of data
publications throughout the quarter
-- effect of unanticipated component of data releases (news) on the now-cast
-- Do for nominal GDP, real GDP and the deflator
-- Analyze two models: daily model and monthly model
4. Analyze relation between nominal and real variables through the analysis of the news and through a conditional forecast exercise
Mimicking Market behavior and Beyond
(a) Construct a joint model for all macroeconomic data releases (b) Update the model in real time, in accordance with the real-time data flow
• Model based forecasts: free of judgement, mood, heading.
• Translate the news in a common unit What s the impact of the news on GDP?
The Real-Time Data-Flow: Yesterday in Europe
A model of Now-Casting
• ytQ: GDP at time t.
• Ωv: vintage of data (quarterly, monthly, possibly daily) available at time v (date of a particular data release)
Nowcasting of ytQ: orthogonal projection of ytQ on the available information:
E h
ytQ|Ωvi ,
The information set Ωv has particular characteristics:
1 it has a “ragged” or “jagged edge” [publication lags differing across series]
2 it contains mixed frequency series, in our case monthly and quarterly
3 it could be large
A model of Now-Casting
• ytQ: GDP at time t.
• Ωv: vintage of data (quarterly, monthly, possibly daily) available at time v (date of a particular data release)
Nowcasting of ytQ: orthogonal projection of ytQ on the available information:
E h
ytQ|Ωvi ,
The information set Ωv has particular characteristics:
1 it has a “ragged” or “jagged edge” [publication lags differing across series]
2 it contains mixed frequency series, in our case monthly and quarterly
3 it could be large
A model of Now-Casting
• ytQ: GDP at time t.
• Ωv: vintage of data (quarterly, monthly, possibly daily) available at time v (date of a particular data release)
Nowcasting of ytQ: orthogonal projection of ytQ on the available information:
E h
ytQ|Ωvi ,
The information set Ωv has particular characteristics:
1 it has a “ragged” or “jagged edge” [publication lags differing across series]
2 it contains mixed frequency series, in our case monthly and quarterly
3 it could be large
Further features
• Projections need to be updated regularly E
h
ytQ|Ωvi , Eh
ytQ|Ωv +1i , ...
v , v + 1, ..., consecutive data releases
Typically the intervals between two consecutive data releases are short (possible couple of days or less) and change over time.
Consequently, v has high frequency and it is irregularly spaced.
News and nowcast revisions
• New release ⇒ the information set expands (new releases): Ωv ⊆ Ωv +1[we are abstracting from data revisions]
• Decompose new forecast in two orthogonal components: E
h
ytQ|Ωv +1i
| {z }
new forecast
= Eh
ytQ|Ωvi
| {z }
old forecast
+ Eh
ytQ|Iv +1i
| {z }
revision
,
Iv +1information in Ωv +1“orthogonal” to Ωv
• If we have a model that can account for joint dynamics of all variables, we can express the forecast revision as a weighted sum of news from the released variables:
E h
ytQ|Ωv +1i
− Eh ytQ|Ωvi
| {z }
forecast revision
= X
j∈Jv +1
bj,t,v +1 xj,Tj,v +1− Exj,Tj,v +1|Ωv
| {z }
news
.
For detailed derivation see Banubra and Modugno, 2008.
News and nowcast revisions
• New release ⇒ the information set expands (new releases): Ωv ⊆ Ωv +1[we are abstracting from data revisions]
• Decompose new forecast in two orthogonal components:
E h
ytQ|Ωv +1i
| {z }
new forecast
= Eh
ytQ|Ωvi
| {z }
old forecast
+ Eh
ytQ|Iv +1i
| {z }
revision
,
Iv +1information in Ωv +1“orthogonal” to Ωv
• If we have a model that can account for joint dynamics of all variables, we can express the forecast revision as a weighted sum of news from the released variables:
E h
ytQ|Ωv +1i
− Eh ytQ|Ωvi
| {z }
forecast revision
= X
j∈Jv +1
bj,t,v +1 xj,Tj,v +1− Exj,Tj,v +1|Ωv
| {z }
news
.
For detailed derivation see Banubra and Modugno, 2008.
News and nowcast revisions
• New release ⇒ the information set expands (new releases): Ωv ⊆ Ωv +1[we are abstracting from data revisions]
• Decompose new forecast in two orthogonal components:
E h
ytQ|Ωv +1i
| {z }
new forecast
= Eh
ytQ|Ωvi
| {z }
old forecast
+ Eh
ytQ|Iv +1i
| {z }
revision
,
Iv +1information in Ωv +1“orthogonal” to Ωv
• If we have a model that can account for joint dynamics of all variables, we can express the forecast revision as a weighted sum of news from the released variables:
E h
ytQ|Ωv +1i
− Eh ytQ|Ωvi
| {z }
forecast revision
= X
j∈Jv +1
bj,t,v +1 xj,Tj,v +1− Exj,Tj,v +1|Ωv
| {z }
news
.
