## _wp0659

## Source details

**Canonical URL:** [_wp0659](https://www.imf.org/-/media/websites/imf/imported-full-text-pdf/external/pubs/ft/wp/2006/_wp0659.pdf)

## Other formats

- [Markdown version](/-/media/websites/imf/imported-full-text-pdf/external/pubs/ft/wp/2006/_wp0659.pdf.md)
- [Structured JSON version](/-/media/websites/imf/imported-full-text-pdf/external/pubs/ft/wp/2006/_wp0659.pdf.json)

---

### Overview and scope
- Analysis period: 1990 to 2003.
- Variables analyzed: real GDP growth, inflation, the current account balance (in percent of GDP), and import and export volume growth.
- Coverage: 178 countries in seven regions (Africa; Central and Eastern Europe; Commonwealth of Independent States (CIS) countries and Mongolia; developing Asia; the Middle East; the Western Hemisphere; and the advanced economies).
- Forecast vintages examined: current-year and next-year forecasts prepared in April and September; Executive Board versions presented in February and July and subsequently revised before publication.
- Comparison set: Consensus forecasts for 1990–2003 covering the G-7 economies, seven Latin American economies (Argentina, Brazil, Chile, Colombia, Mexico, Peru, and Venezuela) and nine Asian economies (China, Hong Kong SAR, India, Indonesia, Korea, Malaysia, Singapore, Taiwan Province of China, and Thailand).

### Key findings — Accuracy by variable
- General: WEO forecasts meet basic forecasting-quality standards in some dimensions; notable issues exist and are variable-specific.
- Real GDP growth
  - Systematic tendency for overprediction (predicted growth on average tends to exceed actual growth).
  - Biases are most statistically significant in next-year forecasts.
  - Overpredictions particularly prevalent in forecasts for countries with an IMF program.
  - Overprediction tendency is persistent over time.
  - Evidence of spillovers: forecasts of U.S. GDP growth are positively and significantly correlated with current-year forecast errors of output growth in many advanced economies; German GDP forecasts have predictive power over output growth forecast errors in some regions.
  - The WEO assumption that the output gap is eliminated after five years may contribute to accuracy problems; a predominant negative relationship is found between the output gap and the forecast error in GDP growth, notably for Germany, France, and Italy.
- Inflation
  - Bias toward underprediction, with biases significant in next-year forecasts for many African, Central and Eastern European, and Western Hemisphere countries.
  - Underprediction bias generally weaker in current-year forecasts.
  - Next-year inflation forecast errors are often linked to U.S. GDP forecasts.
- External current account balances (percent of GDP)
  - Fewer overall problems relative to GDP and inflation forecasts.
  - April next-year forecast errors in some cases are significantly biased or serially correlated.
  - No clear general pattern in direction of biases.

### Directional accuracy and recession performance
- Directional accuracy
  - WEO forecasts are quite successful in predicting directional change for current-year real GDP growth and inflation.
  - Directional accuracy is somewhat weaker for next-year forecasts.
  - Hit rates example for Advanced Economies: next-year GDP 62 percent; April current-year 78 percent; September current-year 90 percent.
- Performance during the 2001 downturn
  - WEO GDP forecasts generally overpredicted growth in 2001 in all regions.
  - For 2002, the April and September next-year WEO forecasts prepared in 2001 overpredicted growth in six of the seven regions.
  - Revisions in the April 2002 WEO greatly reduced forecast errors in four regions.
- Recent snapshot (2001–2003)
  - U.S. GDP growth was 0.25 percent in 2001.
  - 2001 actual regional growth examples: CIS and Mongolia ~8 percent; Africa 4.3 percent; Middle East 3.5 percent.
  - WEO overpredicted 2001 growth in all regions except CIS and Mongolia.
  - Example Advanced Economies: September 2000 forecast error for 2001 growth = 2.17 percent → reduced to 1.32 percent (April 2001) → 0.35 percent (September 2001).
  - 2002 recovery: April and September previous-year forecasts overpredicted growth in six of seven regions; April 2002 current-year forecasts much closer (within 0.4 percent in four regions).
  - 2003 forecasts: very precise for Africa, Central and Eastern Europe, and Developing Asia (forecast errors within 0.7 percent).
  - Inflation for 2001: Africa previous-year April forecasts underpredicted widely (2001: 12.6 percent; 2002: 5.3 percent); Advanced Economies inflation forecasts within 0.45 percent of actuals for 2003.
  - Current account forecasts tended to underpredict during 2001–2003 for CIS and Mongolia, Developing Asia, Middle East, and Advanced Economies; errors modest and shrink as horizon shortens.

### Informational value of revisions (Board → Published)
- Revisions from Executive Board to published forecasts add considerable informational value.
- For same-year forecasts (Advanced Economies):
  - February/April reduction in forecast error: around one-fifth for the advanced economies.
  - July/September reduction in forecast error: nearly 30 percent for the advanced economies.
  - Regression evidence: Board-to-Publication revision correlated with forecast error for ~50 percent of Advanced Economies; revision coefficient positive for 80–90 percent of countries; average R2 ≈ 0.25 for current-year regressions.
- For next-year forecasts:
  - Reduction in forecast error from revisions: only 5 percent (Advanced Economies).
  - Next-year regressions: lower significance; average R2 ≈ 0.10; ~60 percent of β-estimates positive; ~10 percent β-estimates > 2.
- Other regions:
  - Lower fraction of positive and significant β coefficients (≈10–15 percent current-year; 5–10 percent next-year).
  - MSE ratio (publication/Board) tends to be lower than for Advanced Economies for some regions, especially for next-year forecasts.

### Statistical-significance and small-sample inference (bootstrap results)
- Bootstrap: 5,000 experiments to construct p-values and percent-significant measures addressing small-sample issues.
- GDP forecasts (selected bootstrap findings)
  - April current-year: ~40 percent of African countries’ forecasts systematically too large (overpredictions); bootstrapped tests confirm significant bias for close to 30 percent of African countries; joint bias+serial-correlation test reduces to 15 percent.
  - CIS & Mongolia and Western Hemisphere: 10–20 percent of countries show significant bias.
  - Middle East: serial correlation more important (15 percent significant).
  - September current-year: proportion of biased cases generally lower than April.
  - Next-year forecasts: significant upward bias in ~35 percent of African countries and ~30 percent of Western Hemisphere countries; >20 percent of Advanced Economies affected; serial correlation pervasive across regions.
- IMF program effect on GDP forecasts
  - Data: 350 program events; after discarding extreme outliers, between 850 and 960 program-year observations analyzed.
  - Program countries show systematic overpredictions (negative mean forecast errors); bias smallest for September current-year, largest for April next-year.
  - 60 percent of next-year forecast errors are negative for program countries.
- Inflation bootstrap findings
  - Mild inefficiency in current-year inflation forecasts (underprediction) in Africa and Western Hemisphere.
  - Strong evidence against efficiency in next-year inflation forecasts (systematic downward bias) in most regions except Advanced Economies.
  - Serial correlation sizable in Advanced Economies next-year inflation errors (rejection rates ~15 percent).
- IMF program effect on inflation forecasts
  - Program countries show systematic underprediction of inflation; largest bias is 4.5 percent for April next-year forecasts; mean bias 0.9 percent for September current-year predictions.
- Trade volumes and current account
  - Little evidence of systematic inefficiencies for trade volumes.
  - Next-year current account forecasts show some bias and serial correlation—rejection rates up to two to four times expected values (example: 24 percent in April next-year forecasts for Western Hemisphere or Developing Asia).

