## Looking for Underlying Structure in WEO Forecasts — Working Paper No. WP/2026/164

## Source details

**Canonical URL:** [Looking for Underlying Structure in WEO Forecasts — Working Paper No. WP/2026/164](https://www.imf.org/-/media/files/publications/wp/2026/english/wpiea2026164-source-pdf.pdf)

## Other formats

- [Markdown version](/-/media/files/publications/wp/2026/english/wpiea2026164-source-pdf.pdf.md)
- [Structured JSON version](/-/media/files/publications/wp/2026/english/wpiea2026164-source-pdf.pdf.json)

---

### Introduction: scope and research questions
- Data: WEO vintages from 1999 to 2023.
- Core questions:
  - Do forecast errors reflect institutional memory (anchoring to historical averages)?
  - Do they reflect a structural optimism that persists independently of the business cycle?
  - How far are they shaped by a top-down global narrative versus factors specific to individual country desks?
- Two main summarized results:
  - The optimism of WEO growth forecasts is a level bias rather than a by-product of deviation from recent history.
  - Forecasts partially fail to propagate global spillovers and fail to preserve the structural link between growth and inflation.

### Empirical strategy (data, horizons, hypotheses, tests)
- Forecast objects and horizons:
  - Forecast path for real GDP growth (y^F) and inflation (π^F) over horizon h = 0,...,5 (nowcast through five-years-ahead terminal).
  - Two forecast-horizon samples:
    - The 3-year horizon (h ≤ 2): includes vintages 2021–2023; captures post-COVID recovery and 2022 energy shock.
    - The 5-year horizon (h ≤ 4): institutional medium-term standard; five-year evaluation restricted to vintages released before 2019.
  - Realized outcomes (y^A) defined using the January 2024 vintage.
- Three hypotheses tested:
  1. Nature of bias: inertia (anchoring) vs. unconditional optimism.
  2. Anatomy of error: temporal (trends vs. cycles) vs. spatial (synchronized global errors vs. idiosyncratic errors).
  3. Structural coherence: whether forecasts preserve inflation-growth coupling compared to mechanical benchmarks.
- Key tests and benchmarks:
  - Panel fixed-effects regression: E_{i,v} = α + γ_i + β D_{i,v} + ε_{i,v}, with E_{i,v} = y^A_{i,v} − y^F_{i,v} and D_{i,v} = ̄y^F_{i,v} − ̄y^H_{i,v}.
  - Principal-component explained variance of levels for realized vs. forecast growth to measure synchronization/fragmentation.
  - Per-country regressions of growth on US and China growth in data and in forecasts to compare spillover loadings.
  - BVAR (Minnesota prior, one-lag) counterfactual benchmark to assess preservation of the growth–inflation link.

### Key empirical findings — optimism and bias
- Growth forecast bias:
  - Forecasts exceed realized growth by about half a percentage point on average.
  - Aggregate mean error (realized minus forecast) for GDP growth: −0.504 pp.
  - Panel regression (A.I) — GDP Growth:
    - Constant (α): −0.5050 (Robust SE 0.1388, t-stat −3.64, P-value 0.000, CI (95%) [−0.78, −0.23])
    - Deviation (β): 0.1477 (Robust SE 0.0931, t-stat 1.59, P-value 0.113, CI (95%) [−0.04, 0.33])
    - Observations: 1,885 | Entities: 29 | R2 (Within): 0.024
  - Conclusion: The insignificant β coefficient for GDP growth indicates forecast errors are independent of departures from historical averages, rejecting the inertia hypothesis and indicating structural optimism captured by α.
- Inflation bias:
  - Aggregate mean error for inflation: +0.426 pp.
  - Panel regression (A.I) — Inflation:
    - Constant (α): 0.4865 (Robust SE 0.2595, t-stat 1.87, P-value 0.061, CI (95%) [−0.02, 1.00])
    - Deviation (β): 0.0041 (Robust SE 0.0007, t-stat 5.51, P-value 0.000, CI (95%) [0.00, 0.01])
    - Observations: 1,740 | Entities: 29 | R2 (Within): 0.004
  - Deviation coefficient is statistically significant but economically negligible: a one-point departure from history moves the error by only a few thousandths of a point.