For detailed derivation see Banubra and Modugno, 2008.
What kind of framework?
Three desiderata:
1 can capture joint dynamics of inputs and target
2 can be estimated on many series while retaining parsimony
3 can handle jagged edged data and mix frequency
Idea: use parsimonious model that can be cast in state space form and use Kalman filter to project and to handle jagged edged data
Evans 2005 IJCB; Giannone, Reichlin and Small, 2008 JME
Computing projections. What kind of model?
The dynamic factor model
xt = µ + Λft+ εt,
• ft: (unobserved) common factors; εt: idiosyncratic components
• Λfactor loadings
• Factors are modelled as a VAR process:
ft = A1ft−1+ · · · +Apft−p+ut
Parsimonious and robust model for large macroeconomic datasets.
- Forni et al. 2000 REStat, ...
- Stock and Watson, 2002 JASA, JBES ...
- Bernanke and Boivin, 2002 JME;
- Bai, 2003 Econometrica;
- Giannone et al. 2005 Macroeconomic Annual See keynote Lecture by James Stock tomorrow morning
Alternative: Bayesian Vector Autoregression
Problems and solutions
• Missing data
Kalman filter and smoother can be used to obtain, in an efficient and automatic manner, the projection for any pattern of data availability in Ωv as well as the news Iv +1and expectations needed to compute bj,t,v +1
• Mixed frequency
Consider lower frequency variables as being periodically missing
• Estimation: Quasi Maximum likelihood:
- robust and feasible Doz, Giannone and Reichlin., 2008 REStat
- handling missing data Banbura and Modugno, 2010
State space representation with mixed frequencies
Example: Let YtQdenote the vector of (log of) the quarterly flow series.
We assume that YtQ is the sum of daily contributions Xt
YtQ=
t
X
s=t−k +1
Xs, t = k , 2k , . . . .
Hence we will have that the stationary series ytQ =YtQ− Yt−kQ can be written as:
ytQ=k
t
X
s=t−k +1
t + 1 − s k xs+
t−k
X
s=t−2∗(k −1)
s − t + 2 ∗ k − 1
k xs
, t = k , 2k , . . . ,
where xs=Xs− Xs−1can be thought of as an unobserved daily growth rate (or difference).
See also Modugno 2011: Nowcasting Inflation
Data Table
Some details
• Consider also monthly model with higher frequency variables aggregated to monthly and made available at the beginning of the month
• Results based on pseudo-real time
• For daily model the forecast is performed every day, for the monthly model every week
• Nominal GDP derived from aggregating real GDP and the deflator appropriately
Daily now-cast of GDP, evolution of RMSFE, news
• INSERT CHARTS
Now-cast
Evolution of RMSFE over the quarter daily r=1, monthly with r=1 and r=2
0.30 0.35 0.40 0.45 0.50 0.55 0.60 0.65 0.70 0.75
Q-1 M1 D7 Q-1 M1 D14 Q-1 M1 D21 Q-1 M1 D28 Q-1 M2 D7 Q-1 M2 D14 Q-1 M2 D21 Q-1 M2 D28 Q-1 M3 D7 Q-1 M3 D14 Q-1 M3 D21 Q-1 M3 D28 Q0 M1 D7 Q0 M1 D14 Q0 M1 D21 Q0 M1 D28 Q0 M2 D7 Q0 M2 D14 Q0 M2 D21 Q0 M2 D28 Q0 M3 D7 Q0 M3 D14 Q0 M3 D21 Q0 M3 D28 Q1 M1 D7 Q1 M1 D14 Q1 M1 D21
r1p1 r2p1 daily RW
news
Comments
• The model does well
• Daily model outperforms monthly for nominal GDP in forecast and early now-cast : timely oil prices and financial news
matter then
• These news affect nominal GDP because they affect the deflator (to check)
• Some of the features of real GDP forecasting are confirmed -- information matters
-- real variables news have sizeable effect -- timeliness matters