### Predictability tests — external predictors and output-gap diagnostics
- Predictors tested: WEO predicted U.S. GDP growth, WEO predicted German output growth, WEO forecast of oil prices, global current account discrepancy.
- Key predictor results (proportions of significant t-values):
  - U.S. GDP forecast: predictive power for current-year GDP forecast errors in Advanced Economies (31 percent of April current-year and 24 percent of September current-year cases with t-value above two); no evidence for next-year forecast errors.
  - German output forecast: predictive in many CIS and Mongolia cases (especially next-year errors).
  - Oil prices: limited predictive power overall; notable predictive correlations in CIS and Mongolia (positive) and Western Hemisphere/Advanced Economies (negative).
  - Global current account discrepancy: significant for 40 percent and 24 percent of the Advanced Economies in explaining April current-year and next-year forecast errors, respectively.
- Inflation forecast predictability
  - Next-year inflation errors linked to U.S. GDP forecasts across multiple regions.
  - German forecast significant for inflation errors in CIS and Mongolia.
- Output gap tests (Advanced Economies)
  - WEO assumption: output gap eliminated after five years.
  - GDP forecast errors: regression t-values predominantly negative; around 15 percent of t-statistics exceed two in absolute value.
  - Large negative output-gap t-statistics for France (-1.63), Germany (-1.99), Italy (-2.30), Japan (-4.16) implying too-rapid gap closure may create upward bias.
  - Inflation forecast errors: output-gap coefficients predominantly positive (WEO tends to underpredict inflation the larger the output gap).
  - Current account errors: many instances of predictability from output gap with varying signs (examples include Hong Kong, Japan, Netherlands, Singapore, Sweden).

### Aggregate country/region scores and directional-accuracy diagnostics
- Country score: sum of three indicators (bias, serial correlation, predictability by instruments) where each indicator = 1 if |t-statistic| > 2.
- Findings:
  - Inflation forecasts (except Advanced Economies) generate highest average “problem scores” (above one in most regions).
  - Next-year forecasts generally generate higher average scores than current-year forecasts.
- Directional accuracy (Pesaran and Timmermann (1992) PT test)
  - GDP next-year hit rates: between 62 and 71 percent; PT test significant for ~20 percent of countries.
  - GDP current-year hit rates: April 68–78 percent; September 76–90 percent.
  - Inflation shows similar patterns; current-account directional accuracy generally lower than for inflation or output.

### Comparison with Consensus forecasts
- Period: 1990–2003; Consensus timing mostly March/September (Latin America February/August).
- Correlations: WEO and Consensus forecasts and forecast errors are strongly positively correlated in all countries and horizons (current-year correlations often exceed 0.96; weakest reported 0.75 for U.S. next-year level).
- Relative bias and RMSFE comparisons
  - Overall: WEO performance similar to Consensus.
  - Current-year: WEO current-year GDP forecasts for G-7 generally less biased than Consensus.
  - Next-year: WEO next-year bias stronger than Consensus across the board.
  - Timing matters: WEO current-year forecasts perform well against March Consensus and considerably better against February Consensus; WEO performs worse against Consensus reported in April (WEO’s longer gestation lag).
  - RMSFE ratios: current-year WEO GDP forecasts often outperform Consensus for G-7 and Latin America; Consensus often better for Asia at current-year horizon.
  - Informational efficiency regressions: in several cases Consensus forecasts help predict WEO errors (estimated |β| > 2 in specific country cases); reverse regressions also identify cases where WEO helps predict Consensus errors.
- Forecast combinations and weights
  - Granger-Ramanathan regressions estimate combination weights: many regressions generate large positive weights on WEO for GDP, with notable exceptions (Italy, Germany, US, Indonesia, Thailand).
  - For inflation: current-year WEO weights generally close to unity for G-7; next-year Consensus weights higher for some G-7.
  - Re-weighting experiments: moving WEO toward Consensus (equal-weight average a = −0.5) often improves next-year inflation forecasts; for G-7 current-year GDP modest gains from moving away from Consensus for some economies but gains modest overall.

### Long-run performance for G-7 countries and subsample findings
- Longer dataset (early seventies onward) key findings:
  - Forecast accuracy has deteriorated somewhat since prior evaluation.
  - WEO systematically and significantly overpredicted economic growth for European G-7 economies and Japan during 1991–2003.
  - U.S. growth was underpredicted after 1990 (bias not statistically significant).
  - Inflation was strongly and significantly overpredicted for Canada, France, Japan, and the United States during the 1990s and 2000s; inflation underpredicted for Italy.
  - Possible explanations: structural breaks (e.g., higher U.S. productivity) and forecasting assumptions (5-year output-gap elimination).
- G-7 diagnostic metrics (examples)
  - Full-sample current-year mean absolute forecast errors range from 0.64 (France) to 1.21 (Japan).
  - RMSFE-values vary from 0.90 (USA) to 1.53 (Japan).
  - WEO RMSFE about 40 percent of random-walk RMSFE and about 50 percent of recursive-mean RMSFE for current-year GDP forecasts; next-year ratios higher by 0.10–0.20.
  - Evidence of worsening relative performance in subsample splits: recursive mean sometimes outperforms WEO next-year forecasts for France, Germany, Italy in the second subsample.

### Recommendations (policy and operational)
- Timeliness of information is key
  - Systematic gains in forecasting accuracy from using the latest available information.
  - Updating process adds significant value, especially for the G-7 economies.
  - Recommendation: IMF country economists should update their projections just before publication.
- Continuous monitoring of forecasting performance
  - Empirical analysis indicates structural instability in variables, especially real GDP growth and inflation.
  - Recommendation: explore instituting real-time forecasting performance indicators and structural-stability monitoring (reverse-ordered Cusum tests, Pesaran and Timmermann (2002) type tests).
- Use bias-adjusted forecasts as guidance
  - Shrinking forecasts toward bias-corrected values can improve forecasts if bias can be precisely estimated.
  - Caveat: bias-adjustment may be too mechanical and assumes constant bias.
  - Recommendation: compare unadjusted forecasts with bias-adjusted forecasts to understand magnitude and direction of biases.
- Forecasts of risk
  - Given inherent uncertainty, strongly recommend incorporating advances in modeling and forecasting risk/uncertainty.
  - Example recommendation: present the full probability distribution of key variables over time (fan charts, two-piece normal, scenario simulations).
- Review the output gap assumption
  - WEO assumption that the output gap is removed within five years may contribute to overprediction of output growth in countries with large output gaps.
  - Recommendation: analyze costs and benefits of this practice and consider more frequent reviews of potential output estimates.

*Source: Author’s calculations.*

### 1. Sample Size, by Variable and Region.............................................................................48

### 1. Sample Size, by Variable and Region.............................................................................48

### Overview and scope
- Analysis period: 1990 to 2003 (last year for which actual data were available when the report was initiated).
- Variables analyzed: real GDP growth, inflation, the current account balance (in percent of GDP), and import and export volume growth.
- Coverage: 178 countries in seven economic regions (Africa; Central and Eastern Europe; Commonwealth of Independent States (CIS) countries and Mongolia; developing Asia; the Middle East; the Western Hemisphere; and the advanced economies).
- Forecast vintages examined: current-year and next-year forecasts prepared in April and September; Executive Board versions presented in February and July and subsequently revised before publication.
- Comparison set: Consensus forecasts for 1990–2003 covering the G-7 economies, seven Latin American economies (Argentina, Brazil, Chile, Colombia, Mexico, Peru, and Venezuela) and nine Asian economies (China, Hong Kong SAR, India, Indonesia, Korea, Malaysia, Singapore, Taiwan Province of China, and Thailand).

