## wpiea2021136-print-pdf

## Source details

**Canonical URL:** [wpiea2021136-print-pdf](https://www.imf.org/-/media/files/publications/wp/2021/english/wpiea2021136-print-pdf.pdf)

## Other formats

- [Markdown version](/-/media/files/publications/wp/2021/english/wpiea2021136-print-pdf.pdf.md)
- [Structured JSON version](/-/media/files/publications/wp/2021/english/wpiea2021136-print-pdf.pdf.json)

---

### I. INTRODUCTION — Abstract and Key Findings
- Paper: "Patterns in IMF Growth Forecast Revisions: A Panel Study at Multiple Horizons" (WP/21/136).
- Main empirical findings (as stated in the Abstract):
  - (i) Growth revisions in horizons closer to the actual are generally larger, more volatile, and more negative.
  - (ii) On average, growth revisions are in the right direction, becoming progressively more responsive to the forecast error gap as horizons get closer to the actual year.
  - (iii) Growth revisions in systemic economies are relevant for growth revisions in all country groups.
  - (iv) WEO and Consensus Forecast growth revisions are highly correlated.
  - (v) Fall-to-spring WEO revisions are more correlated with Consensus Forecasts revisions compared to spring-to-fall revisions.
  - (vi) Across vintages, revisions for a given time horizon are not autocorrelated; within vintages, revisions tend to be positively correlated, suggesting perception of persistent short-term shocks.
- Research questions addressed:
  - Are WEO growth revisions in the right direction?
  - What are the key drivers explaining WEO growth revisions?
  - How do WEO revisions compare to Consensus Forecast revisions?
  - Is there mean reversion/persistence in WEO growth forecasts?
  - How do results vary across time-horizons and country groupings?
- Role and motivation:
  - Focus on forecast revisions (rather than forecast errors) to gain dynamic insights into the forecasting process and potential efficiency improvements.
  - WEO database: biannual (Spring and Fall) WEO vintages over 1990–2019 for 194 economies; forecasts for year t consist of 12 vintages from year t-5 to year t, covering forecasts until 2024.

### III. DATA AND METHODOLOGY
- Dataset and coverage:
  - Source: WEO growth and ToT forecasts from Celasun et al. (forthcoming).
  - Coverage: Spring and Fall WEO vintages over the period 1990-2019 for 194 economies.
  - For forecasts of year t: 12 vintages from year t-5 to year t (Spring and Fall for each of these six years) — dataset covers forecasts until 2024.
  - Outturns: main analysis uses outturns measured at t+1; robustness shown for outturns measured at t+5.
  - Consensus Forecasts: March and September vintages matched to Spring and Fall WEO vintages.
- Definitions and notation:
  - Forecast revision (REV) of variable X for country and year t: for every i = 0,1,...,4 and t = 1990:2023, REV defined as forecast F made in (t-i) minus forecast made in (t-i-1).
  - "i" denotes the time horizon (how many years ahead the forecast is made).
  - Primary focus often on F0-F1 (F(t-0,t) − F(t-1,t)), revisions to forecasts for the current year.
- Regression specifications (five core panel regressions run for each horizon i; Hausman tests used to choose between RE and FE):
  1. Directional test (are revisions in the right direction?): RREV_Y(t-i,t) = β0 + β1 FR_Y(t-i-1,t) + ε_t, where FR_Y is prior forecast error F(t-i-1,t) − A(t+1,t).
  2. Impact of systemic economies’ GDP revisions: RREV_Y(t-i,t) = β0 + β1 RREV_X1(t-i,t) + ε_t (X1 = US, China, Euro Area, G7, or World).
  3. Impact of terms-of-trade (ToT) revisions: RREV_Y(t-i,t) = β0 + β1 RREV_X2(t-i,t) + ε_t (X2 = percent change in WEO ToT).
  4. Joint impact of systemic GDP and ToT revisions: RREV_Y(t-i,t) = β0 + β1 RREV_X1(t-i,t) + β2 RREV_X2(t-i,t) + ε_t.
  5. WEO versus Consensus Forecast revisions: RREV_Y(t-i,t) = β0 + β1 RREV_CF_Y(t-i,t) + ε_t.
- Panel regressions run for full sample and country-group subsamples; RE and FE coefficients generally quite similar.

