Does Uncertainty Reduce Growth? Using Disasters as Natural Experiments

Transcription

1 Does Uncertainty Reduce Growth? Using Disasters as Natural Experiments Scott R. Baker (Northwestern) Nick Bloom (Stanford and NBER) Stephen Terry (Boston University) Melbourne Institute Macroeconomic Policy Meetings October 2016

2 Uncertainty as a potential driver of business cycles Many business people and policymakers believed this was a key factor in driving recent recessions: participants reported that uncertainty about the economic outlook was leading firms to defer spending projects until prospects for economic activity became clearer. -FOMC minutes, April 2008 Many economists have also a taken a similar view: the main contribution to the decline in output and employment during the recession are estimated to come from financial and uncertainty shocks -Stock and Watson (2012)

3 Uncertainty has also been in the media recently

4 This raises two obvious questions 1) Counter-cyclicality: Does uncertainty usually rise in recessions? 2) Causality: Is uncertainty a cause or effect of recessions?

5 Counter-cyclicality: uncertainty proxies rise in recessions (60 country panel, ) Uncertainty proxies (normalized to mean 0, SD 1) Stock index daily returns volatility Cross-firm daily stock returns spread Sovereign bond yields daily volatility Exchange rate daily volatility GDP forecast disagreement Annual GDP growth deciles (in deviation from country mean) Notes: Volatility indicators constructed from the unbalanced panel of daily data from 1970 to 2012 from 60 countries. Volatility values are calculated across all trading days (forecasts) within each year, and then normalized for presentational purposes so each of the four indicators has a mean of0 and a standard-deviation of1 over the sample. The GDP growth quintilesare calculated using annual values in deviations from the country mean across the sample.

6 Note, uncertainty also appears to be about 1/3 higher in developing countries In developing countries (<$10k GDP per capita) - GDP growth 50% volatile - Bond markets 35% volatile - Stock-markets 12% volatile Source: Baker and Bloom (2015) and Bloom ( fluctuations in uncertainty, 2013)

7 Causality: Do recessions cause uncertainty or the reverse? Here the literature has more mixed views Empirical papers: Uncertainty Slowdown: e.g. Bloom (2009), Arellano et al (2010), Christiano et al (2010), Gilchrist et. al (2010), Fernandez-Villaverde et al (2010, 2011), Basu & Bundick (2011), Ilut and Schneider (2011), Handley and Limao (2012) and Stein and Stone (2012) Slowdown Uncertainty: e.g. Bachmann and Sims (2010), and Bachmann, Elster and Sims (2011), Kehrig (2011) Theory papers: Uncertainty Slowdown: e.g. Bernanke (1983), Dixit & Pindyck (1994), Abel and Eberly (1996), Caballero and Engel (1999) Slowdown Uncertainty: e.g. Van Nieuwerburgh & Veldkamp (2006), Bachmann & Moscarini (2011), Fostel and Geanakoplos (2011), Boedo and D Erasmo (2012), Fajgelbaum, Schaal and Taschereau-Dumouchel (2012)

8 This paper evaluates the causal impact of uncertainty in driving growth using exogenous disaster shocks 1. For 60 countries, build a panel (from ) of stockmarket returns, volatility and GDP growth dataset 2. Regress GDP growth on stock returns and volatility (as 1 st and 2 nd moment proxies) instrumenting with natural disasters, terrorism, and political shocks in single equation and VAR estimation frameworks 3. Find that: IV estimates find large negative uncertainty effects on growth (about 5 times larger than OLS estimates) Uncertainty (2 nd moment) is at least equal in importance as demand and TFP (1 st moment) effects for GDP growth No impact of higher moments (3 rd and 4 th moments)

9 Simulation model Disaster Shocks and Economic Data IV Regressions Vector Autoregressions

10 Use simulated data to confirm our empirical strategy works before moving to real data Can we use stock-market returns and volatility as proxies for first and second moment shocks? Yes! Can a vector of disaster shocks be used to instrument for first and second moment shocks? Yes!