### Key findings — Accuracy by variable
- General: WEO forecasts for many variables and many countries meet basic forecasting quality standards in some dimensions; notable issues exist and are variable-specific.

- Real GDP growth
  - Systematic tendency for overprediction (predicted growth on average tends to exceed actual growth).
  - Biases are most statistically significant in next-year forecasts.
  - Systematic overpredictions are particularly prevalent in forecasts for countries with an IMF program.
  - Overprediction tendency is persistent over time.
  - Evidence suggests forecasts could be improved by paying more attention to international linkages:
    - Forecasts of U.S. GDP growth are positively and significantly correlated with current-year forecast errors of output growth in a substantial number of advanced economies.
    - Forecast of German GDP growth has predictive power over output growth forecast errors in some regions.
  - The WEO assumption that the output gap is eliminated after five years may contribute to accuracy problems; a predominant negative relationship is found between the output gap and the forecast error in GDP growth, notably for Germany, France, and Italy.

- Inflation
  - Bias toward underprediction, with these biases significant in next-year forecasts for many African, Central and Eastern European, and Western Hemisphere countries.
  - Underprediction bias generally weaker in current-year forecasts.
  - Next-year inflation forecast errors are often linked to U.S. GDP forecasts.

- External current account balances (in percent of GDP)
  - Fewer overall problems relative to GDP and inflation forecasts.
  - April next-year forecast errors, in some cases, are significantly biased or serially correlated.
  - No clear general pattern in direction of biases.

### Directional accuracy and recession performance
- Directional accuracy
  - WEO forecasts are quite successful in predicting directional change for current-year real GDP growth and inflation.
  - Directional accuracy is somewhat weaker for next-year forecasts.

- Performance during the 2001 downturn
  - WEO GDP forecasts generally overpredicted growth in 2001 in all regions.
  - For 2002, the April and September next-year WEO forecasts prepared in 2001 overpredicted growth in six of the seven regions.
  - Revisions in the April 2002 WEO greatly reduced forecast errors in four regions.

### Informational value of revisions
- Revisions from Executive Board to published forecasts add considerable informational value.
- For same-year forecasts:
  - February/April reduction in forecast error: around one-fifth for the advanced economies.
  - July/September reduction in forecast error: nearly 30 percent for the advanced economies.
- For next-year forecasts:
  - Reduction in forecast error from revisions: only 5 percent.

### Tabular overview of forecast accuracy (selected entries from Table 1)
- Real GDP Growth — Bias / Serial Correlation (Averages across regions for April current-year forecasts and September next-year forecasts; fraction of economies in region where statistics are significant in parenthesis)
  - Advanced Economies: -0.04/-0.36 (0.07/0.21) ; 0.21/0.24 (0.03/0.14)
  - Africa: -1.17/-1.48 (0.25/0.33) ; 0.21/0.03 (0.15/0.29)
  - Central and Eastern Europe: -1.17/-1.40 (0.07/0.07) ; 0.37/0.34 (0.13/.20)
  - Commonwealth of Independent States: -1.93/-2.39 (0.15/0.00) ; 0.31/0.46 (0.00/.15)
  - Developing Asia: -0.38/-0.53 (0.05/0.04) ; 0.25/0.31 (0.00/.13)
  - Middle East: -1.66/-1.34 (0.08/0.07) ; 0.37/0.31 (0.15/.21)
  - Western Hemisphere: -0.64/-1.16 (0.13/0.24) ; 0.23/0.21 (0.06/.21)

- CPI Inflation — Bias / Serial Correlation (same format)
  - Advanced Economies: -0.08/-0.13 (0.00/0.07) ; 0.20/0.33 (10.07/0.17)
  - Africa: 0.60/1.00 (0.13/0.16) ; 0.29/0.26 (0.06/0.14)
  - Central and Eastern Europe: 2.30/3.27 (0.13/0.16) ; 0.39/0.39 (0.33/0.33)
  - Commonwealth of Independent States: 126.95/153.65 (0.00/0.20) ; 0.69/0.74 (0.69/0.77)
  - Developing Asia: 1.06/0.62 (0.00/0.08) ; 0.33/0.36 (0.15/0.21)
  - Middle East: -0.86/-1.15 (0.00/0.08) ; 0.28/0.27 (0.15/0.00)
  - Western Hemisphere: 1.78/1.60 (0.13/0.27) ; 0.38/0.29 (0.22/0.15)

- Current Account Balance (percent of GDP) — Bias / Serial Correlation
  - Advanced Economies: 0.25/0.22 (0.07/0.11) ; 0.26/0.35 (0.07/0.29)
  - Africa: -0.21/-0.67 (0.02/0.09) ; 0.27/0.24 (0.13/0.09)
  - Central and Eastern Europe: 20.95/8.50 (0.00/0.00) ; 0.30/0.33 (0.07/0.13)
  - Commonwealth of Independent States: -7.49/1.51 (0.00/0.08) ; 0.24/0.39 (0.08/0.23)
  - Developing Asia: 0.58/1.11 (0.11/0.09) ; 0.28/0.34 (0.05/0.09)
  - Middle East: -1.36/-1.69 (0.00/0.00) ; 0.23/0.25 (0.09/0.00)
  - Western Hemisphere: -0.61/-.95 (0.03/0.16) ; 0.25/0.27 (0.10/0.16)

- Memorandum: G-7 Economies (selected entries)
  - Canada: Real GDP Growth -0.21/0.58 ;  -0.14/0.26 ; CPI Inflation 0.0/-0.20 ; -0.26/0.19 ; Current Account -0.01/0.35 ; 0.43/0.51
  - France: Real GDP Growth -0.20/-0.80 ; -0.38/-0.23 ; CPI Inflation 0.04/-0.05 ; 0.32/0.63 ; Current Account 0.14/0.08 ; 0.38/0.64
  - Germany: Real GDP Growth -0.13/-1.12 ; 0.05/-0.26 ; CPI Inflation 0.09/0.12 ; 0.21/0.37 ; Current Account -0.08/-0.35 ; 0.22/0.46
  - Italy: Real GDP Growth -0.43/-1.02 ; -0.20/-0.07 ; CPI Inflation 0.08/0.25 ; 0.08/0.18 ; Current Account -0.29/-0.25 ; -0.25/0.38
  - Japan: Real GDP Growth 0.01/-1.03 ; -0.01/0.23 ; CPI Inflation -0.07/-0.24 ; -0.19/-0.03 ; Current Account -0.02/-0.02 ; -0.06/0.01
  - United Kingdom: Real GDP Growth -0.02/-0.48 ; -0.60/0.46 ; CPI Inflation -0.07/-0.26 ; -0.09/-0.15 ; Current Account 0.33/0.51 ; -0.09/-0.04
  - United States: Real GDP Growth 0.17/0.23 ; 0.16/0.23 ; CPI Inflation 0.01/-0.30 ; -0.06/0.15 ; Current Account 0.01/-0.12 ; -0.59/-0.16

  - Note in Table 1: 1/ Mean forecast error. 2/ Coefficient on lagged forecast error in a regression of the forecast error on a constant and its lagged realization. 3/ In percent of GDP.