### IV. DESCRIPTIVE FINDINGS
- Stylized facts:
  - Growth revisions closer to the actual vintage are generally larger, more volatile, and more negative.
  - Distributional extremes: World growth revisions in the latest horizon (F0-F1) sometimes range from -70 percentage points to +80 percentage points.
  - Visual analyses focus on core segment of distribution within -10 to +10 percentage points.
- Regional and historical highlights:
  - Growth revisions closest to the actual (F0-F1) often negative, especially during the GFC and across income groups.
  - Across EMDE regions, growth revisions are more likely to be negative; large negative adjustments visible during:
    - transition in Eastern Europe in the early 1990s,
    - the Asian Crisis in the late 1990s,
    - the Arab Spring in the 2010s.
  - Emerging and Developing Europe: 25th percentile of the distribution of growth revisions is -28 ppts in 1991 and -11 ppts in 2009.
  - Most action across time horizons occurs in the latest revision (between the Fall WEO in the year for which the forecast is made and the Fall WEO of the previous year); earlier-horizon revisions are considerably more stable.

### V. ARE REVISIONS IN THE RIGHT DIRECTION? (Specification (1)) — Directional Evidence and Robustness
- Key regression interpretation:
  - Negative β1 implies revisions narrow previous period’s forecast error; estimated β1 are negative for all horizons and country groups.
  - Absolute value of β1 increases (becomes more negative) for vintages closer to the actual, implying stronger corrections for forecasts made closer to the actual year.
- Robustness to outturn updates:
  - Using outturns measured at t+5 instead of t+1 yields the negative relationship between prior forecast errors and subsequent revisions robustly, though relationship weakens for EMDEs (suggesting outturn revisions after t+1 incorporate factors not visible in real time).
- Significance convention in figures:
  - Black dots denote significance at 10 percent level; red dots indicate lack thereof.

### VI. ROLE OF SYSTEMIC ECONOMIES’ GDP REVISIONS
- Hypothesis: revisions for systemic economies (e.g., US) affect revisions for other economies via trade and financial linkages.
- Empirical findings:
  - US growth revisions influence growth revisions across the world: for advanced economies across all time horizons; and for other countries in later horizons.
  - Impact is larger for revisions closest to the actual (higher β1 in the latest horizons).
  - In univariate specifications, impact of US growth revisions is significant for all country groups and regions; in multivariate specifications including ToT revisions, the impact of US growth revisions remains significant.
  - Individual-country regressions: distribution of slope coefficients shifts upwards in the last horizon, consistent with panel results.
  - Seasonal-horizon robustness: spring-to-spring results reproduce main conclusions; semi-annual results show stronger impact in Fall-to-Spring revisions than in Spring-to-Fall revisions, suggesting forecast teams better reflect systemic economies’ revisions when updating in Spring.
- Asymmetric effects and notable drivers:
  - Negative US revisions generally have a larger impact on other economies’ growth revisions than positive US revisions.
  - Larger US revisions (in absolute value) have a stronger impact.
  - The year 2009 (peak of the GFC) is a key driver: excluding 2009 makes results for five country groups insignificant.
- Magnitude example:
  - An upward (downward) revision of growth for the US is typically associated with an upward (downward) revision of growth for countries in all major country groups; the estimated coefficient is between 0.2 and 1 (see upper panel in Table 2).

### VII. ROLE OF TERMS-OF-TRADE (ToT) REVISIONS
- Hypothesis: ToT revisions affect growth revisions (upward ToT revision → higher income → upward growth revision).
- Findings:
  - For most time horizons, impact from ToT revisions is generally positive but weak and often of limited significance.
  - ToT revisions play a weaker role relative to systemic economies’ growth revisions.
  - Latest horizon (Table 4): impact of ToT revisions is significant and important for LAC, and positive for most other EM groups.
  - An observed negative coefficient for Emerging Europe may reflect correlation with Advanced Europe growth revisions; ToT coefficient for Emerging Europe becomes insignificant when controlling for Advanced Europe growth.

### VIII. ALTERNATIVE MEASURES OF SYSTEMIC GROWTH REVISIONS
- Alternative benchmarks considered: China, Euro Area, G7, World aggregate, in addition to US.
- Findings:
  - Impact of systemic economies is robust to alternative measures, remaining positive and strongly significant in almost all specifications.
  - Caveat: use of G7, Euro Area, World aggregate can overstate impact because some economies appear on both sides of the regression; regressions using US or China avoid this overlap by excluding these countries from LHS.
  - Joint inclusion of US and China: both matter for individual economies’ growth revisions, but incremental gain from including both relative to separately is small.
- Pairwise correlations of systemic-economy growth revisions (period 1990-2019): most pairs show high correlation except China-US and China-G7 pairs which show very low correlations (correlations below 0.15 shaded in source).

