The NY Fed DSGE: Its (Short but ArguablySuccessful) Past, Present, and (Uncertain) Future

DSGE TeamFederal Reserve Bank of New York

Large-Scale Econometric Models: Do they Have a Future?Royal Economic Society, March 2018

Disclaimer: The views expressed here do not necessarily reflect thoseof the Federal Reserve Bank of New York or the Federal Reserve

System

Outline

1 The NY Fed DSGE in a nutshell

2 Its short but arguably successful past (“DSGE Forecasts of the LostRecovery”)

3 The uncertain future

NY Fed DSGE Team The NY Fed DSGE RES 2

NY Fed DSGE

• Medium/largish-scale model with Smets & Wouters’ nominal andreal rigidities, and financial frictions as in Bernanke, Gertler, andGilchrist, 1999/Christiano, Motto, and Rostagno (2014)

• Observables (1960Q1-2017Q2): the growth rate of real output (bothGDP and GDI ), consumption, investment, real wage, hours worked,inflation (both core PCE and GDP), long run inflation expectations,the FFR, the ten-year Treasury yield, Fernald’s TFP growth, Baaspreads

• Model was developed in 2007-2009 and has been routinely in usesince late 2010 (although it evolved over time)

• Model’s code is available on GitHub and its documentation is keptup to date here

• 5th most popular Julia package ;-). 82 forks (people expanding onthe code on Github)

NY Fed DSGE Team The NY Fed DSGE RES 3

NY Fed DSGE

• Medium/largish-scale model with Smets & Wouters’ nominal andreal rigidities, and financial frictions as in Bernanke, Gertler, andGilchrist, 1999/Christiano, Motto, and Rostagno (2014)

• Observables (1960Q1-2017Q2): the growth rate of real output (bothGDP and GDI ), consumption, investment, real wage, hours worked,inflation (both core PCE and GDP), long run inflation expectations,the FFR, the ten-year Treasury yield, Fernald’s TFP growth, Baaspreads

• Model was developed in 2007-2009 and has been routinely in usesince late 2010 (although it evolved over time)

• Model’s code is available on GitHub and its documentation is keptup to date here

• 5th most popular Julia package ;-). 82 forks (people expanding onthe code on Github)

NY Fed DSGE Team The NY Fed DSGE RES 3

• Model has been “vetted” (?) academically, so to say, in thatversions of the NY Fed DSGE model have been used in variety ofpublished/refereed papers:

• “Safety, Liquidity, and the Natural Rate of Interest” withDomenico Giannone, Marc Giannoni, and Andrea Tambalotti,Brookings Papers on Economic Activity, Spring 2017

• “Inflation in the Great Recession and New Keynesian Models,”with Marc P. Giannoni, and Frank Schorfheide, AmericanEconomic Journal: Macroeconomics, 7(1), 2015

• “DSGE Model-Based Forecasting” with Frank Schorfheide,Handbook of Economic Forecasting Vol. II, 2013

• The prospect of reasonably good academic publications generatesincentives for the economists working on the model

NY Fed DSGE Team The NY Fed DSGE RES 4

• We try to advertise the model by publishing its results (often theoutcome of policy exercises) in blog posts:

• The Macro Effects of the Recent Swing in Financial ConditionsFederal Reserve Bank of New York Liberty Street Economics Blog,May 2016

• Why Are Interest Rates So Low?, Federal Reserve Bank of New YorkLiberty Street Economics Blog, May 2015

• Forecasting with Julia, Federal Reserve Bank of New York LibertyStreet Economics Blog, May 2017

• ...

• Since 2014 we regularly publish the DSGE forecasts on the LibertyStreet Blog: September 2014, May and December 2015, May and

November 2016, and in February, May, August and November 2017, and

March 2018

• We do not forecast because the DSGE is “good” at forecasting —we forecast with the DSGE to test the model

NY Fed DSGE Team The NY Fed DSGE RES 5

• We try to advertise the model by publishing its results (often theoutcome of policy exercises) in blog posts:

• The Macro Effects of the Recent Swing in Financial ConditionsFederal Reserve Bank of New York Liberty Street Economics Blog,May 2016

• Why Are Interest Rates So Low?, Federal Reserve Bank of New YorkLiberty Street Economics Blog, May 2015

• Forecasting with Julia, Federal Reserve Bank of New York LibertyStreet Economics Blog, May 2017

• ...

• Since 2014 we regularly publish the DSGE forecasts on the LibertyStreet Blog: September 2014, May and December 2015, May and

November 2016, and in February, May, August and November 2017, and

March 2018

• We do not forecast because the DSGE is “good” at forecasting —we forecast with the DSGE to test the model

NY Fed DSGE Team The NY Fed DSGE RES 5

How Did the NY Fed DSGE Models Fare in Terms ofForecasting?