### Long-run performance for G-7 countries
- Using a longer dataset starting in the early seventies, findings include:
  - Forecast accuracy has deteriorated somewhat since the last evaluation (Artis, 1996).
  - WEO forecasts systematically and significantly overpredicted economic growth for all European G-7 economies and Japan during 1991–2003.
  - U.S. growth was underpredicted after 1990, although the bias was not found to be statistically significant.
  - Inflation was strongly and significantly overpredicted for Canada, France, Japan, and the United States during the 1990s and 2000s.
  - Inflation was underpredicted by a significant margin for Italy during the same period.
- Possible explanations:
  - Structural breaks in underlying series (for example, a break toward higher productivity growth in the United States).
  - Underlying assumptions in the forecasting framework—such as the 5-year elimination of the output gap—may generate biases.

### Comparison with Consensus forecasts
- Period covered: 1990–2003.
- Overall outcome:
  - Forecast performance of the WEO is similar to that of the Consensus forecast.
  - Current-year WEO forecasts of GDP growth in the G-7 economies are generally less biased than current-year Consensus forecasts.
  - Bias in next-year forecasts is stronger in the WEO than in the Consensus across the board.
- Timing matters:
  - WEO current-year forecasts perform well against March Consensus forecasts and considerably better against February Consensus forecasts.
  - WEO forecasts tend to perform considerably worse against Consensus forecasts reported in April, reflecting the WEO’s longer gestation lag.
- Information content:
  - With the possible exception of next-year inflation forecasts, there is little systematic evidence that WEO forecasts could be improved by modifying them to account for information embodied in the Consensus forecasts.

### Recommendations
- Timeliness of information is key
  - Systematic gains in forecasting accuracy from using the latest available information.
  - Updating process adds significant value, especially for the G-7 economies.
  - Recommendation: IMF country economists should update their projections just before publication.

- Continuous monitoring of forecasting performance
  - Empirical analysis indicates structural instability in variables, especially real GDP growth and inflation.
  - Recommendation: explore instituting real-time forecasting performance indicators.

- Use bias-adjusted forecasts as guidance
  - Shrinking forecasts toward bias-corrected values is a simple approach to improve forecasts if bias can be precisely estimated.
  - Caveat: bias-adjustment may be too mechanical and assumes constant bias.
  - Recommendation: compare unadjusted forecasts with bias-adjusted forecasts to understand magnitude and direction of biases.

- Forecasts of risk
  - Given inherent uncertainty, strongly recommend incorporating advances in modeling and forecasting risk/uncertainty.
  - Example recommendation: present the full probability distribution of key variables over time.

- Review the output gap assumption
  - WEO assumption that the output gap is removed within a relatively short period may contribute to overprediction of output growth in countries with large output gaps.
  - Recommendation: analyze costs and benefits of this practice and consider more frequent reviews of potential output estimates.

*Source: Author’s calculations.*

### Section XIV concludes.

### Section XIV concludes.

### II. DESCRIPTION OF WEO DATASET
- Coverage:
  - Data on 178 countries over the period 1990–2003.
  - Countries grouped into seven regions: Africa (50 countries), Central and Eastern Europe (15), CIS and Mongolia (13), Developing Asia (24), Middle East (14), Western Hemisphere (33), and Advanced Economies (29).
  - Five forecasted variables: GDP growth, export volume, import volume, inflation, and current account balance (in U.S. dollars or as a percentage of GDP).
  - Maximum of 14 annual data points (1990–2003); 13 for next-year forecasts that begin in 1991.
  - Data availability: most complete for the Advanced Economies and least complete for CIS and Mongolia; current account data tends to contain somewhat fewer observations.
- Forecast vintages and timing:
  - WEO publishes April and September current- and next-year forecasts → four forecasts per predicted year (April and September next-year; April and September current-year).
  - Outcome measurement: first-available data in the April WEO issue of year t + 1 for period t (yt); next-year forecasts for period t + 1 compared to realized values in September WEO issue of year t + 2 (yt+1).
  - Notation for forecasts and errors preserved exactly as in source (April/Sep, current/next-year), producing four forecast errors and two revision measures between April and September.

### III. PROPERTIES OF OPTIMAL FORECASTS (benchmark tests)
- Optimal forecasts assumed to minimize mean squared error (MSE) → conditional expectation under quadratic loss.
- Key properties and tests:
  - Unbiasedness and lack of serial correlation:
    - Regression tests: et = α + εt and et+1 = α + βet + εt+1; equivalently Mincer-Zarnowitz levels regression yt+1 = α + βŷt+1,t + εt+1 with unbiasedness requiring α = 0, β = 1.
  - Informational efficiency (orthogonality):
    - No variable zt in information set Ωt should predict future forecast errors: et+1 = α + βzt + εt+1 with α = β = 0 under efficiency.
  - Forecast revisions:
    - Revisions should be unpredictable (martingale difference); E[revt | Apr t Ω] = 0 and E[rev t+1,t | Apr t Ω] = 0 under conditional-expectation forecasts and law of iterated expectations.
    - Predictable revisions imply simple learnable improvements: improved April forecast = April forecast + fitted revision (equation (5)).
  - Non-increasing variance with shortened horizon:
    - Var(e1,Sep t ) ≤ Var(e1,Apr t ) and Var(eSep t ) ≤ Var(eApr t ) (equation (8)); variance ratio/variance-pattern tests appropriate given small samples.

### IV. EMPIRICAL RESULTS — SUMMARY FINDINGS
- General:
  - Use robust statistics where outliers dominate (e.g., inflation, trade volumes).
  - Table 3 (source) reports mean, median, standard deviation, average absolute first-order autocorrelation, and percentage of positive forecast errors by region and variable.
- GDP growth — current-year:
  - Advanced Economies: mean current-year forecast error very close to zero.
  - April current-year biases exceed more than one percent and are negative for Africa, Central and Eastern Europe, CIS and Mongolia, and the Middle East; September forecasts significantly reduce these biases.
  - Standard deviations of April current-year errors: exceed 8 percent for CIS and Mongolia and 6 percent for the Middle East (outliers dominate).
  - Robust measures: African same-year forecasts show between 34 and 40 percent overpredictions; median forecast error for Africa is -0.81.
  - Variance reduction example: Advanced Economies current-year standard deviation reduced from 1.36 percent in April to 0.81 percent in September — representing a 40 percent reduction.
- GDP growth — next-year:
  - Next-year biases generally exceed current-year biases and are negative in every region for mean April or September biases (WEO tends to overpredict next-year GDP growth).
  - Advanced Economies next-year biases: -0.36 and -0.55 percent (depending on reporting date).
  - Standard deviations for April vs September next-year forecasts are much more similar than for current-year forecasts (less learned between April and September for next-year).
  - Proportion of positive next-year forecast errors: Africa 0.33; Western Hemisphere 0.35.
  - Serial correlation: high average absolute first-order autocorrelation in central and eastern Europe and CIS and Mongolia.
  - Forecast revisions (April→September) show systematic negative mean revisions for GDP (September revised downwards relative to April).
  - Revision standard deviations larger for current-year than next-year.
- Inflation:
  - Outliers large; analyses focus on proportion of positive forecast errors.
  - Current-year inflation: proportion positive not far from 50% except Middle East (34–43%) and Advanced Economies (43%).
  - Next-year inflation: 60–70% of April forecast errors are positive for Africa, central and eastern Europe, and CIS and Mongolia (indicating underprediction); September proportions closer to 60%.
  - Forecast revisions for inflation mostly have positive means; more than 50% of revisions positive in many regions — April forecasts can be improved by increasing values.
  - Generally, April forecast error SD > September SD despite outliers.
- Export and import volumes:
  - Triple-digit outliers affect exports in Africa, CIS and Mongolia, and Middle East; other regions less affected.
  - For regions not affected by outliers, biases modest (example Advanced Economies export bias between 0.22 and -0.39).
  - September errors generally have smaller SD than April errors.
  - Import volume: Advanced Economies, Western Hemisphere, and Developing Asia not affected by triple-digit outliers; average biases ≤ 1 percent in absolute value.
  - Proportions of positive forecast errors close to 50% in most regions/horizons.
- Current account balance:
  - Slight tendency for WEO current-year forecasts to underpredict in central and eastern Europe and developing Asia.
  - Forecast revisions mostly have positive means (September > April), indicating a method to improve April forecasts.