Monthly model:
Now-cast nominal GDP
-2.5 -2 -1.5 -1 -0.5 0 0.5 1 1.5 2 2.5 3
r1p1 r2p1 actual
Monthly model:
Now-cast deflator
-0.4 -0.2 0 0.2 0.4 0.6 0.8 1 1.2 1.4
r1p1 r2p1 actual
Monthly model:
Now-cast real GDP
-3 -2.5 -2 -1.5 -1 -0.5 0 0.5 1 1.5 2 2.5
r1p1 r2p1 actual
Nominal GDP
0.30 0.35 0.40 0.45 0.50 0.55 0.60 0.65 0.70 0.75
Q-1 M1 D7 Q-1 M1 D14 Q-1 M1 D21 Q-1 M1 D28 Q-1 M2 D7 Q-1 M2 D14 Q-1 M2 D21 Q-1 M2 D28 Q-1 M3 D7 Q-1 M3 D14 Q-1 M3 D21 Q-1 M3 D28 Q0 M1 D7 Q0 M1 D14 Q0 M1 D21 Q0 M1 D28 Q0 M2 D7 Q0 M2 D14 Q0 M2 D21 Q0 M2 D28 Q0 M3 D7 Q0 M3 D14 Q0 M3 D21 Q0 M3 D28 Q1 M1 D7 Q1 M1 D14 Q1 M1 D21
r1p1 r2p1 RW
Observations
• The monotone behavior of the RMSFE is less clear in the monthly model since the high frequency variables are
aggregated and artificially made available at the beginning of the month
• From the end of the first month of the current quarter (wen the deflator and real GDP from the previous quarter arrive) incremental news reduce uncertainty and the now-cast model clearly outperform the constant growth model
Deflator
0.15 0.17 0.19 0.21 0.23 0.25 0.27 0.29
Q-1 M1 D7 Q-1 M1 D14 Q-1 M1 D21 Q-1 M1 D28 Q-1 M2 D7 Q-1 M2 D14 Q-1 M2 D21 Q-1 M2 D28 Q-1 M3 D7 Q-1 M3 D14 Q-1 M3 D21 Q-1 M3 D28 Q0 M1 D7 Q0 M1 D14 Q0 M1 D21 Q0 M1 D28 Q0 M2 D7 Q0 M2 D14 Q0 M2 D21 Q0 M2 D28 Q0 M3 D7 Q0 M3 D14 Q0 M3 D21 Q0 M3 D28 Q1 M1 D7 Q1 M1 D14 Q1 M1 D21
r1p1 r2p1 RW
Observations
• Not much seems to matter except past real GDP and deflator
• Te second factor helps
Real GDP
0.3 0.35 0.4 0.45 0.5 0.55 0.6 0.65 0.7 0.75 0.8
Q-1 M1 D7 Q-1 M1 D14 Q-1 M1 D21 Q-1 M1 D28 Q-1 M2 D7 Q-1 M2 D14 Q-1 M2 D21 Q-1 M2 D28 Q-1 M3 D7 Q-1 M3 D14 Q-1 M3 D21 Q-1 M3 D28 Q0 M1 D7 Q0 M1 D14 Q0 M1 D21 Q0 M1 D28 Q0 M2 D7 Q0 M2 D14 Q0 M2 D21 Q0 M2 D28 Q0 M3 D7 Q0 M3 D14 Q0 M3 D21 Q0 M3 D28 Q1 M1 D7 Q1 M1 D14 Q1 M1 D21
r1p1 r2p1 RW
Observations
• Monotonic pattern confirmed (see Giannone et al 2008 and Banbura et al 2012)
• Important role of the surveys
Comments
• Inflation - more predictable in an absolute sense since it is less volatile. However harder to beat the random walk benchmark. This is possible only at the end of this first month of the current quarter when the deflator is released
• The release of the deflator has also a major impact on nominal GDP but in the daily model the pattern is declining from previous quarter give the availability of daily financial variables
• Strong case for now-casting nominal GDP – can beat random walk by exploiting timely releases in back-cast, now-cast and forecast
Moreover:
• Two factors fit better the deflator (it captures the price dimension )
• For nominal and real variable the difference between one and two factors is small
Conjecture:
Weak relation between nominal and real side
However both nominal and real side matters for the deflator
In-sample fit at different frequencies: deflator
In-sample fit at different frequencies: nominal GDP
In-sample fit at different frequencies: real GDP
Comments
• Second factor small but relevant for the deflator.
• Fit is good for nominal and real GDP up to a year cycle – very poor for deflator – as known in the literature difficult to
capture the low frequency component of inflation (see Lenza et al for the euro area)
What are the news that matter?
