## _wp1219

## Source details

**Canonical URL:** [_wp1219](https://www.imf.org/-/media/websites/imf/imported-full-text-pdf/external/pubs/ft/wp/2012/_wp1219.pdf)

## Other formats

- [Markdown version](/-/media/websites/imf/imported-full-text-pdf/external/pubs/ft/wp/2012/_wp1219.pdf.md)
- [Structured JSON version](/-/media/websites/imf/imported-full-text-pdf/external/pubs/ft/wp/2012/_wp1219.pdf.json)

---

### I. Introduction — framing and motivation
- Research gap:
  - Relationship between oil-price shocks and the macroeconomy lacks consensus on effects and magnitudes.
  - Vector autoregression (VAR) approaches widely used but subject to endogeneity and predictability weaknesses (Bernanke, Gertler, and Watson (1997); Wu and McCallum, 2005; Chinn, LeBlanc, and Coibon, 2005).
- Motivating concerns:
  - Endogeneity: VARs often cannot separate oil-price movements driven by exogenous shocks from movements reflecting endogenous responses to other structural shocks (example cited for 2002–2008).
  - Predictability: oil futures prices can strongly predict spot oil price movements, implying some observed changes were anticipated.
- Objective:
  - Develop market-information-based measures of exogenous oil-price shocks combining narrative and quantitative approaches to avoid endogeneity and predictability problems.

### II. Methodology summary — narrative plus quantitative, daily event database
- Core approach:
  - Combine narrative event identification from oil-market commentaries with quantitative oil-price forecasting to construct daily measures of exogenous oil-price shocks.
- Data collection and scope:
  - Sample period: January 3, 1984, to October 31, 2007 — a total of 5,971 trading days.
  - Sources: daily oil-market commentaries in Oil Daily, Oil & Gas Journal, Monthly Energy Chronology; cross-checked with Energy Chronology (EIA).
  - Events classified into 22 oil-event types; each trading day assigned one numerical code or multiple codes.
  - Three independent analysts read and classified events to minimize subjective bias.
- Two quantitative shock definitions:
  - “Predicting error” (modified futures-spot spread forecasting model):
    - Rolling sample: previous 200 trading days.
    - Regress realized oil-price changes on spreads between oil spot and futures prices at different horizons quoted by previous trading day end; unanticipated change at t+1 defines the predicting error shock.
    - Futures contracts beyond six months excluded.
  - “Log-price change”:
    - Shock measured as change in logarithm of the spot oil price on the day.
- Exogeneity definitions (three used for robustness):
  - Baseline definition: event-types 1 through 12.
  - Narrow definition: event-types 2 through 9.
  - Broad definition: event-types 1 through 12 and 15 through 17.
- Exclusions common to all three:
  - Event-types 13 and 14 (OPEC/non-OPEC production plan changes) excluded as likely endogenous.
  - Event-types 18 and 19 (global demand/efficiency changes) excluded from exogeneity sets and examined separately.
- Aggregation:
  - Daily shocks divided by event codes if multiple events on a day, aggregated into 22 monthly series, then combined per exogeneity definition to produce monthly oil-price shock measures.
  - Temporal aggregation assumption: monthly aggregation of daily shocks has no effect on impulse response coefficient estimates (caveat noted).

### III. Empirical application, VARX estimation strategy, and validation
- Econometric framework (VARX / rational distributed lag):
  - Xt = A0 + A1 t + A2(L) Xt−1 + B(L) Ot + εt
  - Xt contains: log real GDP, log CPI, level of the federal funds rate, and log real price of oil (log PPI for crude oil − log CPI).
  - Ot: observable exogenous oil-price shock measure from market information.
  - Six lags used; monthly sample: January 1984 to October 2007.
  - Monthly real GDP indicator obtained via Chow and Lin (1971) interpolation (monthly industrial production and total capacity utilization used as interpolators).
- Normalization and inference:
  - Oil-shock measures normalized so the peak response of the real price of oil is 10 percent.
  - Statistical inference: residual-based bootstrap with 1,000 replications; 95 percent standard percentile confidence intervals reported.
- Exogeneity validation:
  - Granger-causality joint test (example: “baseline, log-price change”): p-value = 36 percent (null that all 24 lag coefficients = 0).
  - Strong exogeneity test (Bierens, 2004) joint Likelihood Ratio χ2(28) test: p-value = 0.1035.
  - Individual baseline Granger-causality p-values:
    - Real GDP: 0.5511
    - CPI: 0.2108
    - Federal funds rate: 0.8797
    - PPI crude petroleum: 0.6745
  - Conclusion: lagged endogenous variables do not help predict constructed Ot; strong exogeneity tests support validity.