### V. ANALYSIS OF STATISTICAL SIGNIFICANCE (bootstrapped inference)
- Small-sample concerns addressed via bootstrap of forecast errors with 5,000 experiments to construct p-values and percent-significant measures.
- GDP forecasts:
  - April current-year: ~40 percent of African countries’ forecasts systematically too large (overpredictions); bootstrapped tests confirm significant bias for close to 30 percent of African countries; joint bias+serial-correlation test reduces to 15 percent.
  - Serial correlation significant for ~6 percent of African countries (column four reference).
  - CIS & Mongolia and Western Hemisphere: 10–20 percent of countries show significant bias.
  - Middle East: serial correlation more important (15 percent significant).
  - September current-year: proportion of biased cases generally lower than April.
  - Next-year forecasts: significant upward bias in ~35 percent of African countries and ~30 percent of Western Hemisphere countries; >20 percent of Advanced Economies affected; serial correlation pervasive across regions.
- IMF program effect on GDP forecasts:
  - Data: 350 program events; after discarding extreme outliers, between 850 and 960 program-year observations analyzed.
  - Systematic overpredictions (negative mean forecast errors) in program countries; bias smallest for September current-year, largest for April next-year.
  - 60 percent of next-year forecast errors are negative for program countries.
- Inflation tests:
  - Bootstrap shows mild inefficiency in current-year inflation forecasts (positive bias / underprediction in Africa and Western Hemisphere).
  - Strongest evidence against efficiency in next-year inflation forecasts (systematic downward bias in most regions except Advanced Economies).
  - Serial correlation sizable in Advanced Economies next-year inflation errors (rejection rates ~15 percent).
- IMF program effect on inflation forecasts:
  - Program countries show systematic underprediction of inflation (positive mean forecast errors); largest bias is 4.5 percent for April next-year forecasts; mean bias 0.9 percent for September current-year predictions.
- Trade volumes and current account:
  - Little evidence of systematic inefficiencies for trade volumes current- and next-year forecasts.
  - Next-year forecasts of current account show some bias and serial correlation—rejection rates up to twice or four times expected values (example: 24 percent in April next-year forecasts for Western Hemisphere or Developing Asia).

### VI. CAN THE WEO FORECAST ERRORS BE PREDICTED? (predictor-variable tests)
- Predictor variables used: WEO predicted U.S. GDP growth, WEO predicted German output growth, WEO forecast of oil prices, and a global current account discrepancy (sum of current accounts scaled by 15 global export).
- Key findings (proportions of significant t-values reported in source):
  - U.S. GDP forecast:
    - Advanced Economies: 31 percent of April current-year forecasts and 24 percent of September current-year forecasts of U.S. GDP generate t-value above two in explaining forecast errors → predictive power for current-year forecast errors; no evidence for next-year forecast errors.
    - Central and eastern Europe: 33 percent of current-year t-values exceed two (U.S. GDP forecast predicts revisions).
  - German output forecast:
    - Positively correlated and significant in many CIS and Mongolia cases (especially next-year forecast errors).
  - Oil prices:
    - Limited predictive power overall; exceptions include CIS and Mongolia (predicted oil prices positively correlated with GDP forecast errors) and Western Hemisphere/Advanced Economies (negative correlation).
  - Global current account discrepancy:
    - Significant for 40 percent and 24 percent of the Advanced Economies in explaining April current-year and next-year forecast errors, respectively.
- Inflation forecast errors:
  - Next-year inflation errors linked to U.S. GDP forecasts for central and eastern Europe, CIS and Mongolia, Developing Asia, Western Hemisphere, and Advanced Economies.
  - German forecast significant for inflation forecast errors in CIS and Mongolia.
- Current account forecast errors:
  - U.S. growth and oil prices affect errors in 20–30 percent of Middle East countries; global current account discrepancy affects a greater-than-expected number of cases.

### VI.A Output gap tests (Advanced Economies)
- Output gap used as predictor for 29 Advanced Economies (WEO assumes output gap eliminated after five years).
- Findings:
  - GDP forecast errors: regression t-values predominantly negative; around 15 percent of t-statistics exceed two in absolute value.
  - Large negative t-statistics for Germany, France, and Italy — these countries had large average output gaps during 1990s:
    - France: -1.63
    - Germany: -1.99
    - Italy: -2.30
    - Japan: -4.16
  - Implication: assuming too rapid closing of output gaps may lead to upward bias (overprediction) of growth.
  - Inflation forecast errors: sign of output-gap coefficient predominantly positive (WEO tends to underpredict inflation the larger the output gap).
  - Current account forecast errors: many instances of predictability from output gap with varying signs (examples: Hong Kong, Japan, Netherlands, Singapore, Sweden).

### VI.B Overall country/region scores
- Aggregate score per country: sum of three indicators (bias, serial correlation, predictability by instruments), each indicator = 1 if corresponding t-statistic exceeds two in absolute value.
- Scores produced for current- and next-year forecasts of GDP growth, inflation, and current account.
- Findings:
  - Inflation forecasts (except Advanced Economies) generate highest average “problem scores” (above one in most regions).
  - Next-year forecasts generally generate higher average scores than current-year forecasts.

### VII. DIRECTIONAL ACCURACY (sign/turning-point ability)
- Method: Pesaran and Timmermann (1992) sign test (PT test) comparing proportion of correctly predicted signs (lP) to benchmark l*P that accounts for marginal distributions.
- GDP directional accuracy:
  - Next-year forecasts: hit rates between 62 and 71 percent; PT test significant for ~20 percent of countries.
  - Current-year forecasts: substantial improvement — April forecasts hit rates 68–78 percent; September forecasts 76–90 percent.
  - Advanced Economies example: next-year 62 percent → April current-year 78 percent → September current-year 90 percent.
  - Largest improvements occur in revisions from previous-year forecasts to current-year forecasts; little improvement from April→September next-year revisions.
- Inflation directional accuracy:
  - Similar pattern: small change April→September next-year, larger improvements for current-year forecasts; Advanced Economies next-year hit rate higher than GDP next-year, but similar September current-year hit rates for inflation and GDP.
- Current account directional accuracy:
  - Generally lower than for inflation or output; smaller PT statistics and fewer significant tests.