• We extract the news as described and we present average result for first, second and third month
• Present results for nominal GDP, the deflator and real GDP
Average impact nominal GDP r=2
-0.20 -0.15 -0.10 -0.05 0.00 0.05 0.10 0.15 0.20 0.25 0.30
IPTOT PMI DPRI URATE PYRL_TOT PCETOT HSTARTS HSOLD ORD_MFGD PPI_FG CPITOT EXPORT IMP_CUST PHBOS_GA SALES_RETAIL_NOM CONF_CFB S&P GS10 GS3M MCOILWTICO TWEXMMTH M1SL M2SL GDP
m3 m2 m1
Average impact deflator r=2
-0.04 -0.02 0.00 0.02 0.04 0.06 0.08 0.10 0.12
IPTOT PMI DPRI URATE PYRL_TOT PCETOT HSTARTS HSOLD ORD_MFGD PPI_FG CPITOT EXPORT IMP_CUST PHBOS_GA ES_RETAIL_NOM CONF_CFB S&P GS10 GS3M MCOILWTICO TWEXMMTH M1SL M2SL Real GDP Deflator
m3 m2 m1
Average impact real GDP r=2
-0.20 -0.15 -0.10 -0.05 0.00 0.05 0.10 0.15 0.20 0.25
IPTOT PMI DPRI URATE PYRL_TOT PCETOT HSTARTS HSOLD ORD_MFGD PPI_FG CPITOT EXPORT IMP_CUST PHBOS_GA SALES_RETAIL_NOM CONF_CFB S&P GS10 GS3M MCOILWTICO TWEXMMTH M1SL M2SL Real GDP Deflator
m3 m2 m1
What is it that matters?
Summary
• Nominal GDP:
real variables, interest rates and exchange rate – no monetary aggregates, no prices
• Deflator:
Very small impact of the news; prices, interest rates and exchange rate – CPI, PPI and oil inflation - no real variables, no monetary aggregates / this for r=2 otherwise constant model is best
• Real GDP:
same as nominal except for inflation news which affect it negatively
Remarks
Some of these results to be expected
Little effect of real news on deflator arises from poor relative predictability of the deflator
By the same reasoning, poor relative now-casting predictability of inflation and high relative predictability of real GDP imply that real news move the signal of nominal GDP more than nominal news
Some are interesting
Monetary aggregates news have no role for neither nominal GDP nor the deflator exactly as we had found for real GDP. The yield curve takes their role No Phillips curve relation seems to emerge from the effect of inflation news to real GDP
Considerations on Phillips curve
• Very little Phillips curve relation in the data and in the news analysis → real economy objective does not affect inflation even in the short-run?
• Literature has suggested that Phillips curve relations emerge conditional to large shocks (Giannone et al 2011) or around recessions (Stock and Watson 2012)
Use the model to study effect of different types of shocks: oil prices and unemployment
Report here only persistent effect of an increase in unemployment
Counterfactual analysis
• Use model (balanced panel version) to compute conditional forecasts given an assumed path of a given variable
Two exercises:
1. Condition on future path of oil price – 1 month jump and 6 month increase
2. Condition on future impact of unemployment – 1 month and 6 month increase.
Report here only last exercise as an illustration:
Cumulated change (over 6 month) on level: 1%, 5% and 10%
(10% corresponds approx to 1 point increase)
Persistent shock on unemployment effect on nominal GDP level
1300000.00 1350000.00 1400000.00 1450000.00 1500000.00 1550000.00
1 2 3 4 5
nominal gdp level base nominal gdp level cond 1%
nominal gdp level cond 5%
nominal gdp level cond 10%
Persistent shock on unemployment effect on the level of the deflator
111.50 112.00 112.50 113.00 113.50 114.00 114.50 115.00
1 2 3 4 5
deflator level base deflator level cond 1%
deflator level cond 5%
deflator level cond 10%
Implications in terms of CPI inflation
monthly growth rate
-1.00 -0.80 -0.60 -0.40 -0.20 0.00 0.20 0.40 0.60
1 2 3 4 5 6 7 8 9 10 11 12 13 CPI growth rate base
CPI growth rate cond 1%
CPI growth rate cond 5%
CPI growth rate cond 10%
Comments
• Phillips curve relation emerges in response to persistent shock in unemployment
• Result obtained in Giannone et al 2011 via a different model and euro area data is confirmed
– suggest tradeoff policy relevant during recessions – case for active output stabilization in the short run
Tentative Conclusions
On now-casting nominal GDP
• Now-casting is an appropriate framework for nominal GDP and now-casting model outperforms naïve constant growth when is able to exploit timely information
• Daily financial variables add to now-casting ability f nominal GDP (unlike what found for real GDP by Banbura et al) since they are valuable for capturing the nominal side.
• Monthly monetary aggregates do not have any role
• All the other results found in the real GDP now-casting literature are confirmed On Phillips curve tradeoffs
• No strong tradeoff between nominal and real variables as shown by the model fit and forecasting results for the deflator as opposed to the real GDP as well as the analysis of the news
• Tradeoff emerges for large shocks
Work in progress
• Fully real time
• Uncertainty
Revision errors and statistical errors important to understand economic significance of the news analyis