### IX. WEO VS CONSENSUS FORECAST REVISIONS
- Specification: RREV_Y(t-i,t) = β0 + β1 RREV_CF_Y(t-i,t) + ε_t.
- Findings:
  - WEO and Consensus Forecast annual growth revisions move very closely together for both Fall-to-Fall and Spring-to-Spring revisions across country groups and regions with reasonable data coverage.
  - Semi-annual comparisons:
    - Fall-to-Spring revision (S0-F1) correlation is about 1.
    - Spring-to-Fall current-year revision (F0-S0) correlation is about 0.8.
    - Revisions between Spring and Fall of the previous year (F1-F2) correlation is about 0.6.
  - Data limitations: MENAP and SSA results based on 3 and 3 countries only, respectively; those group results should be ignored.

### X. MEAN REVERSION, AUTOCORRELATION, AND WITHIN-VINTAGE PERSISTENCE
- Mean reversion and learning:
  - Specification linking forecast error to next-year revision shows a negative effect for the following year and null effect for the medium term: discovery of an overestimation correlates with a negative next-year revision; learning has virtually no impact on medium-term forecast revision.
- Serial autocorrelation across vintages:
  - Revisions do not appear serially autocorrelated across vintages (small coefficients, sometimes significant about half the time).
  - Implication: revisions for a given vintage are not systematically predictable from previous vintages — pointing toward efficiency of revisions.
- Revisions for a given year across vintages and horizons:
  - Earlier horizons: coefficients negative — revisions done several years ago tend to be updated with negative correlation over time (back-and-forth pattern).
  - Latest horizons: some groups (AE, LAC, MENAP) show positive coefficients — revisions sequenced in same direction (possibly due to more confidence or better information).
  - Practical implication: if forecasters revise growth for 2023 down in 2021, in 2022 they are more likely to revise growth for 2023 up, indicating potential inefficiency and opportunity to improve forecast efficiency.
- Within-vintage persistence:
  - Shocks are generally perceived as persistent within vintages, especially for LAC and Emerging Asia, and for closest time horizons.
  - For all country groups except MENAP, positive coefficients imply positive growth revisions for the current year are associated with positive growth revisions for the next 1 or 2 years within the same vintage.

### XI. MAIN EMPIRICAL FINDINGS, INTERPRETATION, AND IMPLICATIONS
- Summary of empirical patterns:
  - Growth revisions closer to the actual are larger, more volatile, and more negative; most revision activity happens for next year.
  - Negative bias for next-year revisions implies forecasters enter latest horizons with generally optimistic growth forecasts and subsequently revise downwards.
  - WEO growth revisions are, on average, in the right direction—revisions tend to narrow prior-period forecast errors at any horizon and for any country region.
  - Revisions become progressively more responsive to the forecast error gap as revision horizons get closer to the actual, indicating learning and improved accuracy nearer the target year.
  - Growth revisions in systemic economies are relevant for growth revisions in all regions, reflecting trade and financial linkages; impact mainly concentrated in the horizon closest to the actual.
  - Large and negative systemic revisions have stronger cross-country impacts than smaller or positive revisions; the GFC year 2009 is a key driver amplifying these effects for LICs and some EMDE groups.
  - ToT revisions matter mainly for LAC and in the last horizon; overall ToT role is secondary to systemic growth revisions.
  - Strong co-movement between WEO and Consensus Forecast revisions suggests shared information or common drivers across forecasting institutions.
  - Fall-to-spring WEO revisions are more correlated with corresponding Consensus Forecast revisions than spring-to-fall pairs, indicating timing/seasonality effects in information incorporation.
  - Revisions for a given time horizon are not autocorrelated across vintages, supporting revision efficiency; within vintages, revisions tend to be positively correlated, indicating perceived short-term persistence of shocks.
- Practical implications:
  - Forecasters use cross-country linkages more for next-year forecasts and rely more on country-specific factors for medium-term forecasts.
  - Asymmetric downside sensitivity suggests downside shocks drive cross-country revision dynamics; attention to large negative systemic shocks is important for cross-country forecasting.
  - Opportunities exist to improve forecast efficiency by better accounting for persistent within-vintage signals and by understanding sequencing patterns across vintages.