### IV. Main impulse-response findings — market-information-based baseline (normalized peak oil response = 10 percent)
- Real GDP:
  - Gradual decline with largest absolute response ~18 months after the shock.
  - Response becomes statistically significant three months after the shock and remains significant through 24 months.
  - 24-month cumulative output loss: 6.8 percent of a month’s real GDP (equivalent to about 0.6 percent of annual real GDP in two years).
- CPI:
  - Immediate upward shift; peak three months after the shock.
  - Average increase ~14 basis points over the 24-month horizon.
  - CPI remains 10 basis points higher than pre-shock level at 24 months.
- Federal funds rate:
  - Rises by a few basis points in first three months.
  - With declining real GDP and decelerating inflation, policy becomes accommodative.
  - 24-month cumulative decline in the federal funds rate: 2.6 percentage points (about 11 basis points lower than pre-shock level each month on average).
- Real oil price:
  - Hump-shaped response; peak at one month after the shock; increase remains statistically significant even eight months after the shock.
- Comparable qualitative results obtained when shock sizes computed by “log-price change.”

### V. Comparison with VAR-based measures and quantitative differences
- VAR-based measures constructed:
  - Asymmetric VAR-based: NOPI (Hamilton, 1996) ordered last in recursive VAR.
  - Symmetric VAR-based: change in log oil price ordered last.
- Volatility and correlations:
  - Symmetric VAR-based measure is the most volatile; asymmetric VAR-based measure the least volatile.
  - Correlations:
    - Baseline “predicting error” vs asymmetric VAR-based: 24 percent.
    - Baseline vs symmetric VAR-based: 23 percent.
    - Two VAR-based measures between themselves: 53 percent.
- Impulse-response magnitudes (24-month cumulative output loss, normalized peak = 10 percent):
  - Baseline (market-information based): Output: -6.75; Price: 0.14; Interest rate: -0.11
  - Symmetric VAR-based: Output: -2.91; Price: 0.05; Interest rate: -0.07
  - Asymmetric VAR-based: Output: -1.66; Price: 0.04; Interest rate: -0.04
- Interpretive summary:
  - VAR-based strategies typically yield substantially weaker and statistically insignificant output responses compared with market-information-based narrative measures.
  - Example: baseline 24-month cumulative output loss = 6.8 percent versus asymmetric VAR-based = 1.7 percent and symmetric VAR-based = 2.9 percent.

### VI. Narrative decomposition — effects by event types
- Demand-driven oil-price shocks (event-types 18 and 19 combined):
  - Real GDP increases and CPI declines after such shocks; federal funds rate barely moves.
  - Output response statistically insignificant (consistent with Kilian (2009) and small sample).
  - Timing: primarily occur after 2000 in sample.
  - Over time, adverse effect of higher oil price can dominate initial stimulus, eventually causing GDP decline while CPI can remain below pre-shock level.
- OPEC and non-OPEC production decision shocks (event-types 13 and 14):
  - Real GDP declines and CPI rises; peak oil-price response at three months.
  - Federal funds rate rises significantly in response to CPI increase.
  - Output loss weaker than exogenous shocks and statistically insignificant.
  - OPEC announcements most frequent: 742 trading days.
- Oil market-specific demand shocks (event-types 15–17: SPR changes, inventory adjustments, precautionary demand):
  - Real oil price rise less persistent — returns to pre-shock level in about five months.
  - Real GDP declines but returns to pre-shock level in about six months.
  - 24-month cumulative output loss: 2.2 percent of monthly GDP (compared with 6.8 percent for exogenous shocks).
  - CPI increases immediately but reverts in four months.
  - Short-lived responses imply limited monetary accommodation.
- Military actions in the Middle East:
  - Real oil price rises and real GDP declines substantially.
  - Largest GDP decline arrives 11 months after the shock; decline statistically significant at 95 percent for most horizons.
  - 24-month cumulative output loss: 8.4 percent of monthly GDP (larger than baseline exogenous shocks).
  - CPI increases after the shock but less than baseline exogenous shocks.
  - Monetary policy: 24-month cumulative decline in federal funds rate: 5.2 percentage points (or 22 basis points lower than pre-shock level on average), larger accommodative move than baseline (14-basis point decline).
  - Note: offsetting fiscal/military spending may mask the true oil-channel GDP decline.

### VII. Robustness, alternative exogeneity definitions, and stability across subsamples
- Alternative exogeneity definitions produce similar impulse responses to baseline (narrow and broad definitions; Figures 11 and 12 in appendix referenced).
- Robustness highlights:
  - Output responses remain statistically significant under alternative exogeneity definitions.
  - Under the “broad” definition, 24-month cumulative output loss reaches 8.4 percent (equivalent to 0.7 percent of a year’s real GDP in two years).
  - Alternative price measures and univariate distributed lag specifications generate no significantly different results.
- Stability across subsamples (January 1984–December 1994 vs January 1995–October 2007), reported 24-month aggregates (shock normalized to 10 percent peak real oil price response):
  - Asymmetric VAR-based:
    - 1984–1994: Output = -4.22; Price = -0.24; Interest rate = -0.21
    - 1995–2007: Output = 3.61; Price = 0.10; Interest rate = 0.34
    - Likelihood ratio test: p-value for identical impulse responses across subsamples rejected at 5% significance.
  - Symmetric VAR-based:
    - 1984–1994: Output = 0.48; Price = -0.04; Interest rate = 0.01
    - 1995–2007: Output = -2.68; Price = 0.07; Interest rate = -0.01
  - Market-information based:
    - 1984–1994: Output = -3.33; Price = -0.11; Interest rate = -0.10
    - 1995–2007: Output = -2.44; Price = 0.14; Interest rate = 0.17
    - p-value for identical impulse responses across subsamples = 32 percent (cannot reject null at common significance levels).
- Interpretation:
  - VAR-based measures show time variation (including a post-1995 positive output response for the asymmetric VAR), which may reflect increased prevalence of demand-driven oil-price shocks rather than disappearance of exogenous oil-shock real effects.
  - Market-information based measures indicate persistent adverse effects on real GDP across subsamples.

