Thursday, April 29, 2021

The Pandemic Recession Ended Long Ago

Speaking of today's strong U.S. GDP release:

Question:  When the smoke clears and the Pandemic Recession ends, how will it stack up relative to its ancestors? 

Answer:  The smoke has already cleared -- the recession ended long ago -- and it was likely both the deepest and the shortest of all time.  

The NBER  chronology does not measure recession deepness, but ADS does, and it reveals that the Pandemic Recession is clearly the deepest U.S. recession since 1960 (and probably of all time, although ADS is calculated only from 1960).  The NBER chronology does measure recession duration, but the NBER has not yet announced the Pandemic Recession's ending date, so we don't know its ``official" duration.  But ADS is released in much more timely fashion than the NBER chronology, and ADS has long indicated a clear return to sustained positive growth by mid-May 2020 (as have other leading nowcasts, like the CFNAI).  This would make the Pandemic Recession not only the deepest recession at least since 1960, but also the all-time shortest, by a wide margin (presently the shortest is the six-month recession of early 1980).  In my view, any eventual claim that the recession ended later than May 2020 would be more an indication of reluctance to declare such a short recession than an unbiased assessment of economic reality.

Tuesday, April 27, 2021

Variational Bayes for Network Estimation

Check out the fine paper, "Fast and Accurate Variational Inference for Large Bayesian VARs with Stochastic Volatility," by Chan and Yu (2020), which builds on earlier work by Koop and Korobilis (2018) and Gefang, Koop, and Poon (2019) and much earlier work. By using a global approximating density, the variational approach abandons the naive hope of achieving perfection in the limit of an MCMC sequence, in exchange for massive speed gains even in very high dimensions (without sacrificing much accuracy when done well). 

Superficially Chan-Yu looks like just another "Bayesian VARs with stochastic volatility" paper (not that there's anything wrong with that!).  But here's a key thought chain: 

(1) Networks are productively characterized and understood by interpreting them as VARS and then using standard VAR estimation, decomposition, and visualization technology (see Diebold-Yilmaz et al., here and here and the references therein).  

(2) But many interesting networks are very high-dimensional, which presented a problem for taking network-VAR analysis to the next level, where, for example, one might want a 5000-dimensional network VAR.

(3) But ultra-high-dimensional situations are now much less problematic, thanks to variational Bayes.  Moving in that direction, Chan and Lu do variational Bayes for the DDLY bank network (approximately 100 dimensions), and moreover they incorporate stochastic volatility.

Friday, April 23, 2021

Apr. 30: Bridging Factor and Sparse Models

I blogged before on the Fan-Masini-Medeiros "Bridging Factor and Sparse Models" paper, here.  Now we can see it presented live on April 30 (16:00 Amsterdam time), together with the equally fascinating Pettenuzzo-Sabbatucci-Timmermann paper on firm behavior during the pandemic (15:00).  See you there!

You will have to register for the webinars in advance, at:

On behalf of the organisers Bart Keijsers and Sander Barendse:

We are happy to invite you to the UvA Financial Econometrics Workshop, which consists of two presentations around the topic of financial econometrics. It will be held online on

Friday 30 April, from 15:00 to 17:00.

Below is the list of the speakers, the schedule, and link to register for the workshop. The information is also available on the website:

List of speakers

  • Riccardo Sabbatucci (Stockholm School of Economics)
    Title: Dividend Suspensions and Cash Flows During the COVID-19 Pandemic: A Dynamic Econometric Model


Firms suspended dividend payments in unprecedented numbers and at unparalleled speed in response to the outbreak of the COVID-19 pandemic. We develop a dynamic econometric model that allows dividend suspensions to affect the conditional mean, volatility, and jump probability of growth in daily dividends and demonstrate how the parameters of this model can be estimated using Bayesian Gibbs sampling methods. We find that dividend suspensions had a sharp and immediate impact on the conditional mean and volatility of daily dividend growth. Information embedded in daily dividend suspensions proves valuable in monitoring and predicting the trajectory of broader measures of economic activity during the pandemic.