### VIII. REVISIONS FROM BOARD TO PUBLISHED FORECASTS
- Board forecasts: February (current-year) and July (next-year) versions prior to April/September publications.
- Tests: regress Board forecast errors on Board→Publication revisions (equations (10)).
- Advanced Economies:
  - Current-year Board-to-Publication revision contains valuable information: revision correlated with forecast error for ~50 percent of countries; revision coefficient positive for 80–90 percent of countries.
  - Average R2 ≈ 0.25 for current-year regressions → ~25 percent of current-year February/July forecast error explained by Board→Published revision.
  - Next-year regressions: lower significance; ~60 percent of β-estimates positive; ~10 percent β-estimates > 2; average R2 ≈ 0.10.
  - MSE ratios (publication vs Board) for February/April same-year forecasts: average ≈ 0.80 for Advanced Economies; July/September same-year ≈ 0.70; next-year ≈ 0.95. Values below unity indicate improvement from Board → published.
- Other regions:
  - Lower fraction of positive and significant β coefficients (≈10–15 percent current-year; 5–10 percent next-year).
  - MSE ratio tends to be lower than for Advanced Economies for some regions, especially for next-year forecasts, implying notable improvement Board→published.

### IX. RECENT PERFORMANCE (2001–2003 snapshot)
- Table 11 summarizes 2001 recession and 2002–2003 recovery forecasts (mean actuals and forecast errors in April/September previous year and April/September current year).
- Selected facts:
  - U.S. GDP growth was 0.25 percent in 2001.
  - 2001 actual regional growth: CIS and Mongolia ~8 percent; Africa 4.3 percent; Middle East 3.5 percent.
  - WEO overpredicted 2001 growth in all regions except CIS and Mongolia (and September 2000 forecast of 2001 growth in Middle East).
  - Forecast revisions pattern: little downward revision between April and September of previous year; substantial reductions between September previous year → April current year → September current year.
  - Example Advanced Economies: September 2000 forecast error for 2001 growth = 2.17 percent → reduced to 1.32 percent (April 2001) → 0.35 percent (September 2001).
  - 2002 recovery: April and September previous-year forecasts overpredicted growth in six of seven regions; April 2002 current-year forecasts much closer (within 0.4 percent in four regions).
  - 2003 forecasts: very precise for Africa, central and eastern Europe, and developing Asia (all forecast errors within 0.7 percent).
  - Inflation forecasts: more volatile; Africa previous-year April forecasts underpredicted inflation widely (2001: 12.6 percent; 2002: 5.3 percent); inflation forecasts generally accurate for 2003; Advanced Economies inflation forecasts within 0.45 percent of actuals.
  - Current account forecasts: general tendency to underpredict during 2001–2003 for CIS and Mongolia, Developing Asia, Middle East, and Advanced Economies; errors modest and typically shrink as forecast horizon shortens.

### X. LONG-RUN FORECASTING PERFORMANCE FOR G-7 ECONOMIES (introductory note)
- Extends prior postmortem (Artis, 1997) using newer data.
- Variables available for long-run analysis: GDP growth, inflation, export and import volume, and current account balance.
- In GDP growth and inflation cases, current-year forecasts date back to 1971; next-year forecasts date back to (text truncated in source).

*Source: _wp0659 - Section XIV concludes.*

### 1973. In the case of the import and export volume and the current account balance, the current

### _wp0659 - 1973. In the case of the import and export volume and the current account balance, the current

### Data and sample
- Current-year forecasts for import and export volume and the current account balance begin in 1972; next-year forecasts begin in 1973.
- Outcome data extend up to 2003, providing between 31 and 33 annual observations for series noted.
- WEO reports current-year and next-year forecasts in both April and September. Analysis focuses on:
  - April current-year WEO forecast (denoted ŷt,WEO)
  - September next-year WEO forecast (denoted ŷt+1,WEO)
- Consensus forecasts used: March current-year (ŷt,cons) and September next-year (ŷt+1,cons); Latin America uses February/August due to coverage limitations.
- Consensus forecast is mean across survey participants; data coverage: G-7 current-year 14 / next-year 13 (1991–2003).

### A. GDP growth — forecasting performance and diagnostics
- Full-sample current-year mean absolute forecast errors range from 0.64 (France) to 1.21 (Japan).
- RMSFE-values vary from 0.90 (USA) to 1.53 (Japan).
- Forecast errors for all countries except the US have slightly negative mean (over-predictions).
- Next-year forecast biases negative for six of seven economies; next-year biases larger (in magnitude) than current-year biases.
- Little evidence of serial correlation in full-sample forecast errors except next-year UK (first-order serial correlation coefficient 0.41).
- Relative performance versus naive benchmarks (Table 13):
  - WEO RMSFE about 40 percent of random walk RMSFE and about 50 percent of recursive mean RMSFE for current-year forecasts.
  - Next-year RMSFE ratios increase by 0.10–0.20 relative to current-year ratios (i.e., next-year ≈ 0.60 relative to random walk; 0.60–0.80 relative to recursive mean).
- Subsample split (two equal subintervals) shows RMSFE-performance of WEO forecasts relative to naive benchmarks has generally worsened; in second subsample recursive mean sometimes produces lower RMSFE than WEO next-year forecasts for France, Germany, Italy.
- Efficiency/univariate regressions (Table 14):
  - Tests regress forecast errors on (1) a constant, (2) a constant and lagged error, (3) a constant and lagged realization.
  - Standard errors HAC (Bartlett kernel up to 2 lags).
  - Notable: positive bias in next-year forecast errors significant for France and Germany.

### B. Inflation — forecasting performance and diagnostics
- Inflation RMSFE-values similar to GDP growth RMSFEs for most G-7 countries.
- Mean absolute forecast errors higher for next-year than current-year forecasts.
- Evidence of positive serial correlation at the next-year horizon for six of seven economies.
- Theil statistics (current-year WEO RMSFE ratios):
  - Relative to random walk: between 0.36 and 0.74.
  - Relative to recursive mean: between 0.20 and 0.37 (lowest values for United States).
- Next-year Theil statistics remain well below unity; relative performance only marginally worse than current-year.
- Over time: WEO next-year forecasts perform worse relative to random walk but improve relative to recursive mean (possible structural breaks).
- Efficiency and Mincer-Zarnowitz regressions show scant evidence of biases in current-year inflation forecasts; exception Italy. Some serial correlation for France, Italy, United Kingdom.

### C. Import and export volume and current account
- For current account balance forecasts, best naive forecast is the random walk; performance versus random walk deteriorates somewhat in second half of sample.
- Theil statistics for import and export volume WEO current- and next-year forecasts relative to random walk and recursive mean all < 1 (Table 13).
- Subsample analysis: increasingly difficult to beat naive benchmarks; several instances where RMSFE ratio > 1 against recursive mean for trade volume variables (3–4 cases), indicating simple smooth forecasts perform well for about half of countries.
- Against random walk, RMSFE ratio always below unity — WEO forecasts dominate random walk for trade volumes.

### D. Performance after 1990 (post-1990 subsample)
- Next-year WEO systematically overpredicted economic growth after 1990 for European G-7 (France, Germany, Italy, and to smaller degree UK). Japan growth also overpredicted; high t-values despite short sample.
- US growth underpredicted after 1990 (no statistically significant corroboration).
- Next-year inflation forecasts post-1990:
  - Strong and significant overprediction for Canada, France, Japan, United States.
  - Significant underprediction for Italy.
  - Evidence of serial correlation in forecast errors for Italy.
- Two possible explanations: structural breaks in output growth and inflation series; or conditioning information (e.g., assumption output gap eliminated quickly) inducing biases.