### VIII. Key numerical frequency and correlation statistics (1984–2007)
- Event frequencies (Column 3 of Table 1, 1984–2007):
  - “OPEC announcements on oil production”: 741 trading days, 12 percent of the sample.
  - “U.S. oil inventory announcements”: 730 days, 12 percent of the sample.
  - “Political developments in the Middle East”: 476 days, 8 percent of the sample.
  - Oil production or transportation disruptions in the U.S. and outside the U.S. (types 3 and 4): 486 trading days, about 8 percent of the sample.
- Correlations among constructed measures:
  - Predicting error-based correlations:
    - Baseline vs broad: 78 percent.
    - Baseline vs narrow: 72 percent.
    - Narrow vs broad: 55 percent.
  - Correlations across quantitative approaches:
    - Baseline “predicting error” vs baseline “log-price change”: 90 percent.
    - Broad measures across approaches: 87 percent.
    - Narrow measures across approaches: 88 percent.

### IX. Methodological procedures emphasized and intended advantages
- Procedures:
  - Compiled daily oil-market events database since 1984 from industry trade journals; classified events into 22 types.
  - Isolated events exogenous to the U.S. economy and constructed shock measures via a real-time oil-price forecasting model incorporating oil futures prices.
  - Aggregated daily shocks to monthly series for econometric analysis and preserved primitive information for alternative definitions and future work.
- Intended advantages:
  - Alleviate endogeneity and predictability problems affecting traditional VAR strategies.
  - Provide more reliable measures of exogenous oil-price shocks for macroeconomic analysis.

### X. Conclusions and implications
- Main findings:
  - Exogenous oil-price shocks have had substantial and statistically significant effects on the U.S. economy over the past two and a half decades.
  - Traditional VAR identification strategies imply substantially weaker and often insignificant real effects for the same period.
  - Discrepancy likely stems from VAR-based approaches’ inability to separate exogenous oil-supply shocks from endogenous oil-price fluctuations driven by demand changes.
- Policy and research implications:
  - The U.S. economy may not have become as insulated from oil shocks during the last two and a half decades as some VAR-based studies suggest.
  - Recommended future research: apply the narrative-plus-quantitative approach to the pre-“Great Moderation” period to examine the evolution of the oil price–macroeconomy relationship over the whole postwar period.

*Source: _wp1219 - References / Sections II–IV and conclusions (IMF working paper excerpt).*

### References .  .  .  .  .  .  .  .  .  .  .  .  .  .  .  .  .  .  .  .  .  .  .  .  .  .  .  .  .  .  .  .  .  .  .  .  .

### _wp1219 - References .  .  .  .  .  .  .  .  .  .  .  .  .  .  .  .  .  .  .  .  .  .  .  .  .  .  .  .  .  .  .  .  .  .  .  .  .

### I. Introduction — framing and motivation
- The relationship between oil-price shocks and the macroeconomy has been extensively studied over the past three decades but lacks consensus on effects and magnitudes.
- Vector autoregression (VAR) approaches have been widely used to identify exogenous oil-price shocks and estimate their effects, yet:
  - Estimated impacts of VAR-based measures on output and prices can be quite unstable across samples (Bernanke, Gertler, and Watson (1997)).
  - Traditional VAR-based measures exhibit two recurrent weaknesses: endogeneity and predictability.
- Endogeneity concern:
  - VAR approaches often cannot separate oil-price movements driven by exogenous shocks from movements reflecting endogenous responses to other structural shocks (example: oil price increases over the 2002–2008 period viewed by many as the result of “an expanding world economy driven by gains in productivity” (The Wall Street Journal, August 11, 2006)).
- Predictability concern:
  - Some observed oil price changes may have been anticipated by private agents. Studies (Wu and McCallum, 2005; Chinn, LeBlanc, and Coibon, 2005) find oil futures prices strongly predict spot oil price movements.
- These concerns motivate a different approach to obtain more reliable measures of exogenous oil-price shocks.

### II. Methodology summary — narrative plus quantitative, daily event database
- Core approach:
  - Combine narrative and quantitative approaches to develop new measures of exogenous oil-price shocks that avoid endogeneity and predictability concerns.
- Data collection and event identification:
  - Identify events that have driven oil-price fluctuations on a daily basis from 1984 to 2007.
  - Collect information from daily oil-market commentaries published in oil-industry trade journals such as Oil Daily, Oil & Gas Journal, and Monthly Energy Chronology.
  - Construct a database that identifies oil-related events occurring each day since January 1984.
- Event classification and shock construction:
  - Classify daily events into event types based on features (examples: weather changes in the U.S., military actions in the Middle East, OPEC announcements on oil production, U.S. oil inventory announcements) (see Table 1).
  - For each event type, construct a measure of oil-price shocks by running oil-price forecasting equations on a daily basis.
  - Select and aggregate shock series from exogenous oil events into a single measure of exogenous oil-price shocks.
- Design features:
  - By construction, the shock measures aim to be free of endogeneity and predictability problems.
  - Statistical tests are conducted to confirm exogeneity.
  - Multiple alternative definitions of exogenous oil-price shocks and corresponding shock measures are constructed for robustness.