### XII. LIMITATIONS AND DIRECTIONS FOR FUTURE RESEARCH
- Limitations noted:
  - Role of China’s and other systemic economies’ forecast revisions requires further disentangling.
  - Some regional results have limited sample sizes (e.g., MENAP and SSA for Consensus comparisons).
  - Overlap issues when using aggregates (G7, Euro Area, World) can overstate impacts because economies may appear on both sides of regressions.
- Suggested future work:
  - Investigate differential effects in the case of IMF program and surveillance cases.
  - Further inspect the extent and nature of cross-country revision dependence and the channels (trade, financial, other).
  - Explore improved use of within-vintage persistence signals to enhance forecast efficiency.

*Source: WP/21/136, "Patterns in IMF Growth Forecast Revisions: A Panel Study at Multiple Horizons", Metodij Hadzi-Vaskov, Luca Antonio Ricci, Alejandro Mariano Werner, and Rene Zamarripa (May 2021).*

### Section 1

### I. INTRODUCTION

### Abstract and Key Findings
- Paper: "Patterns in IMF Growth Forecast Revisions: A Panel Study at Multiple Horizons" (WP/21/136).
- Main empirical findings summarized in the Abstract:
  - (i) Growth revisions in horizons closer to the actual are generally larger, more volatile, and more negative.
  - (ii) On average, growth revisions are in the right direction, becoming progressively more responsive to the forecast error gap as horizons get closer to the actual year.
  - (iii) Growth revisions in systemic economies are relevant for growth revisions in all country groups.
  - (iv) WEO and Consensus Forecast growth revisions are highly correlated.
  - (v) Fall-to-spring WEO revisions are more correlated with Consensus Forecasts revisions compared to spring-to-fall revisions.
  - (vi) Across vintages, revisions for a given time horizon are not autocorrelated; within vintages, revisions tend to be positively correlated, suggesting perception of persistent short-term shocks.

### Research questions addressed
- Are WEO growth revisions in the right direction?
- What are the key drivers explaining WEO growth revisions?
- How do WEO revisions compare to Consensus Forecast revisions?
- Is there mean reversion/persistence in WEO growth forecasts?
- How do results vary across time-horizons and country groupings?

### Role and motivation
- Focus on forecast revisions (rather than forecast errors) to gain dynamic insights into the forecasting process and potential efficiency improvements.
- WEO database: biannual (Spring and Fall) WEO vintages over 1990–2019 for 194 economies; forecasts for year t consist of 12 vintages from year t-5 to year t, covering forecasts until 2024.

---

### II. MOTIVATION AND LITERATURE

### Literature dimensions summarized
- Three recurrent dimensions in existing literature:
  - Forecast accuracy and systematic forecast bias.
  - Forecast performance relative to comparators (e.g., Consensus Forecast).
  - Incorporation of available information to improve forecast efficiency.

### Selected literature points reported
- Timmermann (2006, 2007): WEO real GDP growth forecasts display systematic overprediction and could be improved by accounting for cross-country linkages.
- Genberg and Martinez (2014): WEO forecasts not consistently biased except in specific recessions; performance similar to comparators; efficiency gains possible by considering systemic economies.
- Celasun et al. (forthcoming): Absolute forecast accuracy improved over 2004-2017 vs 1990-03; potential improvement for about 10-15 percent of economies by taking into account growth forecasts of systemic economies and terms of trade.
- Ismail, Perrelli, and Yang (2020): Large planned fiscal and external adjustments associated with optimistic WEO growth projections with significant non-linearities.
- Expectation-formation literature: evidence of sticky information and perceived persistence (Coibion and Gorodnichenko (2015); Branch (2007); Bordalo, Gennaioli, and Schleifer (2018); Bluedorn and Leigh (2018)).

---

### III. DATA AND METHODOLOGY

### Dataset
- Source: WEO growth and ToT forecasts from Celasun et al. (forthcoming).
- Coverage: Spring and Fall WEO vintages over the period 1990-2019 for 194 economies.
- For forecasts of year t: 12 vintages from year t-5 to year t (Spring and Fall for each of these six years) — dataset covers forecasts until 2024.
- Outturns: WEO growth outturns collected from Spring and Fall WEO vintages; outturn of year t collected as measured from year t+1 to year t+5. Main analysis focuses on outturns measured at t+1; robustness shown for outturns measured at t+5.
- Consensus Forecasts: growth forecasts from Consensus Forecast collected from March and September vintages to match Spring and Fall WEO vintages.

### Definitions and notation
- Forecast revision (REV) of any variable X for a given country and year t:
  - For every i = 0,1,...,4 and t = 1990:2023, REV defined as forecast F made in (t-i) minus forecast made in (t-i-1).
- "i" denotes the time horizon (how many years ahead the forecast is made).
- "Vintage" denotes the year when forecasts or revisions were made (first subscript); second subscript denotes the year for which the forecast was made.
- Primary focus often on F0-F1 (F(t-0,t) − F(t-1,t)), i.e., revisions to forecasts for the current year.