11 A micro-macro model with disaster shocks (1/5) Firms produce using a Cobb-Douglas technology as in Bloom, Floetotto, Jaimovich, Saporta-Eksten, and Terry (2014) Y = zak L β Business conditions are stationary AR(1) processes in logs with idiosyncratic firm & macro components: log z, = ρ.. log z,/0 + σ,/0 ε.,, ε., ~N 0,1 log A, = ρ 9 log A,/0 + σ 9,/0 ε 9,, ε 9, ~N 0,1

12 A micro-macro model with disaster shocks (2/5) Uncertainty process σ, 9 and σ,. are 2-point Markov chains Because micro and macro uncertainty highly correlated in data assume σ, 9 and σ,. move together but have different baseline values

13 A micro-macro model with disaster shocks (3/5) Assume adjustment costs for factors using values estimated in Cooper and Haltiwanger (2006) and Bloom (2009) Assume uncertainty Markov chains estimated from US Census manufacturing micro-data on the variance of establishment-level total factor productivity -TFP measured using SIC 4-digit factor share approach Assume stock market returns are noisy with same variance of noise as true variance (e.g. IID noise-trading)

14 A micro-macro model with disaster shocks (4/5) Add disaster shocks to the simulation. Have 4 types in the data: Natural disasters: e.g. earthquakes, volcanoes, hurricanes, fires, storms, floods and typhoons Terrorist attacks: eg. bombings, coordinated attacks Revolutions: overthrow or attempted overthrow of the government by a group outside the ruling elite Political coups d etats: overthrow or attempted overthrow of the government by a ruling elite, generally the military

15 A micro-macro model with disaster shocks (5/5) Assume macro disaster shocks to have the following properties: First moment (A, ) Second moment (σ, ) Natural -0.5σ L 0 Terrorism -0.5σ L Jump to σ H Revolution -2σ L Jump to σ H Coups +0.5σ L Jump to σ H Shock size and frequency calibrated to match real data

16 Solve at the firm level and aggregate Simulated a panel of firms quarterly for 500 countries for 25 years, clustering within country for SE s Generate simulated quarterly data for each country: - GDP data is total output summed across firms - Stock return levels are average mean returns - Stock return volatility are the standard deviation of returns Consider measures of returns, volatility and growth over the last quarter ( quarterly ) and over the last year ( yearly )

17

18 Simulation model Disaster Shocks and Economic Data IV Regressions Vector Autoregressions

19 Economic Data Unbalanced quarterly panel of 60 countries from 1970 onwards, defined as all country-quarter observations with: Quarterly real GDP data Daily stock index data Disaster Data Centre for Research on the Epidemiology of Disasters (joint UC Louvain and WHO institute based in Brussels, founded in 1971) Center for Systemic Peace (George Mason Institute based in Washington DC, founded 1997)

20 These disaster shocks seem unpredicted based on forecast tests (below) and media tests Forecast test (Appendix A1): No predictive power for shocks from lagged GDP growth, stock volatility and returns

21 Media tests examine the frequency of the country in search results from Access World News

22 Media test: Jump in news articles around the shock Count of country word counts (each event normalized so that before averages to one) days before 15 days after Days before and after the shock Notes for the figure: Shows the daily count of the name of the impacted country in the two weeks before and after the shock, averaged over the 1353 shocks studied in the regression analysis. For graphing purposes the series for each event is normalized so that over the 15 days before the shock it has a mean of one. In the regressions events are weighted by the increase in cites in the 15 days after the event compared to the 15 days before to focus on the jump in cites after an event.

23 In comparison, anticipated events get no jump: news articles around General Elections Count of country word counts (each event normalized so that before averages to one) days before 15 days after Days before and after the election Notes for the figure: Shows the daily count ofthe name ofthe impacted country in the two weeks before and after the election, averaged over the 133 elections in the G20 countries our sample. The series for each event is normalized for graphing so that over the 15days beforethe election it has a mean of one.

24 In comparison, anticipated events get no jump: news articles around the World Cup & Super Bowl Count of articles mentioning country of shock (each event normalized so that before averages to one) days before 15 days after Days before and after the World Cup or Super Bowl Notes for the figure: Shows the daily count of the name of the impacted country in the two weeks before and after the World Cup and Super Bowl, averaged over 52 events since The series for each event is normalized for graphing so that the 15 days beforethe election hasa mean of one.

25 Use this jump in media coverage to weight shocks Instrument in each quarter is a disaster type indicator scaled by the % jump in news articles containing the country name from the 15 days before to the 15 days after e.g. use the 322% increase in the count of the word Japan in fifteen days after the March 11 th 2011 earthquake to weight this shock by 3.22 This weighting method means that we focus on the largest and most surprising disaster shocks

26 Simulation model Disaster Shocks and Economic Data IV Regressions Vector Autoregressions

27 Regress GDP growth on stock market returns and volatility = β B + β 0 + β F + + T, + Instrumenting returns and volatility with: Natural disaster indicator % jump in citations Terrorist attack indicator % jump in citations Political coup indicator % jump in citations Revolution indicator % jump in citations Cluster standard errors by country