### III. Empirical application and comparative results
- Targets of analysis:
  - Study responses of U.S. output, consumer prices, and monetary policy to exogenous oil-price shocks identified by the market-based methodology.
  - Compare estimated responses with those obtained using two traditional VAR-based identification strategies popular in the literature.
- Main empirical findings:
  - Market-based (new) measures yield substantial and statistically significant output and price responses to exogenous oil-price shocks.
  - Responses implied by VAR-based approaches are much weaker, statistically insignificant, and unstable over time.
  - Following a demand-driven oil-price shock, real GDP increases and the price level declines — consistent with scenarios where oil-price fluctuations are endogenous responses to changes in economic activity rather than exogenous shocks.
- Interpretation:
  - Traditional VAR-based approaches may fail to separate effects of exogenous oil shocks from endogenous demand-driven oil-price movements, leading to biased estimates of dynamic responses.

### IV. Relation to existing literature and contributions
- Connections to narrative approach literature:
  - Approach similar in spirit to narrative methods used by Romer and Romer (2004, 2010), Alexopoulos (2011), Alexopoulos and Cohen (2009), and Ramey (2009).
- Comparisons with prior oil-shock studies:
  - Earlier studies (Hamilton (1983, 1985); Hoover and Perez (1994); Bernanke, Gertler, and Watson (1997); Hamilton (2003); Kilian (2008)) identified discrete oil-related episodes or geopolitical events and examined effects on U.S. economy.
  - This study extends prior work by constructing a comprehensive daily database of all oil-related events and extracting the “unpredictable” component of oil-price fluctuations using an oil futures price-based forecasting model.
- External corroboration:
  - Kilian (2009) uses market information to disentangle oil-supply, global aggregate demand-driven, and oil market-specific demand shocks in a tri-variate VAR; effects estimated there are quite close to the empirical estimates from the present market-based approach, corroborating its validity.
- On time variation and the “Great Moderation” literature:
  - VAR studies (e.g., Hooker (1996), Blanchard and Galí (2009)) often find weaker, statistically insignificant relationships between identified oil-price shocks and real GDP growth in recent decades and cite reasons such as better economic policy or lower energy dependence (the “Great Moderation”).
  - Despite the “Great Moderation” characterization, estimation results from the market-based approach reveal substantial and significant adverse effects of exogenous oil shocks on the U.S. economy even during the last two and a half decades.
  - Time variation in VAR coefficient estimates may reflect inadequate identification rather than true disappearance of oil-shock effects.

### V. Organization of the paper (from the source)
- The paper proceeds as follows:
  - Section II describes the methodology to identify oil-related events and construct oil-price shock measures.
  - Section III illustrates the procedure to estimate macroeconomic effects of oil-price shocks.

*Source: _wp1219 - References .  .  .  .  .  .  .  .  .  .  .  .  .  .  .  .  .  .  .  .  .  .  .  .  .  .  .  .  .  .  .  .  .  .  .  .  .*

### Section IV presents our empirical results and compares them with those of earlier studies. Ro-

### Section IV presents our empirical results and compares them with those of earlier studies. Ro-

### II. MEASURES OF EXOGENOUS OIL-PRICE SHOCKS BASED ON MARKET INFORMATION
- Methodology consists of three key steps:
  - Collect daily information from oil-industry trade journals and other sources and classify events into 22 oil-event types.
  - For each event type, construct daily measures of oil-price shocks via an oil-price forecasting exercise to capture the unpredictable component of daily oil-price fluctuations.
  - Aggregate shock series corresponding to exogenous event types into a single measure of exogenous oil-price shocks; provide alternative definitions of exogeneity and construct corresponding aggregated measures.

### A. A Comprehensive Study of Daily Oil-Related Events
- Data and sources:
  - Information collected from oil-industry trade journals such as the Oil Daily and the Oil & Gas Journal; cross-checked with sources including the Energy Chronology published by the EIA.
  - Sample runs from January 3, 1984, to October 31, 2007, a total of 5,971 trading days.
- Event classification:
  - After reading market commentaries, oil-related events are classified into 22 types (examples: weather changes in the U.S., military actions in the Middle East, OPEC announcements on oil production, U.S. oil inventory announcements).
  - Each trading day is assigned one numerical code, or multiple codes if more than one event type occurred on the same day.
- Frequency findings (Column 3 of Table 1, 1984–2007):
  - “OPEC announcements on oil production”: 741 trading days, 12 percent of the sample.
  - “U.S. oil inventory announcements”: 730 days, 12 percent of the sample.
  - “Political developments in the Middle East”: 476 days, 8 percent of the sample.
  - Oil production or transportation disruptions in the U.S. and outside the U.S. (types 3 and 4): 486 trading days, about 8 percent of the sample.
- Content analysis and objectivity:
  - Three independent analysts read and classified events to minimize subjective bias.
- Justification of daily frequency:
  - Daily frequency chosen because the oil market is highly volatile and reacts immediately; daily is the highest frequency with available market information.