### Anatomy of optimism: level vs. timing
- Error decomposition (A.IV, percentage points):
  - Level (trend) error, mean absolute: 1.054 pp
  - Shape (cycle) error, mean absolute: 1.700 pp
  - Signed level bias (the optimism): −0.56 pp
  - Signed shape bias (zero by construction): 0.00 pp
- Over-prediction by cyclical starting point:
  - economy above trend (boom): −0.53 pp
  - economy below trend (downturn): −0.48 pp
  - slope on real-time HP gap: −0.08 (t-stat = −6.2)
- Interpretation:
  - Optimism is entirely a level error: average projected growth exceeds average realized growth by about half a percentage point.
  - Shape errors are larger in absolute terms but cancel within the path; only the average error produces systematic optimism.
  - Level bias is present regardless of cyclical starting point (reports for forecasts begun above trend: −0.53 pp; below trend: −0.48 pp).

### Spillovers and forecast fragmentation (synchronization)
- Principal-component explained variance (A.II):
  - Realized growth (Reality, 1999–2023): 61.31%
  - Forecasts (mean over h≤4): 45.29% (Gap vs. reality: 16.02 pp)
  - Excluding the US and China: Realized 62.14%, Forecasts 46.19% (Gap 15.96 pp)
- Forecasts by fixed horizon (target-year basis), first PC explained variance and gap vs. reality:
  - h=0 (nowcast): 66.12% (Gap −4.81 pp)
  - h=1: 32.29% (Gap 29.02 pp)
  - h=2: 46.27% (Gap 15.03 pp)
  - h=3: 48.85% (Gap 12.46 pp)
  - h=4: 47.38% (Gap 13.93 pp)
- US and China spillover slopes (Table 3, mean-group per-country OLS slopes; 27 economies other than US and China):
  - Data: US slope 0.837 (SE 0.101), China slope 0.327 (SE 0.047)
  - Forecast: US slope 0.337 (SE 0.078), China slope 0.209 (SE 0.052)
  - forecast − data: −0.500 (US) (SE 0.070); −0.118 (China) (SE 0.047)
  - Efficiency regression (error on realized US, China): US slope 0.895 (SE 0.139), China slope 0.202 (SE 0.124) (R2 = 0.35)
  - Wild-cluster bootstrap p-value for US slope in efficiency row: 0.14 (on the conservative side)
- Persistence and time variation (A.II):
  - Pre-2010 (1999–2009): data US slope 0.640, forecast 0.422, Captured 66%
  - Post-2010 (2010–2023): data 1.142, forecast 0.432, Captured 38%
  - 2010–2019 (ex-COVID): data 0.516, forecast 0.048, Captured 9%
  - 2020–2023 (COVID)†: data 2.466, forecast 0.851, Captured 35% († directional only, barely identified)
  - Pooled (1999–2023): data 0.837, forecast 0.337, Captured 40%
- Interpretation:
  - Forecasts embed under half of the US spillover on average; China loading is understated too.
  - A single global factor explains about 61 percent of cross-country variance in realized growth versus about 45 percent in the forecasts — a gap of roughly 16 percentage points.
  - The fragmentation gap is widest one to two years ahead and absent at the nowcast horizon.

### Growth–inflation coupling and model coherence
- Within-forecast growth–inflation correlation (A.III):
  - Positive co-movement (r>0.3): Count 744, Share 37.7%
  - Negative co-movement (r<−0.3): Count 994, Share 50.4%
  - Weak or no co-movement (−0.3 ≤ r ≤ 0.3): Count 234, Share 11.9%
  - Average correlation: −0.116
- Structural (BVAR and MG) benchmarks (A.III):
  - Local GDP and inflation: β GDP→π 0.1348
  - Local and global GDP and inflation: β GDP→π 0.0741
  - Mean Group estimator: β GDP→π 0.0714
  - Within forecasts (conditional): β GDP→π −0.0164
- Additional robustness (Table 14, IT subsample):
  - Data BVAR slope β (GDP→π): 0.126 (IT sample)
  - Within-forecast conditional slope: 0.009 (IT sample)
- Interpretation:
  - The data carry a positive conditional growth→inflation slope (approximately 0.07–0.13); the forecasts carry essentially none (within-forecast slope near zero or slightly negative).
  - Medium-term normalization assumptions and independently-anchored real and nominal blocks produce a mildly negative projected correlation in many forecast paths.
  - Discarding the conditional growth–inflation signal reduces accuracy, especially at medium-term horizons.