28

29 Third and fourth moments not significant

30 Main results similar for other uncertainty proxies Bond Yields Use 10-year bond yields across all countries Mean and volatility of daily returns Exchange Rates Exchange Rate vs. USD For USA, use trade-weighted rates Mean and volatility of daily returns Consensus Forecasts GDP Mean and Spread Loses power, only 23 country panel covered

31 Results seem robust to a variety of tests, like dropping media weighting, population weighting, and wealth and time splits

32 Simulation model Disaster Shocks and Economic Data IV Regressions Vector Autoregressions

33 Estimate a parsimonious panel VAR system for growth, stock returns and volatility = ) Include time and country FE: = + g, + + Cannot see true structural shocks: Estimating the response of the economy to underlying structural shocks e it - especially e vol it - is hard because we see the reduced form combination = Beit.

34 Can now use IV approach with the VAR OLS Impulse Responses Assumes that B = I, analyze the response to the average observed volatility innovation based on estimated A alone. -Standard VAR approach Disaster Instrument Impulse Responses Use the disaster instruments covariance with the observed innovations to identify the underlying shocks e vol it to volatility and the response matrix B via GMM. -Requires similar exclusion restrictions as in a traditional IV setting

35 A Disaster Instruments VAR reveals that volatility shocks drive recessions GDP Growth, Percent Year-on-Year Data VAR Model VAR Quarter, Shock in Period 1 Notes: The figure shows the response of GDP growth to a one-standard deviation innovation in volatility. The sample is an unbalanced panel of 60 nations with approximately 6,000 quarterly observations and an average of 25 years of data per nation. GDP growth in t is the percentage growth from quarter t-4 to t. The estimated VAR includes country and year dummies. Standard errors are based on a block bootstrap at the country level, with 90% intervals plotted. Model and data use identical panel structures, sample sizes, and datadefinitions.

36 Conclusion 1. IV and VAR results suggest uncertainty has a negative causal impact on GDP growth (5x larger than OLS estimates), using model-validated identification strategies 2. Uncertainty appears at least as important in explaining growth fluctuations as first moment shocks (in fact about three times as important in our baseline estimation) 3. Variations in higher (third and fourth) moments appear to play little empirical role at this frequency

37 Currently working on building US and international policy uncertainty data (with Steve Davis)

38 Significant variation of Economic Policy Uncertainty over time Policy Uncertainty Index Gold Standard Act Great Depression, New Deal and FDR Assassination of McKinley Versailles conference Start of WW I McNary Haughen farm bill Great Depression Relapse Post-War Strikes, Truman- Dewey Watergate Black Monday OPEC II OPEC I Berlin Conference Gulf War I Lehman and TARP 9/11 and Gulf War II Asian Fin. Crisis Debt Ceiling Notes: Index of Policy-Related Economic Uncertainty composed ofquarterly news articles containing uncertain or uncertainty, economic or economy, and policy relevant terms (scaled by the smoothed total numberof articles) in 5 newspapers (WP, BG, LAT, WSJ and CHT). Data normalized to 100 from

39 Does Uncertainty Reduce Growth? Using Disasters as Natural Experiments Scott Baker (Northwestern) Nick Bloom (Stanford and NBER) Stephen Terry (Boston University) Melbourne Institute Macroeconomic Policy Meetings October 2016

40 BACK-UP SLIDES

41 Top Newspaper Frequencies from Searches Newspaper Percent of Articles New York Times 7.86 Los Angeles Times 3.72 BBC 2.32 Chicago Tribune 1.98 Washington Post 1.66 Toronto Star 1.37 Boston Globe 1.21 New Straits Times 1.19 USA Today 1.12 Deseret News 1.11 Pittsburgh Post-Gazette 1.11 Christian Science Monitor 1.03 Toledo Blade 1.01 The Sun.99 St. Petersburg Times.97 Sydney Morning Herald.93 Sarasota Herald-Tribune.90

42 Descriptive statistics (60 countries, ) Obs. Mean Median St. Dev. Min Max GDP Growth Stock Returns Stock Ret. Vol. (log) Ex. Rate Vol. (log) Natural Disasters Political Shocks Revolution shock GDP Per Capita (2005 $US) ,450 18,477 16, ,559

43 Economic Data Unbalanced quarterly panel of 60 countries from 1970 onwards, defined as all country-quarter observations with: Quarterly real GDP data Daily stock index data Disaster Data Centre for Research on the Epidemiology of Disasters (joint UC Louvain and WHO institute based in Brussels, founded in 1971) Center for Systemic Peace (George Mason institute based in Washington DC, founded 1997) Keep disasters that meet one of the following: More than.001% of a country s population dead More than.01% of a country s GDP in damage Successful coup or regime change