### B. Two Measures of Oil-Price Shocks
- Approach 1: “Predicting error” from a modified futures-spot spread forecasting model (based on Wu and McCallum 2005)
  - Estimation uses a rolling sample of the previous 200 trading days.
  - Estimating equation (informal description): regress realized oil-price changes on spreads between oil spot and futures prices at different horizons quoted by the end of the previous trading day; unanticipated change (realized at t+1) defines the predicting error shock for the day.
  - Futures contracts beyond six months are excluded due to liquidity concerns.
  - Changing the length of the rolling sample has negligible effects on forecasting results.
- Approach 2: “Log-price change” measure
  - Shock measured as the change in the logarithm of the spot oil price on the day.
  - Consistent with the view that futures prices may not have predictive content and that log oil price follows a random walk.
- Use in analysis:
  - Both measures are formulated daily and, via event identification, constructed around days of exogenous events; both are treated as legitimate measures of exogenous oil shocks.

### C. What Does Exogeneity Mean?
- Strict exogeneity criterion:
  - A genuine exogenous shock must be exogenous with respect to the U.S. economy; events potentially correlated with the U.S. economy are excluded.
  - Examples excluded as not exogenous: U.S. weather changes (event-type 1), U.S. production/transport disruptions (event-type 3), military conflicts (event-types 10–12), political developments (event-types 7–9) — because they may correlate with U.S. real GDP or defense spending.
  - Empirical note: over the past 25 years, there were only six trading days in which non-U.S. weather changes (event-type 2) significantly affected the oil market; no new oil field discoveries (event-types 5 and 6) had a noticeable impact.
- Three definitions of “exogeneity” provided and used for robustness:
  - Baseline definition: event-types 1 through 12 (includes U.S. and non-U.S. weather changes, oil production/transport disruptions, political and military actions).
  - Narrow definition: event-types 2 through 9 (non-U.S. weather changes, oil production/transport disruptions, political developments); excludes U.S. weather changes and military actions.
  - Broad definition: event-types 1 through 12 and 15 through 17 (adds events such as oil inventory announcements or changes in market expectations of oil inventories, described as “precautionary demand shocks”).
- Exclusions from all three definitions:
  - Event-types 13 and 14 (OPEC or non-OPEC oil exporters’ changes of production plans or proposals) excluded because they likely reflect producers’ endogenous responses.
  - Event-types 18 and 19 (changes in oil demand such as global economic growth, improvements in oil usage efficiency) excluded from these exogeneity definitions but examined separately in subsequent analysis.
- Analytic flexibility:
  - Authors construct and preserve primitive information to allow other researchers to select alternative definitions of exogenous events and construct corresponding measures.

### D. Constructing Monthly Oil Shock Series
- Aggregation procedure:
  - Daily shocks attributed to event types based on assigned codes; if multiple codes on a day, shocks equally divided among corresponding event types.
  - 22 daily shock series aggregated into 22 monthly series.
  - For each exogeneity definition (and other combinations of interest), a monthly oil-price shock measure is constructed for econometric analysis.
  - Assumption: temporal aggregation of daily shocks into monthly series has no effect on impulse response coefficient estimates (caveat noted; Marcellino 1999 referenced).
- Comparative properties of constructed measures:
  - Figures 1A and 1B display market-information-based measures (annual averages of monthly series) for shocks defined as the “predicting error” and the “log-price change.”
  - Correlations among measures (predicting error-based):
    - Correlation between baseline and broad measures: 78 percent.
    - Correlation between baseline and narrow measures: 72 percent.
    - Correlation between narrow and broad measures: 55 percent.
  - Correlations between the two quantitative approaches:
    - Correlation between baseline “predicting error” and baseline “log-price change”: 90 percent.
    - Correlation between broad measures across the two approaches: 87 percent.
    - Correlation between narrow measures across the two approaches: 88 percent.
- VAR-based comparative measures:
  - Two VAR-based measures widely used in the literature are displayed in Figure 1C:
    - “Asymmetric” VAR-based measure: based on the “net oil price increase” (NOPI) indicator (Hamilton 1996).
    - “Symmetric” VAR-based measure: based on the log change in the producer price index (PPI) for crude oil (as in Bernanke, Gertler, and Watson 1997; Blanchard and Galí 2009).
  - Both VAR-based measures are estimated residuals from a recursive VAR that includes macroeconomic variables and an oil-price indicator, with the oil-price indicator ordered as the last variable in the VAR system.