### Forecast accuracy vs. mechanical benchmarks
- RMSE comparison (A.VI, percentage points):
  - 3-Year (Recovery Included): WEO RMSE 2.469, BVAR 2.922, Naïve MA 2.921; WEO vs. BVAR: 15.5% reduction; WEO vs. MA: 15.5% reduction
  - 5-Year (Historical Standard): WEO RMSE 2.456, BVAR 2.791, Naïve MA 2.746; WEO vs. BVAR: 12.0% reduction; WEO vs. MA: 10.5% reduction
- Interpretation:
  - Over a three-year, recovery-inclusive window, expert forecasts reduce RMSE relative to a mechanical two-variable BVAR by about 15 percent.
  - At the five-year horizon the expert advantage is about 12 percent.
  - The BVAR closes most of the gap between expert forecasts and a naive moving average, implying modest marginal return to very large expert information sets for medium-term horizons.

### Evolution of forecast performance over time (sub-periods)
- GDP Bias (3Y) by sub-period (A.VIII):
  - Early (1999–2010): −0.249 pp
  - Pre-COVID (2011–19): −0.597 pp
  - Post-COVID (2020–23): +0.044 pp
- GDP Bias (5Y) by sub-period:
  - Early: −0.388 pp
  - Pre-COVID: −0.681 pp
  - Post-COVID: −0.051 pp
- Inflation Bias by sub-period:
  - 3Y Early: +0.670 pp; Pre-COVID: −0.441 pp; Post-COVID: +1.785 pp
  - 5Y Early: +0.862 pp; Pre-COVID: −0.178 pp; Post-COVID: +1.917 pp
- Scaled forecast performance (5-year window, h≤4) selected metrics:
  - GDP growth, 2011–2019: Bias −0.68, RMSE 3.25, Realized SD 3.58, Theil’s U 0.93
  - GDP growth, 1999–2010: Bias −0.39, RMSE 2.79, Realized SD 3.54, Theil’s U 0.66
  - Inflation, 2020–2023: Bias +1.92, RMSE 3.85, Realized SD 4.47, Theil’s U 0.83
- Interpretation:
  - The low-inflation decade 2011–2019 shows smaller raw errors but worse relative skill (Theil’s U rose from 0.63 to 0.73 for inflation), indicating apparent improvements were regime-driven rather than forecasting skill gains.
  - Forecasts adapt slowly to prolonged regime shifts; anchoring to steady states explains small errors in calm periods and large misses during regime reversals (e.g., post-2020 inflation surge).

### Practical implications and suggested role for structural models
- Main takeaways:
  - Growth forecasts exhibit persistent optimism: average projection about half a percentage point above realized growth.
  - Projections understate cross-country comovement: forecasts carry a weaker common component than outcomes (forecast fragmentation).
  - Projections link real and nominal sides more loosely than the data: the positive conditional growth–inflation slope present in the data is essentially absent in the forecasts.
- Practical role for structural models:
  - Structural models can estimate and test medium-term properties such as comovement with systemic economies and the growth–inflation link.
  - Used alongside expert judgment, structural models can provide disciplined, testable medium-term anchors (years 3–5) that are routinely checked, while leaving expert judgment to lead on near-term developments.
  - Such anchors guard against slow adaptation to prolonged regime shifts (examples documented: medium-term inflation over-prediction in the low-inflation 2010s and under-prediction post-2020).
- Open questions and further work:
  - Why medium-term baselines sit above recent outturns.
  - Why country projections move less closely with systemic economies.
  - Why real and nominal sides are loosely linked.
  - Addressing these questions requires more granular information: comparisons across country groups, desk-level heterogeneity, and before-and-after studies of specific modeling adoptions.

### Data, sample, and variable definitions
- Sample: 29 economies spanning advanced, emerging, and developing countries intensively covered in the WEO process (list of 29 economies provided in source).
- Variable definitions (B.I):
  - GDP Growth (NGDP_R): Gross domestic product, constant prices (Percent change). Sourced from WEO Database variable NGDP_R.
  - Inflation (PCPI): Inflation, average consumer prices (Percent change). Sourced from WEO Database variable PCPI.
  - WEO Vintages: All semi-annual (Spring/Fall) and available quarterly updates from 1999 to 2023.