Coauthors: Davide Pettenuzzo (Brandeis) and Allan Timmermann (UCSD)

Personal website | SSRN link

 Marcelo Medeiros (Pontifical Catholic University of Rio de Janeiro)

  • Title: Bridging factor and sparse models
    Factor and sparse models are two widely used methods to impose a low-dimensional structure in high-dimension. They are seemingly mutually exclusive. In this paper, we propose a simple lifting method that combines the merits of these two models in a supervised learning methodology that allows to efficiently explore all the information in high-dimensional datasets. The method is based on a flexible model for panel data, called factor-augmented regression model with both observable, latent common factors, as well as idiosyncratic components as high-dimensional covariate variables. This model not only includes both factor regression and sparse regression as specific models but also significantly weakens the cross-sectional dependence and hence facilitates model selection and interpretability. The methodology consists of three steps. At each step, the remaining cross-section dependence can be inferred by a novel test for covariance structure in high-dimensions. We developed asymptotic theory for the factor-augmented sparse regression model and demonstrated the validity of the multiplier bootstrap for testing high-dimensional covariance structure. This is further extended to testing high-dimensional partial covariance structures. The theory and methods are further supported by an extensive simulation study and applications to the construction of a partial covariance network of the financial returns and a prediction exercise for a large panel of macroeconomic time series from FRED-MD database.

Coauthors: Jianqing Fan (Princeton) and Ricardo Masini (Princeton)

Personal website | SSRN link


15:00 - 16:00 Webinar Riccardo Sabbatucci (45 min talk + 15 min questions)

16:00 - 17:00 Webinar Marcelo Medeiros (45 min talk + 15 min questions)


You will have to register for the webinars in advance, via the following link:

Feel free to share this with your colleagues, and please let us know if you have any questions.

Kind regards,

Bart Keijsers ( and Sander Barendse (


---------------------------------------------------------------------------------- Netherlands Econometrics Study Group: NESG List: View, Join or Leave using

Wednesday, April 14, 2021

To my loyal email subscribers

Hi Email Subscribers,

Thanks so much for your No Hesitations support, and email subscriptions, over the years.  The message below just arrived.  Not sure what I'll be able to do -- we will see.  In any event your email subscription may be temporarily or even permanently interrupted or cancelled.  You can of course always view the blog directly at, or follow my Twitter posts (@FrancisDiebold) which link to the blog posts. I hope you will!


Hi Francis,
FeedBurner has been a part of Google for almost 14 years, and we're making several upcoming changes to support the product's next chapter. Here’s what you can expect to change and what you can do now to ensure you’re prepared.
Starting in July, we are transitioning FeedBurner onto a more stable, modern infrastructure. This will keep the product up and running for all users, but it also means that we will be turning down most non-core feed management features, including email subscriptions, at that time.
For those who use FeedBurner email subscriptions, we recommend downloading your email subscribers so that you can migrate to a new email subscription service.
For many users, no action is required. All existing feeds will continue to serve uninterrupted, and you can continue to create new accounts and burn new feeds. Core feed management functionality will continue to be supported, such as the ability to change the URL, source feed, title, and podcast metadata of your feed, along with basic analytics.
Learn More

Tuesday, April 13, 2021

Practical Guide to Climate Econometrics

This tutorial site is very nice for  climate data sources, computing environments for climate data manipulation, special climate data issues, etc., even if it's shallow on actual modeling.  Good for students. Hats off to the lead  authors, two of whom are indeed Ph.D. Students: Azhar HussainJames RisingKevin Schwarzwald, and Ana Trisovic.  Thanks to Glenn Rudebusch for alerting me to the site.  