### Visual illustration (described)
- Figure 1: Fall forecasts over 1990-2024 for US; solid black line: actual outcomes measured at t+1; grey line: outcomes measured at t+5; earliest horizons yellow circles (t-5) to darkest blue circles (t-0). Revisions captured by distances between circles.
- Figure 2: Full set of revision series (fall-to-fall) with actual outcomes; lines darken as horizons approach actual outcome; red line references actual outturns.

### Regression specifications
- Five core regression specifications (panel regressions run for each horizon i with RE and FE; Hausman tests used to choose preferred method). Notation: Y_t is percentage change of real GDP at time t.

1) Directional test (are revisions in the right direction?)
   - RREV_Y(t-i,t) = β0 + β1 FR_Y(t-i-1,t) + ε_t
   - FR_Y(t-i-1,t) is the forecast error: F(t-i-1,t) − A(t+1,t) where A(t+1,t) is actual outturn for year t reported at t+1.

2) Impact of systemic economies’ GDP revisions
   - RREV_Y(t-i,t) = β0 + β1 RREV_X1(t-i,t) + ε_t
   - X1 is percentage change of real GDP for a systemic economy (US, China, Euro Area, G7, or World).

3) Impact of terms-of-trade revisions
   - RREV_Y(t-i,t) = β0 + β1 RREV_X2(t-i,t) + ε_t
   - X2 is percentage change in WEO terms-of-trade.

4) Joint impact of systemic GDP and ToT revisions
   - RREV_Y(t-i,t) = β0 + β1 RREV_X1(t-i,t) + β2 RREV_X2(t-i,t) + ε_t

5) WEO versus Consensus Forecast revisions
   - RREV_Y(t-i,t) = β0 + β1 RREV_CF_Y(t-i,t) + ε_t
   - RREV_CF_Y denotes corresponding revision from Consensus Forecast.

- Panel regressions cover full sample and country-group subsamples. Coefficients from RE and FE generally quite similar.

---

### IV. DESCRIPTIVE FINDINGS

### Stylized facts from descriptive analysis
- Growth revisions closer to the actual vintage:
  - Are generally larger.
  - More volatile.
  - More negative on average.
- Distributional extremes:
  - World growth revisions in the latest horizon (F0-F1) sometimes range from -70 percentage points to +80 percentage points (reflecting exceptional events).
- Visual analysis subsequently zooms into the core segment of distribution: revisions within -10 to +10 percentage points.

---

*Source: WP/21/136, "Patterns in IMF Growth Forecast Revisions: A Panel Study at Multiple Horizons", Metodij Hadzi-Vaskov, Luca Antonio Ricci, Alejandro Mariano Werner, and Rene Zamarripa (May 2021).*

### Section 2

### Section 2 — Visual and Econometric Evidence on Growth Revisions

### Visual evidence on growth revisions
- Growth revisions closest to the actual (F0-F1) have often been negative, especially during the GFC, and across income groups.
- Across EMDE regions, growth revisions are more likely to be negative in general; large negative adjustments are visible during:
  - the transition in Eastern Europe in the early 1990s,
  - the Asian Crisis in the late 1990s,
  - the Arab Spring in the 2010s.
- For Emerging and Developing Europe, the 25th percentile of the distribution of growth revisions is -28 ppts in 1991 and -11 ppts in 2009.
- Across time horizons, most action takes place during the latest revision (between the Fall WEO in the year for which the forecast is made and the Fall WEO of the previous year); earlier-horizon revisions are considerably more stable.
- Summary finding (Figure 7): most recent growth revisions are generally larger, more volatile, and more negative.

### Are revisions in the right direction? (Specification (1))
- Panel regressions estimate: RY_{t−i,t} = β0 + β1 FRY_{t−i−1,t} + ε_t.
- Key interpretation: a negative slope coefficient β1 implies revisions narrow the previous period’s forecast error (when prior forecast error FRY_{t−i−1,t} > 0, then RY_{t−i,t} < 0).
- Empirical findings:
  - Estimated slope coefficients β1 are negative for all revision horizons and all country groups, implying on average revisions move forecasts toward the actual (revisions are generally in the right direction).
  - The absolute value of β1 increases (coefficients become more negative) for vintages closer to the actual, implying stronger corrections for forecasts made closer to the actual year.
- Significance notation used: black dots denote significance at 10 percent level; red dots indicate lack thereof.