• The NY Fed DSGE model was a simple three-equations NK DSGEfrom 2004 to 2008, and was not used for forecasting.

• In 2008 we began to build a medium-size DSGE with financialfrictions, which has been routinely used for forecasting since late2010.

• This was a a challenging period – also from a forecasting point ofview

• Deep recession not followed by a swift recovery — unlike inprevious post-war deep recessions. Instead, large and persistentoutput gaps (lost recovery)

• These large gaps were however not associated with negativeinflation, as a traditional Phillips curve relationship would havepredicted (“missing deflation”)

• Short term interest rates stuck at the ZLB. Unconventionalmonetary policy: QE and forward guidance

NY Fed DSGE Team The NY Fed DSGE RES 6

How Did the NY Fed DSGE Models Fare in Terms ofForecasting?

• The NY Fed DSGE model was a simple three-equations NK DSGEfrom 2004 to 2008, and was not used for forecasting.

• In 2008 we began to build a medium-size DSGE with financialfrictions, which has been routinely used for forecasting since late2010.

• This was a a challenging period – also from a forecasting point ofview

• Deep recession not followed by a swift recovery — unlike inprevious post-war deep recessions. Instead, large and persistentoutput gaps (lost recovery)

• These large gaps were however not associated with negativeinflation, as a traditional Phillips curve relationship would havepredicted (“missing deflation”)

• Short term interest rates stuck at the ZLB. Unconventionalmonetary policy: QE and forward guidance

NY Fed DSGE Team The NY Fed DSGE RES 6

How Did the NY Fed DSGE Models Fare?

“... most people outside the discipline who take one look at these models[DSGEs] immediately think they’re kind of a joke. They contain so manyunrealistic assumptions that they probably have little chance of capturingreality. Their forecasting performance is abysmal. Some of their coreelements are clearly broken. Any rigorous statistical tests tend to rejectthese models instantly, because they always include a hefty dose offantasy.”

NY Fed DSGE Team The NY Fed DSGE RES 7

How Did the NY Fed DSGE Models Fare?RMSEs: Jan 2011-Jan 2016

Blue Chip SPF SEP

1N = 18

2N = 18

3N = 18

4N = 18

5N = 18

6N = 18

7N = 14

8N = 7

0.2

0.4

0.6

0.8

1.0

Real GDP Growth

1N = 13

2N = 13

3N = 13

4N = 13

5N = 13

6N = 0

7N = 0

8N = 0

0.2

0.4

0.6

0.8

1.0

Real GDP Growth

1N = 18

2N = 17

3N = 13

4N = 3

0.2

0.4

0.6

0.8

1.0

Real GDP Growth

1N = 13

2N = 13

3N = 13

4N = 13

5N = 13

6N = 0

7N = 0

8N = 0

0.0

0.1

0.2

0.3

0.4

0.5

0.6

Core PCE Inflation

1N = 18

2N = 17

3N = 13

4N = 3

0.0

0.1

0.2

0.3

0.4

0.5

0.6

Core PCE Inflation

“Real” Real Time Forecasts from NY Fed DSGEEvolution of 4Q forecasts

Output Growth Core PCE Inflation2011Q1

2010 2011 2012 2013 20141

2

3

4

Real GDP Growth

2010 2011 2012 2013 20140.8

1.0

1.2

1.4

1.6

1.8

2.0

Core PCE Inflation

• Unlike the SEP participants, the NY Fed DSGE model projects aslow recovery from the financial crisis (Reinhard and Rogoff, 2009)

• Because it attributes the crisis to a financial shock

• Since June 2011, the NY Fed DSGE forecasts have been part of amemo produced four times a year for the FOMC. You can find onhttps://www.federalreserve.gov/monetarypolicy/fomc-memos.htm

the NY Fed DSGE forecasts for 2011 and 2012.