### XI. Comparison of WEO and Consensus forecasts — data, correlations, biases
- Consensus dataset (1990–2003): covers G-7, seven Latin American economies, nine Asian economies. March/September timing except Latin America (February/August).
- Pairwise correlations (Table 16): WEO and Consensus forecasts and forecast errors strongly positively correlated in all countries and horizons.
  - Weakest reported correlation: 0.75 for United States next-year forecast levels.
  - Current-year forecast correlations exceed 0.96 generally; next-year correlations somewhat lower.
- Bias comparisons (Table 17):
  - Current-year: WEO forecasts generally less biased than Consensus for many regions.
  - Next-year: WEO next-year bias stronger than Consensus for every country (G-7, Latin America, Asia).
  - Examples: Venezuela next-year bias -3.8 percent (WEO); G-7 biases > 1 percent observed for Germany, Italy, Japan (WEO overpredicted growth).
  - Inflation biases: next-year biases higher than current-year in majority; G-7 WEO values < 0.3 percent, Consensus < 0.4 percent.
  - WEO forecasts perform better than Consensus in absolute bias for 8/9 current-year and 7/9 next-year inflation forecasts in Asia.
- Directional/sign proportions (Table 18): in most cases differences between WEO and Consensus forecast signs not systematic (proportions near 0.50), with notable exceptions listed in text (e.g., next-year GDP for Germany, China, India, Singapore; next-year inflation for France, Peru; current account for France, Japan, Indonesia).

### C–E. Sources of relative performance, statistical tests, informational efficiency
- Bootstrapped p-values for bias and joint bias/serial-correlation tests (Table 19):
  - Majority show no significant evidence of bias or serial correlation.
  - Exceptions: next-year WEO GDP forecasts for Germany and Italy show strong bias and serial correlation; France and Japan show bias evidence.
  - Consensus forecasts also problematic for Germany and Italy (bias), serial correlation for Italy.
  - Inflation forecasts: many rejections of efficiency in Asia — WEO and Consensus biased/serially correlated for China, Hong Kong, Malaysia, Taiwan; Consensus inflation forecasts biased/serially correlated for Singapore.
  - Current account forecasts mostly unbiased/unserially correlated with some exceptions (e.g., WEO current-year Brazil serial correlation; UK next-year biased and serially correlated).
- RMSFE ratios (Consensus over WEO, Table 19):
  - Current-year WEO GDP forecasts generally better than Consensus for G-7 and Latin America (RMSFE ratio > 1 indicates Consensus worse).
  - Consensus better for Asian economies (only 2/9 current-year RMSFE ratios > 1).
  - Next-year GDP: performance similar (RMSFE ratios between 0.90 and 1.10 mostly); China notable exception (RMSFE ratio 0.85, Consensus better).
  - WEO current-year inflation forecasts perform well vs Consensus in all three regions, especially Latin America. Next-year inflation: Consensus generally better for Latin America (except Peru); similar quality for G-7 and Asia.
  - WEO current- and next-year current account forecasts generally better in G-7 and Latin America; similar in Asia.
- Diebold-Mariano framework with bootstrapped p-values (Table 20):
  - Small samples; not many significant cases.
  - More significant cases where WEO dominates Consensus than reverse, particularly for current-year forecasts.
  - Specific significant cases noted: WEO current-year GDP best for UK, Colombia, Mexico, Indonesia, Thailand; WEO current-year inflation forecasts for Argentina, Venezuela, China dominate; Consensus next-year inflation better for Italy, Japan, Brazil, Hong Kong, Singapore, South Korea.
- Informational efficiency regressions (Table 21):
  - Regress WEO forecast errors on Consensus forecasts (and vice versa).
  - If β = 0, WEO encompasses Consensus information.
  - Estimated |β| > 2 in 6 of 46 cases for GDP (Brazil, Mexico, Venezuela, China, Indonesia, South Korea), in 8 cases for inflation, 5 cases for current account — Consensus forecasts help predict WEO errors in these cases.
  - Reverse regressions: estimated WEO coefficient significant in explaining Consensus forecast errors in 3 (GDP), 12 (inflation), and 3 (current account) cases — Consensus fails to encompass WEO in these cases.

### F. Timing sensitivity and Board updates
- Timing exercises:
  - Comparing WEO (April/September) to Consensus (February/August) — dating advantage to WEO: February Consensus GDP current-year forecasts generate higher RMSFE than WEO for many countries (6 G-7, 7 Latin America, 8 Asia).
  - Comparing WEO to April/October Consensus (Consensus has informational advantage) — Consensus dominates in majority of cases (Table 23), except current-year inflation in Latin America.
- WEO Board forecasts (February/July) vs published (April/September):
  - RMSFE ratios (published / Board) reported for WEO and analogous ratios for Consensus.
  - For current-year G-7 GDP forecasts, in six of seven cases ratio lowest for WEO — WEO revisions between Board and Publication date relatively efficient.
  - Next-year G-7 GDP forecasts: WEO and Consensus RMSFE ratios similar.

### XII. Forecast combinations and weights
- Theoretical variance-minimizing weights shown; empirical Granger-Ramanathan regressions (equation (14)) estimate combination weights (Table 25).
- Findings:
  - Current- and next-year GDP combination regressions generate large positive weights on WEO forecasts for many economies (notable exceptions: Italy, Germany, US, Indonesia, Thailand with negative WEO weight).
  - More than half of weights on Consensus forecasts are negative in GDP regressions.
  - Many coefficients on current-year WEO forecasts for G-7 and Latin America significant.
  - For inflation:
    - Current-year WEO weights generally close to unity for G-7.
    - Next-year: weight on Consensus higher for six of G-7 economies.
    - Latin America: WEO weights high for current-year inflation; next-year Consensus weights higher in 4/7 economies.
    - Asia: WEO weights higher than Consensus for current-year forecasts in 5/9 countries; generally lower for next-year inflation.
  - Current account regressions: WEO weights larger than Consensus for most Latin American countries; no clear pattern for G-7 and Asia.
- Combining (or moving WEO towards/away from Consensus) — “too much (or too little) Consensus” construct:
  - Represent reported WEO ŷWEO = γ*ŷcons + (1−γ)*ŷ*WEO where γ measures pull toward Consensus.
  - Indexed re-weighting parameter a used to explore modified forecast ỹ = ŷWEO + a*(ŷWEO − ŷcons) (Table 26):
    - a < 0 moves forecast toward Consensus (a = −0.5 = equal-weight average).
    - For G-7 current-year GDP, modest evidence that pushing WEO further away from Consensus slightly improves forecasts for 4/7 economies but gains modest.
    - No evidence improvements for next-year GDP by moving away from Consensus.
    - For next-year inflation, large gains often obtained by pulling WEO strongly toward Consensus (a = −0.25 or −0.5), i.e., equal-weighted average works well across groups (except China).

### XIII. Recommendations (explicit)
- Streamline process to ensure use of the latest available information; timing crucial — update process adds significant value, mostly for current-year forecasts.
  - Evidence: next-year ratio of RMSFE for Consensus relative to WEO decreased relative to current-year ratios for many cases (15/21, 18/21, 17/27 for G-7, Latin America, Asia respectively), indicating WEO performance deteriorates moving to next-year horizon.
- Real-time monitoring of forecasting performance; detect structural instability:
  - Use reverse-ordered Cusum tests (Pesaran and Timmermann (2002)) or structural stability tests on forecast errors.
- Use bias-adjusted forecasts as guidance, recognizing caveats:
  - Simple add-back of estimated bias is mechanical and assumes constant bias and precise estimate; comparison of unadjusted vs bias-adjusted forecasts can illustrate magnitude/direction of bias.
- Incorporate quantitative probability/risk forecasting methods (fan charts, two-piece normal, scenario simulations) to present risks and imbalances.
- Re-examine implications of using conditional forecasts (conditioning on removal of imbalances or output gap elimination trajectories) — could induce systematic overprediction where large output gaps exist.