### Robustness to outturn updates
- Baseline forecast errors used actual ex-post growth outturns for year t reported at t+1.
- Robustness check uses outturns reported at t+5 (instead of t+1).
- Findings using t+5 outturns (Figures 10 and 11):
  - The negative relationship between prior forecast errors and subsequent revisions is robust when using outturns revised at t+5 (albeit with fewer observations).
  - Relationship weakens somewhat for EMDEs: for EMDEs it is easier to forecast the actual as it will be measured at t+1 than at t+5, suggesting subsequent outturn revisions incorporate factors not visible in real time.

### What factors drive growth revisions?

Role of systemic economies’ GDP revisions
- Hypothesis: growth revisions for systemic economies (e.g., US) affect growth revisions for other economies via trade and financial linkages.
- Descriptive example: relationship between US growth revisions and LA5 economies:
  - Impact of US revisions on LA5 increases as the forecast vintage gets closer to the actual (relationship steepens for later vintages; darker lines/points depict later vintages).
- Formal regressions (univariate and multivariate specifications):
  - Equations evaluated include RY_{t−i,t} = β0 + β1 RXX_{t−i,t}^1 + ε_t and multivariate extensions (including additional systemic revisions and ToT).
  - Findings (Figure 13 and related analysis):
    - Growth revisions in the US influence growth revisions across the world: for advanced economies across all time horizons; and for other countries in later horizons.
    - Impact is larger for revisions closest to the actual (higher β1 in the latest horizons).
    - Results from univariate (dotted) and multivariate (solid, including ToT) regressions show a similar picture.
- EMDE regions (Figure 14):
  - US growth revisions are relevant for growth revisions in countries from all EMDE regions.
  - Impact is larger for revisions closer to the actual (steeper lines across horizons) for both univariate and multivariate specifications.
- Latest revision horizon (Table 2):
  - In univariate specifications, the impact of US growth revisions is significant for all country groups and regions.
  - In multivariate specifications including ToT revisions, the impact of US growth revisions remains significant for all country groups and regions.
- Individual-country regressions (Figure 15):
  - Distribution of estimated slope coefficients from individual-country regressions shifts upwards in the last horizon, consistent with panel results.
- Seasonal horizon robustness:
  - Spring-to-Spring regressions reproduce the main conclusions.
  - Semi-annual revisions (Spring-to-Fall and Fall-to-Spring) are robust but reveal a puzzling pattern: in the latest two years, impact is especially important for Fall-to-Spring revisions, while Spring-to-Fall revisions show almost no impact (oscillating pattern in last horizons). This implies forecast teams better reflect systemic economies’ revisions when updating forecasts in Spring rather than in Fall.

Asymmetric impact of systemic GDP revisions
- Table 3 differentiates impacts by:
  - positive vs negative US revisions,
  - large vs small US revisions,
  - sample with vs without the peak of the GFC (year 2009).
- Findings:
  - Negative US revisions generally have a larger impact on other economies’ growth revisions than positive US revisions.
  - Relatively larger US revisions (in absolute value) have a stronger impact on growth revisions for other economies.
  - The year 2009 (peak of the GFC) is a key driver: excluding 2009 makes results for five country groups insignificant.

Role of Terms of Trade (ToT) revisions
- Hypothesis: ToT revisions could affect growth revisions (an upward ToT revision could imply higher income and upward growth revision).
- Regression framework includes ToT (percent change) revisions as controls in multivariate specifications.
- Descriptive evidence (Figure 18, LA5 example):
  - Relationship between ToT revisions and growth revisions steepens for horizons closer to the actual (darker lines), suggesting ToT revisions generally have a positive impact on growth revisions in LA5, at least in the latest horizon.

*Source: WEO database and authors’ calculations.*

### Section 3

### wpiea2021136-print-pdf - Section 3

### Impact of Terms of Trade (ToT) Revisions on Growth Revisions
- Definitions and methodology:
  - Large (small) deviations defined as revisions that are, in absolute value, above (below) one standard deviation from the sample mean.
  - Univariate specification (3): RRRR Y_{t−i,t} = β0 + β1 RRRR X_{t−i,t}^2 + ε_t.
  - Multivariate specification (4): RRRR Y_{t−i,t} = β0 + β1 RRRR X_{t−i,t}^1 + β2 RRRR X_{t−i,t}^2 + ε_t, where RRRR X^1 are systemic economies’ growth revisions.
- Empirical findings:
  - For most time horizons, the impact from ToT revisions is generally positive but weakly associated with growth revisions, with limited significance.
  - ToT revisions point towards a weaker role in affecting growth revisions relative to the revisions to systemic economies’ growth.
  - Table 4 (latest horizon) shows the impact of ToT revisions is significant and important for LAC, and positive for most other EM groups.
  - An observed negative coefficient for Emerging Europe may be due to ToT being negatively influenced by a positive revision to Advanced Europe growth; the ToT coefficient for Emerging Europe becomes insignificant when controlling for Advanced Europe growth.
- Interpretation:
  - ToT revisions have some explanatory power but are generally secondary to systemic growth revisions in explaining individual economies’ growth revisions.