*Source: _wp1219 (IMF working paper excerpt).*

### Section III.

### Section III.

### Comparison of market-information-based and VAR-based oil-shock measures
- Both market-information-based measures and traditional VAR-based measures capture major oil-price spikes (March–April 1986, August–September 1990, December–February 1991, April 1999, September–October 2004), but magnitudes differ.
- Volatility ranking: symmetric VAR-based measure is the most volatile series; asymmetric VAR-based measure is the least volatile.
- Correlations:
  - correlation between the “baseline” oil-shock measure (“predicting error”) and the asymmetric VAR-based measure: 24 percent
  - correlation between the “baseline” and the symmetric VAR-based measures: 23 percent
  - correlation between the two VAR-based measures: 53 percent
- Key methodological distinctions:
  - VAR-based measures are residuals from vector autoregressions that include macroeconomic variables.
  - Market-information-based measures are residuals from an oil-price forecasting equation without macroeconomic variables or simply log changes of the oil price.
  - Narrative/event-study identification using market information classifies oil-price increases driven by global demand expansions (event-type 18 or 19) as non-exogenous and excludes them from exogenous-shock measures, whereas recursive VAR ordering may interpret such increases as shocks.

### VARX estimation strategy for assessing macroeconomic effects
- Estimating system (VARX/rational distributed lag):
  - Xt = A0 + A1 t + A2(L) Xt−1 + B(L) Ot + εt
  - Xt contains: log real GDP, log CPI, level of the federal funds rate, and log real price of oil (log PPI for crude oil − log CPI).
  - Ot is an observable exogenous oil-price shock measure (constructed from market information).
  - A2(L) and B(L) are finite-order polynomials in lag operator L; t is a time trend; εt is mean-zero i.i.d. white noise.
  - Dynamic responses k periods ahead obtained from expansion of [I − A2(L) L]−1 B(L).
- Practical implementation:
  - Six lags used in estimation.
  - Monthly sample period: January 1984 to October 2007.
  - Monthly real GDP indicator obtained using Chow and Lin (1971) interpolation (monthly industrial production and total capacity utilization used as interpolators).
  - Six alternative oil-price shock series available (baseline, broad, narrow definitions × log-price change or predicting error); exposition focuses on baseline.

### Exogeneity and validation of market-information-based shock series
- Granger-causality tests:
  - Tested whether lagged values of real GDP, CPI, federal funds rate, and PPI for crude petroleum jointly predict Ot.
  - Example: for the “baseline, log-price change” shock series, the p-value for null that all 24 coefficients (4 variables × 6 lags) are zero in a joint test is 36 percent.
  - Conclusion: lagged endogenous variables do not help predict constructed Ot.
- Strong exogeneity test (Bierens, 2004):
  - Joint test of zero coefficients plus zero covariance between residuals in Xt equations and residuals in Ot equation.
  - Likelihood Ratio test statistics (χ2(28) — 24 zero-coefficient restrictions plus 4 zero-covariance restrictions) fail to reject null of zero coefficients and uncorrelated residuals.
  - Conclusion: both Granger causality and strong exogeneity tests support validity and exogeneity of market-information-based shock measures.

### Impulse responses to market-information-based exogenous oil-price shocks (baseline)
- Normalization:
  - Oil-shock measures normalized so peak response of the real price of oil is 10 percent (roughly 1.75 times estimated standard deviations of market-information-based shocks).
  - Statistical inference via residual-based bootstrap with 1,000 replications; 95 percent standard percentile confidence intervals reported.
- Key impulse-response findings (baseline, shock magnitudes computed by “predicting error” and “log-price change”):
  - Real GDP:
    - Gradual decline; largest absolute response ~18 months after the shock.
    - Response becomes statistically significant three months after the shock and remains significant through the 24-month horizon.
    - 24-month cumulative output loss: 6.8 percent of a month’s real GDP (equivalent to about 0.6 percent of annual real GDP in two years).
  - CPI:
    - Shifts up immediately on impact; peak three months after the shock.
    - Average increase ~14 basis points over the 24-month horizon.
    - CPI remains 10 basis points higher than pre-shock level even 24 months after the shock.
  - Federal funds rate:
    - Rises by a few basis points in the first three months.
    - With declining real GDP and decelerating inflation, policy becomes accommodative.
    - 24-month cumulative decline in the federal funds rate: 2.6 percentage points (about 11 basis points lower than pre-shock level each month on average).
  - Real oil price:
    - Hump-shaped response; peak one month after the shock.
    - Oil price increase remains statistically significant even eight months after the shock.
- Similar qualitative responses obtained when shock sizes computed by “log-price change.”