### XIV. Conclusion (summary of paper’s scope and open issues)
- Paper evaluated WEO forecasting performance since 1990 addressing:
  - Precision relative to outcomes.
  - Simple improvements (spillovers, combination).
  - Performance during 2001 downturn and recovery; structural breaks.
  - Relative performance versus Consensus forecasts.
  - Potential of simple combination schemes using WEO and Consensus.
- Caveat: the paper did not assume loss is symmetric; asymmetric loss implies biased forecasts may be optimal (Elliott, Komunjer, Timmermann (2004); Patton and Timmermann (2004) references noted in text).

*Source: Author’s calculations and analysis in the provided IMF document excerpt.*

### References

### _wp0659 - References

### Forecast combination, evaluation, and predictive performance
- Aiolfi, Marco, and Allan Timmermann, 2005, “Persistence in Forecasting Performance and Conditional Combination Strategies,” forthcoming in the Journal of Econometrics.
- Bates, J.M., and Clive W. J. Granger, 1969, “The Combination of Forecasts,” Operations Research Quarterly, Vol. 20, pp. 451–68.
- Clemen, Robert T., 1989, “Combining Forecasts: A Review and Annotated Bibliography,” International Journal of Forecasting, Vol. 5 No. 4, pp. 559–81.
- Diebold, Francis X., and José A. López, 1996, “Forecast Evaluation and Combination,” in Handbook of Statistics, edited by G.S. Maddala and C.R. Rao (Amsterdam: Elsevier).
- Diebold, Francis X., and Roberto Mariano, 1995, “Comparing Predictive Accuracy,” Journal of Business and Economic Statistics, Vol. 13 (November), pp. 253–63.
- Elliott, Graham, Ivana Komunjer, and Allan Timmermann, 2005, “Estimation and Testing of Forecast Rationality under Flexible Loss,” Review of Economic Studies, Vol. 72 (April), pp. 1107–25.
- Patton, Andrew J., and Allan Timmermann, 2004, “Properties of Optimal Forecasts under Asymmetric Loss and Nonlinearity” (London School of Economics and University of California of San Diego).
- Pesaran, M. Hashem, and Allan Timmermann, 1992, “A Simple NonParametric Test of Predictive Performance,” Journal of Business and Economic Statistics, Vol. 10 (October), pp. 461–65.
- Pesaran, M. Hashem, and Allan Timmermann, 2002, “Market Timing and Return Prediction Under Model Instability,” Journal of Empirical Finance, Vol. 9 (December), pp. 495–510.
- Pesaran, M. Hashem, and Allan Timmermann, 2005, “Small Sample Properties of Forecasts from Autoregressive Models under Structural Breaks,” Journal of Econometrics, Vol. 129, Nos. 1 and 2, pp. 183–217.
- Timmermann, Allan, 1993, “How Learning in Financial Markets Generates Excess Volatility and Predictability in Stock Prices,” Quarterly Journal of Economics, Vol. 108 (April), pp. 1135–45.
- Timmermann, Allan, 2005, “Forecast Combinations,” forthcoming in Handbook of Economic Forecasting, edited by Graham Elliott, Clive W.J. Granger and Allan Timmermann (Amsterdam: North Holland).

### Forecasting methods, real-time data, and macroeconomic models
- Croushore, Dean, 2005, “Forecasting with Real-Time Data,” forthcoming in Graham Elliott, Clive W.J. Granger, and Allan Timmermann eds, Handbook of Economic Forecasting (North Amsterdam: Holland).
- Stock, James H., and Mark W. Watson, 1996, “Evidence on Structural Instability in Macroeconomic Time Series Relations,” Journal of Business and Economic Statistics, Vol. 14 (January), pp. 11–30.
- Stock, James H., and Mark W. Watson, 2001, “A Comparison of Linear and Nonlinear Univariate Models for Forecasting Macroeconomic Time Series,” in Festschrift in Honor of Clive Granger, edited by Robert F. Engle and Halbert White, (Oxford: Oxford University Press) pp. 1–44.
- Gallo, Giampiero M., Clive W. J. Granger, and Yongil Jeon, 2002, “Copycats and Common Swings: The Impact of the Use of Forecasts in Information Sets,” Staff Papers, International Monetary Fund, Vol. 49, No. 1, pp. 4–21.

### Evaluations of institutional and private sector forecasts
- Artis, Michael J., 1997, “How Accurate Are the WEO’s Short-Term Forecasts? An Examination of the World Economic Outlook,” Staff Studies for the World Economic Outlook (Washington: International Monetary Fund).
- Britton, Erik, Paul Fisher, and John Whitley, 1998, “The Inflation Report Projections: Understanding the Fan Chart,” Bank of England Quarterly Bulletin, Vol. 38 (February), pp 30–37.
- Juhn, Grace, and Prakash Loungani, 2002, “Further Cross-Country Evidence on the Accuracy of the Private Sector’s Output Forecasts,” Staff Papers, International Monetary Fund, Vol. 49, No. 1, pp. 49–64.
- Musso, Alberto, and Steven Phillips, 2002, “Comparing Projections and Outcomes of IMF-Supported Programs,” Staff Papers, International Monetary Fund, Vol. 49, No. 1, pp. 22–48.
- Mincer, Jacob, and Victor Zarnowitz, 1969, “The Evaluation of Economic Forecasts” in Economic Forecasts and Expectations: Analyses of Forecasting Behavior, NBER Studies in Business Cycles, Vol. 19, ed. by Jacob Mincer (Cambridge, Massachusetts: National Bureau of Economic Research).

### Robust inference, heteroskedasticity-autocorrelation consistent (HAC) methods, and testing
- Den Haan, Wouter J., and Andrew T. Levin, 1997, “A Practitioner’s Guide to Robust Covariance Matrix Estimation, Handbook of Statistics, Vol. 15, Chapter 12, pp. 291–341.
- Kiefer, Nicholas M., Timothy J. Vogelsang, and Helle Bunzel, 2000, “Simple Robust Testing of Regression Hypotheses,” Econometrica, Vol. 68 pp. 695–714.
- Kiefer, Nicholas M., and Timothy J. Vogelsang, 2002, “Heteroskedasticity-Autocorrelation Robust Standard Errors Using the Bartlett Kernel Without Truncation,” Econometrica, Vol. 70, pp. 2093–95.
- Kiefer, Nicholas M., and Timothy J. Vogelsang, 2005, “A New Asymptotic Theory for Heteroskedasticity-Autocorrelation Robust Rests,” Econometric Theory, Vol. 21, pp. 1130–64.
- Newey, Whitney K., and Kenneth D. West, 1987, “A Simple Positive Semi-Definite, Heteroskedasticity and Autocorrelation Consistent Covariance Matrix,” Econometrica, Vol. 55, pp. 703–708.
- Pesaran, M. Hashem, and Allan Timmermann, 1992, “A Simple NonParametric Test of Predictive Performance,” Journal of Business and Economic Statistics, Vol. 10 (October), pp. 461–65.

*References as listed in the source PDF _wp0659 - References*

---


_Source: https://www.imf.org/-/media/websites/imf/imported-full-text-pdf/external/pubs/ft/wp/2006/_wp0659.pdf_