### Role of Alternative Measures of Systemic Growth Revisions
- Alternative systemic benchmarks evaluated: China, Euro Area, G7, World aggregate, in addition to US.
- Findings:
  - The impact of systemic economies is robust to alternative measures, remaining positive and strongly significant in almost all specifications.
  - Caveat: The impact of G7, Euro Area and the World aggregate is overstated in most specifications because some economies appear on both sides of the regression equation.
  - Regressions using US or China for systemic economy revisions are immune to this overlap issue as they exclude these countries from the left-hand side.
- Joint effects and correlations:
  - Table 6 reports pairwise correlations of growth revisions in systemic economies (period 1990-2019); most pairs show high correlation except China-US and China-G7 pairs which show very low correlations (correlations below 0.15 are shaded in grey in the source).
  - Table 7 (latest vintage) explores joint inclusion of China and US growth revisions: both US and China growth revisions are relevant for individual economies’ growth revisions, but the “incremental gain” from including them jointly relative to separately is rather small.
- Note on significance convention:
  - Black dots denote significance at 10 percent level; red dots indicate lack thereof.

### WEO and Consensus Forecast Growth Revisions Moving Together?
- Regression specification used: RRRR Y_{t−i,t} = β0 + β1 RRRR_CF Y_{t−i,t} + ε_t (equation (5)); focus on slope coefficient β1 capturing co-movement.
- Findings (Table 8 and Table 9):
  - WEO and Consensus Forecast annual growth revisions move very closely together for both Fall-to-Fall and Spring-to-Spring revisions across all country groups and regions with reasonable data coverage.
  - Semi-annual comparisons (up to 3 vintages of semi-annual revisions) also show strong co-movement:
    - The Fall-to-Spring revision (S0-F1) has a correlation of about 1.
    - The Spring-to-Fall current-year revision (F0-S0) correlation is about 0.8.
    - Revisions between Spring and Fall of the previous year (F1-F2) have a smaller correlation of about 0.6.
- Data limitations:
  - MENAP and SSA results are based on 3 and 3 countries only, respectively, hence results for these groups are to be ignored.

### Mean Reversion and Persistence in WEO Growth Forecasts
- Relationship between forecast errors and forthcoming revisions:
  - Specification (6): RRRR Y_{t+1,t+ j} = β0 + β1 FRR Y_{t,t} + ε_t; where Forecast Error now defined as (F_Y(t,t) − A_Y(t+1,t)).
  - Results (Figure 20):
    - A negative effect for the following year (closest horizon).
    - A null effect for the medium term.
    - Interpretation: discovery of an overestimation (under) correlates with a negative (positive) growth forecast revision for the next year; learning about a forecast mistake has virtually no impact on medium-term forecast revision.
- Serial autocorrelation of revisions across vintages:
  - Specification (7): RRRR Y_{t−i,t} = β0 + β1 RRRR Y_{t−i−1,t−1} + ε_t; ∀ i = 0,1,2,3,4.
  - Results (Figures 21–22):
    - Revisions do not appear to be serially autocorrelated (small coefficients, positive and negative, significant only about half of the time).
    - Forecast revisions are not persistent across vintages; the revision done this year for forecasting this year is not systematically related to the revision done last year for forecasting last year.
- Relationship of revisions for a given year across vintages and horizons:
  - Specification (8): RRRR Y_{t−i,t} = β0 + β1 RRRR Y_{t−i−1,t} + ε_t; ∀ i = 0,1,2,3,4.
  - Results (Figures 23–24):
    - For earlier horizons, coefficients are negative: revisions to forecasts for a given year done several years ago tend to be updated with negative correlation over time (a back-and-forth pattern).
    - For latest horizons, some groups (AE, LAC, MENAP) show positive coefficients, suggesting revisions are sequenced in the same direction (possibly due to more confidence or better information).
    - Practical implication: if forecasters revise growth for 2023 down in 2021, in 2022 they are more likely to revise growth for 2023 up, indicating opportunity to improve forecast efficiency.
- Persistence of growth shocks within forecast vintages:
  - Specification (9): RRRR Y_{t,t+ j} = β0 + β1 RRRR Y_{t,t} + ε_t; ∀ j = 1,2,3,4.
  - Results (Figures 25–26):
    - Shocks are generally perceived as persistent, especially for LAC and Emerging Asia, and for closest time horizons.
    - For all country groups except MENAP, positive coefficients imply that positive growth revisions for the current year are associated with positive growth revisions for the next 1 or 2 years.
    - Practical interpretation: if forecasters revise growth for 2021 up this year, they are likely to revise growth for 2022 up as well within the same vintage.