### Impulse responses to VAR-based oil-price shock measures
- Two VAR-based shock measures constructed:
  - Asymmetric VAR-based measure: NOPI (Hamilton, 1996) as last-ordered variable in recursive four-variable VAR (order: log real GDP, log CPI, federal funds rate, NOPI).
  - Symmetric VAR-based measure: change in log oil price as last-ordered variable.
- Normalization: peak oil-price response set to 10 percent.
- Asymmetric VAR-based measure (Figure 4, left column) implications:
  - Real GDP declines and CPI rises, but output response is substantially weaker and no longer statistically significant.
  - 24-month cumulative output loss: 1.7 percent (about one quarter of the 6.8 percent loss implied by baseline).
  - CPI and federal funds rate responses weaker and statistically insignificant.
- Symmetric VAR-based measure (Figure 4, right column) implications:
  - Slightly stronger output response than asymmetric measure but still weaker than baseline.
  - 24-month cumulative output loss: 2.9 percent (still less than half of the 6.8 percent baseline loss).
  - Real GDP response statistically insignificant for most of the 24-month horizon.
- Comparison with Blanchard and Galí (2009):
  - Their second subsample (1984:Q1 to 2005:Q4) cumulative real GDP loss ~1.6 percent of quarterly GDP over three years (~0.13 percent of annual GDP each year).
  - Symmetric VAR-based estimate here (2.9 percent of monthly GDP in two years) corresponds to ~0.12 percent of annual GDP loss each year — quantitatively similar given overlapping samples and normalization choices.
- Overall conclusion: VAR-based identification strategies typically yield weaker and statistically insignificant output responses compared with market-information-based narrative measures.

### Narrative decomposition and types of shocks identified via market information
- Rationale:
  - VAR identification may conflate exogenous oil-price shocks with endogenous oil-price fluctuations driven by productivity or demand shocks.
  - Narrative approach enables classification of events (e.g., event-types 18 and 19 for demand-driven changes in oil usage efficiency, technology, etc.) and separate estimation of their macroeconomic effects.
- Demand-driven oil-price shocks (event-types 18 and 19 combined):
  - Impulse responses (Figure 6) indicate:
    - Real GDP increases and CPI declines following such shocks, despite non-core CPI components rising.
    - Federal funds rate barely moves (offsetting forces from higher output and lower general price level).
    - These responses match expectations for positive productivity shocks.
  - Note: output response statistically insignificant (consistent with Kilian (2009) and small number of observations).
  - Timing: such shocks occur primarily after 2000 in the sample, consistent with arguments that 2002–2008 oil-price increases were driven by expanding world economy and productivity gains.
  - Over time, adverse effect of higher oil price can dominate initial stimulus, causing eventual GDP decline, while CPI can remain below pre-shock level for extended periods.
- OPEC and non-OPEC production decision shocks (event-types 13 and 14):
  - Impulse responses (Figure 7):
    - Real GDP declines and CPI rises.
    - Peak oil-price response arrives three months after shock.
    - Federal funds rate rises significantly in response to CPI increase amid modest output decline.
    - Output loss weaker than exogenous shocks and statistically insignificant (OPEC announcements most frequent event type with 742 trading days).
    - Monetary policy response lies between responses to exogenous shocks (more accommodative) and demand-driven shocks (more restrictive).
- Oil market-specific demand shocks (event-types 15–17; e.g., SPR changes, inventory adjustments, precautionary demand):
  - Impulse responses (Figure 8):
    - Real price of oil rises but is less persistent than exogenous shocks — returns to pre-shock level in about five months.
    - Real GDP declines but returns to pre-shock level in about six months.
    - 24-month cumulative output loss: 2.2 percent of monthly GDP (compared with 6.8 percent for exogenous shocks).
    - Output decline statistically significant at 95 percent only for first three months after the shock.
    - CPI increases immediately but reverts to pre-shock level in four months.
    - Monetary policy accommodation limited due to short-lived responses.
    - These shocks are less persistent than precautionary-demand shocks identified by Kilian (2009), possibly because Kilian’s identification may encompass some exogenous event types.
- Oil-price shocks related to military actions in the Middle East:
  - Impulse responses (Figure 9):
    - Real price of oil rises and real GDP declines substantially.
    - Largest GDP decline arrives 11 months after the shock; decline statistically significant at 95 percent for most horizons.
    - 24-month cumulative output loss: 8.4 percent of monthly GDP (larger than 6.8 percent from baseline exogenous shocks).
    - CPI increases after the shock, but less than for baseline exogenous shocks.
    - Monetary policy: federal funds rate becomes more accommodative than baseline — 24-month cumulative decline of 5.2 percentage points (or 22 basis points lower than pre-shock level on average), larger than the 14-basis point decline for baseline exogenous shocks.
    - Note: potential offset from increased U.S. military spending could mean actual oil-channel GDP decline is even larger.

### Robustness and stability
- Authors perform a number of robustness and stability checks (details to follow in Section IV.D).
- Estimation alternatives referenced:
  - Univariate distributed lag models (Ramey and Shapiro, 1998; Kilian, 2009) previously estimated and produced effects similar to the VARX model.
  - Impulse responses for “broad” and “narrow” definitions of exogeneity are similar to baseline and discussed in robustness checks.

*Source: _wp1219 - Section III.*

### conclusions. First, we estimate the implied impulse responses using two alternative defini-

### _wp1219 - conclusions. First, we estimate the implied impulse responses using two alternative defini-

### Alternative definitions of exogeneity and robustness checks
- Two alternative definitions of “exogeneity” from Section II.C were used:
  - “Narrow” definition: event types 1 through 9 in Table 1.
  - “Broad” definition: event types 1 through 12, as well as types 15 through 17.
- Implied impulse responses using the narrow and broad definitions are very similar to those from the “baseline” definition (Figures 11 and 12 of the appendix).
- Key robustness findings:
  - Output responses to oil-price shocks remain statistically significant under alternative exogeneity definitions.
  - The 24-month cumulative output loss under the “broad” definition reaches 8.4 percent, or 0.7 percent of a year’s real GDP in two years following the shock.
  - Alternative measures of price level (e.g., the personal consumption expenditure deflator) and a univariate, distributed lag model specification generate no significantly different results and are omitted.