### Appendix A — Selected detailed statistics (highlights)
- Panel regression (A.I) final estimates reproduced:
  - GDP Growth: Constant −0.5050; Deviation 0.1477; Observations 1,885; Entities 29.
  - Inflation: Constant 0.4865; Deviation 0.0041; Observations 1,740; Entities 29.
- Principal-component explained variance (A.II):
  - Realized: 61.31%; Forecasts (mean h≤4): 45.29% (Gap 16.02 pp).
- Growth–inflation structural slopes (A.III):
  - Data BVAR pooled β GDP→π: 0.1348 (local), 0.0741 (local+global); Mean Group β: 0.0714; Within-forecast β: −0.0164.
- Forecast accuracy (A.VI) RMSEs:
  - 3-Year WEO: 2.469; BVAR: 2.922; Naïve MA: 2.921.
  - 5-Year WEO: 2.456; BVAR: 2.791; Naïve MA: 2.746.

*Source: https://www.imf.org/-/media/files/publications/wp/2026/english/wpiea2026164-source-pdf.pdf*

### References18

### References18

### A. Detailed statistical results (Appendix overview)
- Sections listed:
  - A.I Structural optimism (page 19)
  - A.II Spillovers and synchronization (page 20)
  - A.III Inflation-GDP coupling (page 23)
  - A.IV Error decomposition (page 24)
  - A.V The reality check (page 24)
  - A.VI Counterfactual benchmark (page 24)
  - A.VII Forecast bias statistics (page 25)
  - A.VIII Evolution of forecast performance (page 27)
  - A.IX Robustness checks (page 29)
- Section B (Data and sample) begins on page 31 and includes:
  - B.I Sample composition (page 31)
  - B.IV Variable definitions (page 31)

### Glossary (acronyms preserved exactly)
- BVAR Bayesian Vector Autoregression
- DSGE Dynamic Stochastic General Equilibrium
- GDP Gross Domestic Product
- GPM Global Projection Model
- HP Hodrick–Prescott filter
- IMF International Monetary Fund
- MG Mean Group (estimator)
- MONA Monitoring of Fund Arrangements (database)
- OECD Organisation for Economic Cooperation and Development
- PCA Principal Component Analysis
- PCPI Consumer Price Index
- RMSE Root Mean Square Error
- VAR Vector Autoregression
- WEO World Economic Outlook

### Introduction (scope and research questions)
- Data: WEO vintages from 1999 to 2023.
- Core questions:
  - Do forecast errors reflect institutional memory (anchoring to historical averages)?
  - Do they reflect a structural optimism that persists independently of the business cycle?
  - How far are they shaped by a top-down global narrative versus factors specific to individual country desks?
- Two main results highlighted:
  - The optimism of WEO growth forecasts is a level bias rather than a by-product of deviation from recent history.
  - Forecasts partially fail to propagate global spillovers and fail to preserve the structural link between growth and inflation.

### Key empirical findings (from the Introduction and IV.A summary)
- Growth forecast bias:
  - Forecasts exceed realized growth by about half a percentage point on average.
  - The bias is unconditional: holds whether the forecast stance was optimistic or pessimistic relative to recent history.
  - The deviation of a forecast from recent history carries no predictive power for the eventual error.
- Forecast fragmentation and spillovers:
  - Average country’s growth moves with US growth by 0.84 point-for-point in the data but only 0.34 in the forecasts.
  - Forecasts embed under half of the US spillover; China loading is understated too; both shortfalls are statistically significant.
  - A single global factor explains about 61 percent of the cross-country variance in realized growth versus about 45 percent in the forecasts.
  - Forecast errors remain correlated with realized US growth (on the verge of conventional significance).
  - Pre-2010 the forecasts captured about two-thirds of the US spillover; after 2010 they captured well under half.
- Growth–inflation coupling:
  - The conditional growth–inflation slope within the forecasts is essentially flat.
  - This contrasts with the positive structural link present in the data, indicating a disconnection between the real and nominal sides of the projections.