Wednesday, April 7, 2021

Economic Forecasting Mini-Course


DEADLINE: Friday 7 May 2021
Is this email not displaying correctly?
View it in your browser.
Euro Area Business Cycle Network Training School


 Recent Developments in Forecasting


Graham Elliott (UC San Diego)
Allan Timmermann (UC San Diego)

Hosted online with Banca d’Italia, Italy

 1-8 June 2021


Deadline: 6pm (UK time), Friday 7 May 2021

General Description

We are pleased to announce details of the latest EABCN Training School; a three-day course entitled “Recent Developments in Forecasting”. Professors Graham Elliott and Allan Timmermann will teach the course. It is primarily aimed at participants in the Euro Area Business Cycle Network but applications will also be considered from doctoral students, post-doctoral researchers and economists working in central banks and government institutions outside of the network, as well as commercial organisations (fees applicable for non-network organisations).
Course Outline

The course introduces participants to a variety of advanced topics and recent developments in economic forecasting. The first part of the course examines the forecasting problem in general, showing that point forecasting is parameter estimation with a conditional model of the outcome and density forecasting is estimation of a conditional density. We clarify what we mean by optimal forecasting and relate classical and Bayesian approaches.
Understanding these issues provides a foundation for all forecasting problems. Binary forecasting or classification is most closely related to decision making. The simplicity of the loss function allows many strong results. Parametric, Semiparametric and nonparametric methods will be discussed and properties of the approaches examined.
Often the difference between good and bad forecasting approaches hinges on how they deal with changes to the underlying data generating process. The course therefore next addresses the consequences of model instability on forecasting performance and discusses strategies for dealing with such instability, using empirical illustrations from macroeconomics and finance. We also discuss how one can use multivariate (panel) information to better deal with model instability in a forecasting context.
A major issue in modern forecasting is the large number of potential predictors. Much work has been undertaken in econometrics, statistics and computer science in recent years. We provide a framework for thinking about methods as well as explain how some of the popular machine learning methods work and their properties. With this in place, we cover a variety of variable selection and model aggregation methods.

The final part of the course covers how to choose among competing forecasts and formally compare forecasting performance across two or possibly large numbers of forecasts. Ignoring the search across multiple models for a good forecasting model can introduce data mining biases, and we discuss ways to handle this problem.
The course draws on material from the following book (referred to as ET):
G. Elliott and A. Timmermann, 2016, Economic Forecasting. Princeton University Press.

Part I: Foundations and the Binary Forecasting Problem

  1. The Forecasting Problem
    1. Economic loss functions and ‘optimal’ forecasting  (ET chapter 2-3)
    2. Classical and Bayesian Forecasts (ET chapter 4-5)
  2. The Binary Forecasting Problem (ET chapter 12)
    1. Loss functions
    2. Point and Density Forecasting
    3. Methods for Classification/Binary Forecasting
Part II: Predictive Modelling Methods and Model Instability
  1. Forecasting under model instability
    1. Detection of breaks in time-series forecasting models (Rossi, 2013, Elliott and Mueller, 2006)
    2. Choice of estimation window in the presence of instability (Pesaran and Timmermann, 2007)
    3. Ad-hoc Strategies vs. Parametric Models of the Change Process (ET chapter 19, Pettenuzzo and Timmermann, 2011, 2017)
    4. Exploring Panel Data for Detecting and Forecasting under Breaks (Smith and Timmermann, 2017)
  2. Forecasting with Many Regressors
    1. Sparse vs. Dense Models: PCA, PLS, LASSO and variants
    2. Machine Learning Methods: Trees and neural nets (Gu, Kelly, and Xia, 2020, Coulombe et al, 2020)
  3. Model Selection and Forecast Combination Methods
    1. Model Selection Methods (ET, chapter 6)
    2. Model Aggregation approaches (Elliott, Gargano, and Timmerman, 2013).
Part III: Evaluating and Comparing Forecasting Performance
  1. Comparing forecasting performance: Horse races and p-hacking
    1. Comparisons of forecast performance.  (ET chapter 17, Clark and McCracken 2001, Diebold and Mariano 1995, Giacomini and White, 2006)
    2. Evaluating and comparing many forecasting models (White, 2000, Sullivan, Timmermann and White, 1999, Romano and Wolf, 2005, Hansen, Lunde, and Nason, 2011)
    3. Data mining and p-hacking (Harvey, Liu, and Zhu, 2016)
    4. Comparing forecasting performance in a single cross-section (Qu, Timmermann and Zhu, 2020)