Responses to a Financial Shock

5 10 15 20 25 30

0.10.20.30.40.5

BAA - 10yr Treasury Spread

5 10 15 20 25 30-0.7-0.6-0.5-0.4-0.3-0.2-0.1

Nominal FFR

5 10 15 20 25 30-1.0-0.8-0.6-0.4-0.2

Real GDP

5 10 15 20 25 30

-0.18-0.15-0.12-0.09-0.06

GDP Deflator

5 10 15 20 25 30-2.0-1.5-1.0-0.5

0.00.5

Real Investment

5 10 15 20 25 30-1.0-0.8-0.6-0.4-0.2

Hours

Responses to a Monetary Shock

5 10 15 20 25 30

-0.5

-0.4

-0.3

-0.2

-0.1

0.0

Real GDP

5 10 15 20 25 30-0.020

-0.015

-0.010

-0.005

0.000

GDP Deflator

5 10 15 20 25 30-2.5

-2.0

-1.5

-1.0

-0.5

0.0

0.5

Real Investment

5 10 15 20 25 30

-0.5

-0.4

-0.3

-0.2

-0.1

0.0

Hours

• Output reverts much more quickly following a monetary shock

• Model’s forecast after the 1982 recession was for a fast recovery

SWFF Forecast of the 1982 Recession

1980 1981 1982 1983 1984 1985

-7.5

-5.0

-2.5

0.0

2.5

5.0

7.5

Real GDP Growth

Storytelling, Unobservables, and Counterfactuals

• DSGEs can produce unobservables, and explain them in terms of theshocks that hit the economy:

Natural Rate of Interest ... and Its Decomposition

1985 1995 2005 2015

−4

−2

0

2

4

Productivity Financial Other

−4

−2

0

2

4

• DSGEs, being full-fledged econometric models, can characterize theuncertainty surrounding these objects (unlike Cowles foundationmodels).

NY Fed DSGE Team The NY Fed DSGE RES 13

Storytelling, Unobservables, and Counterfactuals

• DSGEs can produce unobservables, and explain them in terms of theshocks that hit the economy:

Natural Rate of Interest ... and Its Decomposition

1985 1995 2005 2015

−4

−2

0

2

4

Productivity Financial Other

−4

−2

0

2

4

• DSGEs, being full-fledged econometric models, can characterize theuncertainty surrounding these objects (unlike Cowles foundationmodels). Can we trust DSGE models-implied uncertainty?

NY Fed DSGE Team The NY Fed DSGE RES 13

Time to Move Past Representative Agents/PerfectInformation/Linear DSGEs

• DSGEs are not all the same, but still, most (if not all) arenear-clones of Smets and Wouters.

• While there are good reasons for this, it is time to move on

1 Not enough diversification/too much homegeneity

2 Disconnect from academia

3 Vast swaths of the economy are completely left out of theanalysis (e.g., inequality – and in general anything to do withheterogeneity, financial markets,...)

• Laundry list of possible avenues: Agents heterogeneity withborrowing and liquidity constraints, financial considerations such asthe demand for safe assets, deviations from perfectinformation/rational expectations ...

NY Fed DSGE Team The NY Fed DSGE RES 14

• This does not mean discarding Smets and Wouters -typerepresentative agents models. That would be unwise as they willlikely be better than the new crop of models at many things (e.g.,forecasting, and describing aggregate data in general) for a while

• It is not easy – both computationally and econometrically – but let’sstart walking (as central bank’s modelers) along some of theseavenues!

NY Fed DSGE Team The NY Fed DSGE RES 15

From Talking the Talk to Walking the Walk

• Moving to non-linear and/or heterogeneous agents models involves abig computational investment

1 Choose the appropriate computing language: Matlab ⇒ Julia

2 Build the infrastructure:

• On the econometrics side:1) Tempered Particle Filter (Herbst and Schorfheide 2017);2) Sequential Monte Carlo (Herbst and Schorfheide 2014)

• On the model solution side:1) Dolo (Pablo Winant)2) Functional Linearization (David Childers)

NY Fed DSGE Team The NY Fed DSGE RES 16

From Talking the Talk to Walking the Walk

• Moving to non-linear and/or heterogeneous agents models involves abig computational investment

1 Choose the appropriate computing language: Matlab ⇒ Julia

2 Build the infrastructure:

• On the econometrics side:1) Tempered Particle Filter (Herbst and Schorfheide 2017);2) Sequential Monte Carlo (Herbst and Schorfheide 2014)

• On the model solution side:1) Dolo (Pablo Winant)2) Functional Linearization (David Childers)

NY Fed DSGE Team The NY Fed DSGE RES 16

Conclusions: Do DSGE Models have a Future?

• Medium-scale DSGE models kept some of the promises made in ealy2000s (Smets and Wouters): Theory-based models flexible enoughto fit the data out-of-sample, and useful in many dimensions forpolicy analysis (e.g., r*, tracing the macroeconomic impact offinancial shocks, ...)

• DSGE models — in the broad sense of structural models of theeconomy, based on modern dynamic macro theory but also empiricalmodels, estimated and tested on actual data — do have a future

• But only if they evolve: 1) embrace model diversity; 2) move beyondthe linear/perfect information/representative agent framework

NY Fed DSGE Team The NY Fed DSGE RES 17