### Concluding Remarks (Summary of Key Findings)
- Descriptive patterns:
  - Growth revisions in horizons closer to the actual are generally larger, more volatile, and more negative; most forecast revision activity happens for next year.
  - The negative bias for next-year revisions implies forecasters enter the latest horizons with generally optimistic growth forecasts and subsequently revise downwards.
- Forecast responsiveness and learning:
  - WEO growth revisions are, on average, in the right direction—revisions tend to narrow the forecast error from the previous period at any horizon and for any country region.
  - Forecast revisions become progressively more responsive to the forecast error gap as revision horizons get closer to the actual, indicating learning and improved accuracy nearer the target year.
- Cross-country linkages and systemic economies:
  - Growth revisions in systemic economies are relevant for growth revisions in all regions, reflecting cross-country trade and financial linkages.
  - Example magnitude: an upward (downward) revision of growth for the US is typically associated with an upward (downward) revision of growth for countries in all major country groups (the estimated coefficient is between 0.2 and 1—see upper panel in Table 2).
  - Revisions in systemic economies are mainly relevant for the forecast horizon closest to the actual; forecasters use cross-country linkages more for next-year forecasts and rely more on individual country characteristics for medium-term forecasts.

*Source: wpiea2021136-print-pdf - Section 3*

### Section 4

### Section 4

### Main empirical findings on growth forecast revisions
- Large and negative revisions have a stronger impact than smaller and/or positive revisions.
- Revisions at the peak of the GFC (year 2009) has been a key driver of these results, especially for LICs and some EMDE groups.
- The relevance of ToT revisions for growth revisions is mainly robust for LAC and again in the last horizon.
- WEO & Consensus Forecast growth revisions are strongly correlated, both at annual and semi-annual horizons, suggesting a commonality in the movement of these forecasts.
- The fall-to-spring WEO revisions are more correlated with the corresponding fall-to-spring Consensus Forecasts revisions compared to spring-to-fall revision pairs.
- If projections were optimistic this year in forecasting growth for this year, then next year forecasters will revise down forecasts for next year (forecasters “act upon mistakes”).
- Revisions for a given time horizon are not autocorrelated across vintages, suggesting that forecast cannot be improved upon by looking at past vintages, a result that points toward efficiency of revisions.
- Within vintages, revisions tend to be positively correlated—positive (negative) growth revisions for the current year are associated with positive (negative) growth revisions for the next 1 or 2 years—suggesting that forecasters generally perceive shocks to have short-term persistence.

### Interpretation and implications
- The asymmetric impact of negative versus positive revisions implies downside shocks drive cross-country revision dynamics, with the GFC year 2009 amplifying these effects for LICs and some EMDEs.
- Term-of-trade (ToT) revision relevance varies by region and horizon, being most salient for LAC and at longer horizons noted as “the last horizon.”
- Strong correlation between WEO and Consensus Forecast revisions indicates shared information or common drivers across forecasting institutions and vintages.
- Greater correlation in fall-to-spring pairs than in spring-to-fall pairs suggests seasonal or timing effects in how information is incorporated across forecasts.
- Lack of autocorrelation across vintages supports the notion that vintages efficiently incorporate available information; positive within-vintage correlation indicates expected short-lived persistence of shocks.

### Limitations noted and directions for future research
- The role of China’s and other systemic economies’ forecast revisions is left for future research; disentangling the effect of these economies is not addressed here.
- Future work could fruitfully investigate the differential effect in the case of the IMF program and surveillance cases (similar to Kareem, Perelli, and Yang (2020)).
- Further inspection of the extent of cross-country revision dependence is suggested.

*Source: wpiea2021136-print-pdf - Section 4*

---


_Source: https://www.imf.org/-/media/files/publications/wp/2021/english/wpiea2021136-print-pdf.pdf_