Administrative information:

The course will take place online, in the evenings for Europe, from 5pm CEST:

  • June 1st lecture (3 hours)
  • June 3rd lecture (3 hours)
  • June 4th practice (1.5 hours)
  • June 7th lecture (3 hours)
  • June 8th practice (3 hours).

Candidates who have a CEPR profile should apply by submitting their CV online at by 6pm (UK time), 7 May, 2021. If you do not currently have a CEPR profile, please create a new one here and then click on the registration link.
PhD students should also send a statement that specifies the ways participating in the school will be useful for their current research (max 300 words).
Participants from non-academic institutions where the employer is not a member of the EABCN network are charged a course fee of €1000.
About the Instructors:
Graham Elliott is a professor of economics. He works in the field of econometrics, developing statistical methods for economic and other applications. He is a fellow of the Center for Applied Macroeconomic Analysis (CAMA), author of the reference/text "Economic Forecasting" jointly with Allan Timmermann, former co-editor of the International Journal of Forecasting (IJF) and co-editor of Volumes 1 and 2 of the Handbook of Forecasting.
Allan Timmermann holds a Atkinson/Epstein Chair in Management Leadership at the Rady School of Management and is also a professor in economics at UC San Diego's department of economics since 1994. He obtained his PhD from University of Cambridge after initial economics training at the University of Copenhagen. Timmermann is a very productive scholar in finance and applied econometrics. He serves as an associate editor on leading journals in finance, economics and forecasting including Journal of Business and Economic Statistics, Journal of Economic Dynamics and Control, Journal of Financial Econometrics, and Journal of Forecasting. He has published in journals such as Journal of American Statistical Association, Review of Economic Studies, Journal of Finance, and Journal of Econometrics.

For more information on EABCN, visit the website.

Monday, April 5, 2021

Local Projections vs. VARs

Interesting paper, showing an interesting LP (higher variance) vs. VAR (higher bias) tradeoff.

The authors conclude that "Unless researchers are overwhelmingly concerned with bias, shrinkage via Bayesian VARs or penalized LPs is attractive."  

A key point in terms of whether researchers are "overwhelmingly concerned with bias" is that it's not so much about researchers' preferences (innate feelings about bias) as it is about the data -- dynamic environments with large moving-average roots force concern with bias, because that's where low-ordered VARs are poor approximations, injecting large amounts of bias.

So: Much depends on how important large moving-average roots are in macroeconomic dynamics.  In principle they can be be very important (so un-penalized LP may be attractive).  In practice, well, usually it seems not so much (in which case penalized LP may be attractive).

Local Projections vs. VARs: Lessons From Thousands of DGPs

By:Dake LiMikkel Plagborg-M{\o}llerChristian K. Wolf
Abstract:We conduct a simulation study of Local Projection (LP) and Vector Autoregression (VAR) estimators of structural impulse responses across thousands of data generating processes (DGPs), designed to mimic the properties of the universe of U.S. macroeconomic data. Our analysis considers various structural identification schemes and several variants of LP and VAR estimators, and we pay particular attention to the role of the researcher's loss function. A clear bias-variance trade-off emerges: Because our DGPs are not exactly finite-order VAR models, LPs have lower bias than VAR estimators; however, the variance of LPs is substantially higher than that of VARs at intermediate or long horizons. Unless researchers are overwhelmingly concerned with bias, shrinkage via Bayesian VARs or penalized LPs is attractive.