### Stability across subsample periods (January 1984–December 1994 vs January 1995–October 2007)
- Shock magnitudes are normalized so that the largest response of the real price of oil is 10 percent. Reported 24-month aggregates:
  - “Output” = 24-month cumulative output loss (sum of impulse response coefficients for output).
  - “Price” = sum of 24 monthly CPI impulse responses divided by 24 (average CPI increment over two years).
  - “Interest rate” = sum of 24 monthly federal funds rate impulse responses divided by 24 (average monetary policy response over two years).
- Asymmetric VAR-based measure:
  - 1984–1994: 24-month cumulative output loss = 4.2 percent (real GDP declines).
  - 1995–2007: 24-month cumulative output gain = 3.6 percent (real GDP increases).
  - Likelihood ratio test: the p-value for the null of identical impulse responses across the two subsamples can be rejected at the 5% significance level.
- Symmetric VAR-based measure:
  - 1984–1994: 24-month cumulative output = 0.5 percent (real GDP barely declines; consistent with literature on mid-1980s).
  - 1995–2007: 24-month cumulative output loss = 2.7 percent (real GDP declines).
- Market-information based measures:
  - Output response has not changed substantially across the two subsample periods; real GDP declines significantly after exogenous oil-price shocks by a similar amount in both periods.
  - p-value for the null of identical impulse responses across the two subsample periods = 32% (the null cannot be rejected at common significance levels).
- Interpretation note in text:
  - The positive output response implied by the asymmetric VAR-based measure for the post-1995 subsample may reflect a predominance of demand-driven oil-price shocks since the late 1990s (likely from global economic expansion), rather than genuine exogenous oil-price shocks.

### Numerical robustness table highlights (Table 3: 24-month cumulative effects; shock normalized so largest real oil price response = 10 percent)
- Baseline estimates:
  - Market-information based — Output: -6.75; Price: 0.14; Interest rate: -0.11
  - Symmetric VAR-based — Output: -2.91; Price: 0.05; Interest rate: -0.07
  - Asymmetric VAR-based — Output: -1.66; Price: 0.04; Interest rate: -0.04
- Subsamples 1984–1994:
  - Market-information based — Output: -3.33; Price: -0.11; Interest rate: -0.10
  - Symmetric VAR-based — Output: 0.48; Price: -0.04; Interest rate: 0.01
  - Asymmetric VAR-based — Output: -4.22; Price: -0.24; Interest rate: -0.21
- Subsamples 1995–2007:
  - Market-information based — Output: -2.44; Price: 0.14; Interest rate: 0.17
  - Symmetric VAR-based — Output: -2.68; Price: 0.07; Interest rate: -0.01
  - Asymmetric VAR-based — Output: 3.61; Price: 0.10; Interest rate: 0.34

### Tests of exogeneity (Table 2)
- Baseline definition — Granger causality test p-values (null: endogenous VARX variables have zero coefficients in equation of oil-shock series Ot):
  - Real GDP: 0.5511 (individual Granger causality test)
  - CPI: 0.2108
  - Federal funds rate: 0.8797
  - PPI crude petroleum: 0.6745
  - Joint Granger causality test (All four variables): 0.3628
- Strong exogeneity test (zero coefficients and zero correlations): p-value = 0.1035

### Methodological procedures emphasized in the conclusions
- Combined narrative and quantitative approach:
  - Collected oil-market related information from oil-industry trade journals and compiled a daily-event database affecting the global oil market since 1984.
  - Isolated events exogenous to the U.S. economy and constructed measures of exogenous oil-price shocks.
  - Calculated shock magnitudes via a real-time oil-price forecasting model incorporating oil futures prices.
- Intended methodological advantages:
  - Procedures help alleviate endogeneity and predictability problems affecting traditional VAR identification strategies.
  - Database preserves primitive information on oil-market developments to facilitate future studies.

### Main findings and implications
- Exogenous oil-price shocks have had substantial and statistically significant effects on the U.S. economy during the past two and a half decades.
- Traditional VAR identification strategies imply a substantially weaker and insignificant real effect for the same period.
- Discrepancy likely stems from VAR-based approaches’ inability to separate exogenous oil-supply shocks from endogenous oil-price fluctuations driven by changes in oil demand.
- The U.S. economy may not have become as insulated from oil shocks during the last two and a half decades as earlier studies have suggested.
- Recommended future research: apply the same narrative and quantitative approach to the period prior to the “Great Moderation” to examine how the oil price–macroeconomy relationship evolved over the whole postwar period.

*Source: _wp1219 - conclusions. First, we estimate the implied impulse responses using two alternative defini-*

---


_Source: https://www.imf.org/-/media/websites/imf/imported-full-text-pdf/external/pubs/ft/wp/2012/_wp1219.pdf_