### Empirical strategy (data, hypotheses, and tests)
- Data construction and horizons:
  - Forecast path for real GDP growth (y^F) and inflation (π^F) over horizon h = 0,...,5 (nowcast through five-years-ahead terminal).
  - Two forecast-horizon samples:
    - The 3-year horizon (h ≤ 2): includes vintages 2021–2023; captures post-COVID recovery and 2022 energy shock.
    - The 5-year horizon (h ≤ 4): institutional medium-term standard; five-year evaluation restricted to vintages released before 2019.
  - Realized outcomes (y^A) defined using the January 2024 vintage.
- Three hypotheses tested:
  1. Nature of bias: inertia (anchoring) vs. unconditional optimism.
  2. Anatomy of error: temporal (trends vs. cycles) vs. spatial (synchronized global errors vs. idiosyncratic errors).
  3. Structural coherence: whether forecasts preserve inflation-growth coupling compared to mechanical benchmarks.
- Inertia vs. optimism regression (panel fixed-effects):
  - E_{i,v} = α + γ_i + β D_{i,v} + ε_{i,v}
  - E_{i,v} = y^A_{i,v} − y^F_{i,v}; negative value indicates over-prediction.
  - D_{i,v} = ̄y^F_{i,v} − ̄y^H_{i,v}, where ̄y^F is average forecast over horizon and ̄y^H is average realized growth over preceding five years.
  - Interpretation:
    - β > 0 implies inertia (under-reacting to structural improvement).
    - β < 0 implies conditional optimism.
    - α < 0 with insignificant β implies structural optimism (persistent over-prediction regardless of cyclical position).
- Spillovers and synchronization tests:
  - Compute share of variance explained by first principal component of growth levels for realized outcomes and forecasts (forecast fragmentation if lower share for forecasts).
  - For each country, regress GDP growth on US and China growth in data and in forecasts; compare average slopes.
  - Regress forecast errors on realized US and China growth to test for predictability.
- Model coherence test:
  - Compute corr(y^F, π^F) and compare to a Bayesian VAR (BVAR with Minnesota prior, one-lag VAR) estimated on the same historical data as a counterfactual benchmark.

### Empirical results — further details
- Nature of bias (IV.A):
  - Deviation term (β) is small and statistically insignificant for growth (Table 1); departures from recent history do not predict forecast errors.
  - Forecasts partially anchor medium-term level to recent history: baseline absorbs roughly a quarter of a country’s recent growth (Appendix Table 13).
  - The constant term captures the optimism: forecasts exceed realized growth by about half a percentage point (Table 1).
- Inflation:
  - Deviation coefficient is statistically significant but economically negligible: a one-point departure from history moves the error by only a few thousandths of a point (Table 1).
  - Mean inflation error is a slight under-prediction (Table 10).
- Long-run steady state bias (IV.B):
  - Medium-term forecast (v+5) compared to preceding 10-year historical average.
  - For GDP growth: forecasts exhibit a mild persistent positive bias (median of 0.21 pp; Figure 5).
  - For inflation: medium-term forecasts sit modestly below recent history at the center, with a long left tail (median of −0.47 pp, mean of −1.18 pp; Figure 5).
  - Cross-country dispersion:
    - Current-year forecast spread tracks realized spread closely.
    - Medium-term spread sits well below both.
    - Over 2011–2019 the realized cross-country interquartile range averaged (text truncated in source).

### Methodological notes and caveats (from source text)
- Using January 2024 realized vintage means forecast errors embed subsequent data revisions as well as genuine forecast misses; longer-horizon errors partly reflect revision noise.
- Principal components provide model-agnostic measure of comovement but do not identify structural shocks (demand versus supply).
- The BVAR benchmark applies light shrinkage to a one-lag VAR (Minnesota prior) to discipline estimation in short samples.

*Source: https://www.imf.org/-/media/files/publications/wp/2026/english/wpiea2026164-source-pdf.pdf*

### 3.1 percentage points, but the medium-term forecast range averaged 2.3, about 27 percent narrower;

### wpiea2026164-source-pdf - 3.1 percentage points, but the medium-term forecast range averaged 2.3, about 27 percent narrower;

### Cross-country dispersion and medium-term convergence
- Realized versus forecast dispersion of inflation:
  - The medium-term forecast range averaged 2.3, about 27 percent narrower than the realized cross-country dispersion of 3.1 percentage points.
  - At long horizons projections compress every country toward a common low-inflation anchor, well inside the dispersion that actually materializes.
  - The convergence is largely assumed: the medium-term normalization assumption pulls each country’s terminal inflation toward target regardless of its starting point, narrowing the forecast cross-section even when realized inflation does not.
- Figure 2 (described): interquartile range of inflation across 29 economies by target year for realized outcome, current-year forecast (h=0), and medium-term forecast (v+5); medium-term spread is markedly narrower; vertical line marks 2008.

### The anatomy of optimism: level versus timing
- Decomposition of forecast path errors:
  - Level error: difference between path’s average projected growth and average realized growth.
  - Shape error: deviation of each year’s forecast from the path average (timing/cyclical profile).
- Key findings:
  - Optimism is entirely a level error: average projected growth exceeds average realized growth by about half a percentage point.
  - Shape errors are larger in absolute terms but cancel within the path; only the average error produces systematic optimism.
  - Level bias is present regardless of cyclical starting point:
    - Forecasts begun with the economy above trend over-predict by 0.53 pp on average.
    - Forecasts begun below trend over-predict by 0.48 pp on average.
    - Regressing the path’s mean error on a real-time HP output gap yields a small tilt (−0.08, t=−6.2), negligible relative to the common half-point optimism.

### Spillovers and forecast fragmentation
- Comovement with systemic economies:
  - In realized data, a one-percentage-point increase in US growth accompanies about a 0.84-point increase in the average country’s growth.
  - In the forecasts, the corresponding loading is only 0.34 (under half the realized loading).
  - The forecast loading on China is understated too, though by less; the US shortfall is statistically significant.
- Principal-component summary:
  - A single global factor explains about 61 percent of variance in realized growth but about 45 percent in forecasts — a gap of roughly 16 percentage points.
  - The principal-component gap has a wide bootstrap interval and corroborates the spillover-loading result.
- Time variation:
  - Splitting the sample at 2010: forecasts captured about two-thirds of the US spillover before 2010 and well under half after.
  - The pandemic widened the gap (realized US loading reaches about 2.5 over 2020–2023, versus roughly 0.5 over 2010–2019), but under-capture is present even in the calm 2010–2019 window where forecast US loading is near zero.
- Robustness checks:
  - The gap remains, and is if anything larger, when the US and China are excluded.
  - The fragmentation is a medium-term phenomenon: widest one to two years ahead and absent at the nowcast horizon.
  - Forecast errors remain correlated with realized US growth; on a conservative bootstrap the US slope sits on the verge of conventional significance.

### Model coherence and information use (growth–inflation link)
- Expected structural relationship:
  - A baseline forecast conditions out contemporaneous supply shocks, so it should show a positive growth–inflation correlation close to the conditional slope in the data.
- Empirical findings:
  - Within projections, the average correlation between projected growth and projected inflation is weakly negative and the conditional slope is essentially flat.
  - Majority of forecast paths show a mildly negative growth–inflation correlation.
  - A panel BVAR on data (conditioning out contemporaneous supply noise) yields a pooled βGDP→π = 0.13; a Mean Group estimator gives 0.07 pooled heterogeneously.
  - Regressing projected inflation on lagged projected growth yields a slope of essentially zero (−0.016).
  - Thus, where the data carry a positive structural link of 0.07 to 0.13, the forecasts carry none.
- Interpretation:
  - Medium-term normalization (inflation converging to target while output recovers to potential) and independently-anchored real and nominal blocks produce a mildly negative projected correlation.
  - The realized data correlation is near zero for different reasons (positive demand-side comovement offset by negative supply-side comovement).
  - Discarding the conditional growth–inflation signal reduces accuracy, especially at medium-term horizons.
- Relative accuracy versus mechanical models:
  - Over a three-year, recovery-inclusive window, expert forecasts reduce root-mean-square error relative to a mechanical two-variable BVAR by about 15 percent.
  - At the five-year medium-term horizon the expert advantage over the BVAR is about 12 percent.
  - The BVAR alone closes most of the gap between expert forecasts and a naive moving average, implying the marginal return to large expert information sets is modest.

### Evolution of forecast performance over time
- Institutional developments: adoption of Global Projection Model (GPM) and more sophisticated DSGE models; growth in staff expertise.
- Sub-period decomposition (Table 11; three sub-periods: 1999–2010, 2011–2019, 2020–2023):
  - GDP growth optimism bias widened from −0.25 pp (early period) to −0.60 pp over 2011–2019, before largely dissipating in post-2020 vintages.
  - Inflation bias by period:
    - 2011–2019 (low-inflation decade): modest over-prediction (−0.44 pp at the 3-year horizon).
    - Post-2020: large under-prediction (+1.79 pp), reflecting difficulty anticipating the global inflation surge of 2021–2022.
    - Early period: under-prediction (+0.67 pp).
  - The over-prediction in the 2010s is specific to that decade; sub-period figures are descriptive averages without formal inference.
  - Theil’s U (scale-free comparison, Appendix Table 12) shows forecasts’ skill relative to a naive no-change benchmark did not improve in the low-inflation 2010s despite smaller raw errors.

### Conclusion and practical implications
- Main takeaways:
  - Growth forecasts exhibit persistent optimism: average projection about half a percentage point above realized growth.
  - Projections understate cross-country comovement: forecasts carry a weaker common component than outcomes (forecast fragmentation).
  - Projections link real and nominal sides more loosely than the data: the positive conditional growth–inflation slope present in the data is essentially absent in the forecasts.
- Practical role for structural models:
  - Structural models can estimate directly and test medium-term properties such as comovement with systemic economies and the growth–inflation link.
  - Used alongside expert judgment, structural models can provide disciplined, testable medium-term anchors (years 3–5) that are routinely checked, while leaving expert judgment to lead on near-term developments.
  - Such anchors guard against slow adaptation to prolonged regime shifts (e.g., medium-term inflation over-prediction in the low-inflation 2010s and under-prediction post-2020).
- Open questions and future work:
  - Several mechanisms remain interpretive hypotheses: why medium-term baselines sit above recent outturns, why country projections move less closely with systemic economies, and why real and nominal sides are loosely linked.
  - Distinguishing among hypotheses requires more granular information: comparisons across country groups, desk-level heterogeneity, and before-and-after studies of specific modeling adoptions.

*Source: https://www.imf.org/-/media/files/publications/wp/2026/english/wpiea2026164-source-pdf.pdf*

### References

### References and Appendix A — Detailed statistical results

### References (selected)
- Alichi, Ali et al. (forthcoming). Is It Mostly Fiscal? An Inconvenient Finding About Forecasters’ Fiscal Multipliers. IMF Working Paper. International Monetary Fund.
- Ball, Laurence, João Tovar Jalles, and Prakash Loungani (2015). “Do forecasters believe in Okun’s Law? An assessment of unemployment and output forecasts”. In: International Journal of Forecasting 31.1, pp. 176–184.
- Blanchard, Olivier (2016). “The Phillips curve: Back to the ’60s?” In: American Economic Review 106.5, pp. 31–34.
- Carrière-Swallow, Yan and José Marzluf (2023). “Macrofinancial causes of optimism in growth forecasts”. In: IMF Economic Review 71.2, pp. 509–537.
- Celasun, Oya et al. (2021). An Evaluation of World Economic Outlook Growth Forecasts, 2004–17. IMF Working Paper WP/21/216. International Monetary Fund.
- Ciccarelli, Matteo and Benoît Mojon (2010). “Global inflation”. In: The Review of Economics and Statistics 92.3, pp. 524–535.
- Coibion, Olivier and Yuriy Gorodnichenko (2015). “Information rigidity and the expectations formation process: A simple framework and new facts”. In: American Economic Review 105.8, pp. 2644–2678.
- Eicher, Theo S et al. (2019). “Forecasts in times of crises”. In: International Journal of Forecasting 35.3, pp. 1143–1159.
- Forbes, Kristin J (2019). “Inflation dynamics: Dead, dormant, or determined abroad?” In: Brookings Papers on Economic Activity 2019.2, pp. 257–338.
- Hellwig, Klaus-Peter (2018). Overfitting in judgment-based economic forecasts: The case of IMF growth projections. International Monetary Fund.
- Kontogeorgos, Georgios and Kyriacos Lambrias (2019). An analysis of the Eurosystem/ECB projections. ECB Working Paper 2291. European Central Bank.
- Lewis, Christine and Nigel Pain (2014). “Lessons from OECD forecasts during and after the financial crisis”. In: OECD Journal: Economic Studies 2014.1, pp. 9–39.
- McLeay, Michael and Silvana Tenreyro (2020). “Optimal inflation and the identification of the Phillips curve”. In: NBER Macroeconomics Annual 34.1, pp. 199–255.
- Pesaran, M Hashem and Ron Smith (1995). “Estimating long-run relationships from dynamic heterogeneous panels”. In: Journal of Econometrics 68.1, pp. 79–113.
- Timmermann, Allan (2007). “An evaluation of the World Economic Outlook forecasts”. In: IMF Staff Papers 54.1, pp. 1–33.
- Tsuchiya, Yoichi (2023). “Assessing the World Bank’s growth forecasts”. In: Economic Analysis and Policy.

### A. Detailed statistical results — A.I Structural optimism
- Panel regression (fixed effects) final parameter estimates:
  - Panel (a): GDP Growth
    - Constant (α): −0.5050 (Robust SE 0.1388, t-stat −3.64, P-value 0.000, CI (95%) [−0.78, −0.23])
    - Deviation (β): 0.1477 (Robust SE 0.0931, t-stat 1.59, P-value 0.113, CI (95%) [−0.04, 0.33])
    - Observations: 1,885 | Entities: 29 | R2 (Within): 0.024
  - Panel (b): Inflation
    - Constant (α): 0.4865 (Robust SE 0.2595, t-stat 1.87, P-value 0.061, CI (95%) [−0.02, 1.00])
    - Deviation (β): 0.0041 (Robust SE 0.0007, t-stat 5.51, P-value 0.000, CI (95%) [0.00, 0.01])
    - Observations: 1,740 | Entities: 29 | R2 (Within): 0.004
- Note: Driscoll-Kraay standard errors with Bartlett kernel. The constant approximately equals the sample-mean forecast error; country effects are mean-zero by construction.
- Conclusion: The insignificant β coefficient for GDP growth indicates forecast errors are independent of the departure from historical averages, rejecting the inertia hypothesis.

### A.II Spillovers and synchronization (GDP growth)
- Principal-component explained variance (first PC, levels):
  - Realized growth (Reality, 1999–2023): 61.31%
  - Forecasts (mean over h≤4): 45.29% (Gap vs. reality: 16.02 pp)
  - Excluding the US and China:
    - Realized growth: 62.14%
    - Forecasts: 46.19% (Gap vs. reality: 15.96 pp)
- Forecasts by fixed horizon (target-year basis), explained variance and gap vs. reality:
  - h=0 (nowcast): 66.12% (Gap −4.81 pp)
  - h=1: 32.29% (Gap 29.02 pp)
  - h=2: 46.27% (Gap 15.03 pp)
  - h=3: 48.85% (Gap 12.46 pp)
  - h=4: 47.38% (Gap 13.93 pp)
- Table 3: US and China spillovers — mean-group per-country OLS slopes (27 economies other than US and China):
  - Data: realized i on realized US, China: US slope 0.837 (SE 0.101), China slope 0.327 (SE 0.047)
  - Forecast: forecast i on forecast US, China: US slope 0.337 (SE 0.078), China slope 0.209 (SE 0.052)
  - forecast − data (within country): −0.500 (US) (SE 0.070); −0.118 (China) (SE 0.047)
  - Efficiency: error i on realized US, China: US slope 0.895 (SE 0.139), China slope 0.202 (SE 0.124) (R2 = 0.35; SEs clustered by year)
  - Note: Wild-cluster bootstrap gives the US slope a p-value of 0.14 for the efficiency row.
- Persistence of external-spillover gap: pre- vs. post-2010 (GDP growth), mean-group US spillover slopes:
  - Pre-2010 (1999–2009): US slope data 0.640, forecast 0.422, Gap (t) −0.217 (−2.9), Captured 66%
  - Post-2010 (2010–2023): US slope data 1.142, forecast 0.432, Gap (t) −0.709 (−5.6), Captured 38%
  - 2010–2019 (ex-COVID): data 0.516, forecast 0.048, Gap −0.468 (−2.4), Captured 9%
  - 2020–2023 (COVID)†: data 2.466, forecast 0.851, Gap −1.615 (−7.2), Captured 35%
  - Pooled (1999–2023): data 0.837, forecast 0.337, Gap −0.500 (−7.1), Captured 40%
  - Note: † The 2020–2023 slope is directional only and barely identified; pooled full-sample forecast slope (0.34) sits below pre-2010 and post-2010 forecast slopes because of per-country regression structure.
- Interpretation:
  - Forecasts under-capture external spillovers from the US and China relative to realized comovement.
  - The under-capture remains after 2010 and is proportionally most severe in 2010–2019 (9% captured), with COVID amplifying realized comovement but not eliminating under-capture.

### A.III Inflation–GDP coupling
- Within-forecast growth–inflation correlation (classification of forecast paths, one per country and vintage, over horizon h≤4):
  - Positive co-movement (r>0.3): Count 744, Share 37.7%
  - Negative co-movement (r<−0.3): Count 994, Share 50.4%
  - Weak or no co-movement (−0.3 ≤ r ≤ 0.3): Count 234, Share 11.9%
  - Average correlation: −0.116
  - Note: 1,972 paths exceed the 1,885 observations of the optimism panel because the panel requires five years of preceding history.
- Structural coherence: Phillips curve from BVAR (β GDP→π and related metrics):
  - Local GDP and inflation: β GDP→π 0.1348; Structural Residual Corr. −0.0578; Trough IRF −0.3138
  - Local and global GDP and inflation: β GDP→π 0.0741; