Discussion of: Ditch the Ex Measure core inflation with a disaggregate ensemble by Francesco Ravazzo

1
Discussion of Ditch the Ex! Measure core
inflation with a disaggregate ensembleby
Francesco Ravazzolo (NB) and Shaun P. Vahey
(MBS)James Mitchell
2
Ravazzolo and Vahey

• Focus on density forecasts or fan charts
• Improved methods of constructing density
  forecasts given uncertain instabilities
• Ensemble forecasting a means of integrating out
  model uncertainty and telling us the weight
  (maybe time-varying) on a given component
• Application to core US PCE inflation
• Leave aside the question of whether one should
  target the PCE deflator or something broader
  (incl. asset or house prices? e.g. see Wadhwani
  Oct. 2008 NIER)
• Ensemble methodology can be applied to any
  inflation measure

3
Structure of comments

• 1 slide on RVs application to core inflation
• Core inflation is trend inflation, and trend
  inflation is a forecast of future inflation
• Agnostic on whether trend is I(1), I(0), I(0)
  with breaks. or whatever
• Focus more generally on RVs suggestions for
  macro modelling
• Density forecasting
• Ensemble modelling or combination forecasts

4
Core inflation

• RV define core inflation as forecasted
  inflation
• Consistent with a BN trend pt E(pth Ot)
• Evaluate core inflation estimates relative to
  inflation outturns
• The difference between actual and core (trend)
  inflation (i.e. the cycle) should predict future
  changes in inflation (with an opposite sign)
  e.g. see Cogley (2002, JMCB)
• The role of h and O
• Core inflation will vary depending on the
  horizon (h) of interest. RV focus on h1
• Ensemble modelling can integrate out uncertainty
  about the appropriate information set to use
• Core inflation restricts O to disaggregate
  inflation measures
• Is this a helpful partitioning of the information
  set?
• If not, whats the point of core inflation?
• What about the possible role of other macro
  information (e.g. output gap, asset prices,
  judgment) when forecasting inflation?

5
Density forecasts

• Density forecasts provide a full impression of
  the uncertainty associated with a forecast
• RVs focus on them and their better construction
  is welcome
• We shouldnt be surprised by the unreliability of
  point forecasts
• What should we expect from our density
  forecasters?
• Supply, like RV, well-calibrated densities?
• And/or tell a story? Depends on the components
• Even with ensemble modelling the two can clash
  what does a weight of 0.5 on the DSGE mean? Which
  half of the story should I believe?

6
Uncertainty about uncertainty Well-calibrated
densities a refresher course

• While density forecasts cannot capture unknowable
  (Knightian) uncertainty and only capture knowable
  uncertainty, these uncertainty assessments can
  and should be evaluated ex post - just like point
  forecasts
• Tests can be constructed based on the inflation
  outturns, circumventing the problem that we dont
  know the true density that determines inflation

7
Testing the densities minor technical
suggestions

• RV consider subset of available PIT tests
• Maintained normality assumption
• Ignore misspecification in moments higher than
  the second
• Efficiency tests have been proposed to
  discriminate between competing densities
  Mitchell Wallis (2009)
• Do food and energy prices explain the density
  forecasting errors made by the Ex?
• Or even the output gap?

8
Use of density forecasts

• But well-calibrated densities are not a solution
  (for IT) per se, as good policy is contingent
  on how they are used
• Regardless of a users loss function, a
  well-calibrated density (in all regions) is
  superior to all others. Hence RVs drive for
  well-calibrated densities is sensible
• Use (or the abuse) of density forecasts depends
  on the policymakers loss function
• Should you blame the (weather) forecaster, if
  they predicted a non-zero (or even zero)
  probability of rain, if you get wet because you
  chose not to take an umbrella?
• How much does an umbrella cost?
• Policymaking with uncertainty Brainard (1967,
  AER)

9
The appropriate benchmark

• Ensembles as the new benchmark
• Better than an AR which is competitive for RMSE
  loss only see RV and Jore, Mitchell Vahey
  (Forthcoming. JAE)

10
Ensemble density forecasting

• Ensemble density forecasts can accommodate
  future, model (including parameter and policy)
  and data uncertainty
• Data uncertainty seems to be ignored in RV -
  despite Garratt and Vahey (2006, EJ)
• PCE inflation is revised considerably and is
  forecastable see Croushore (2007)
• Flexible. Ensembles can accommodate time
  variation even if the components are
  time-invariant so long as the weights are time
  varying

11
Output gap uncertainty
12
Rubbish In, Rubbish Out

• We still need to select good components ex ante
• Choosing the weights on each component density
• Expect weight on components to change
• What if we have no data on tail events which,
  with hindsight, prove to be of interest? How
  quickly will ensemble methods adapt to a crisis?
• Forecasting GDP growth in the current recession.
  It appears to pay to give more weight, say, to
  qualitative survey data than history suggests
• Ensembling maybe adaptive (learning), but
  uncertainties are acknowledged upfront
• Whats the role of our prior (or judgment) over
  (i) the preferred components to include in the
  ensemble (ii) their weight and (iii) their
  dependence?

13
What components should we use?

• Use an expert combination framework to combine
  forecast densities from VARs and a DSGE model
  (NEMO - the Norges Bank core policymaking model)
• Joint with Ida Wolden Bache (NB), Anne Sofie Jore
  (NB) and Shaun Vahey (MBS)
• Forecasting Norwegian inflation
• Helps to include components that allow for
  time-variation
• On-going work with ensembles of DSGEs

14
Recursive logarithmic score weights on the DSGE
in the grand ensemble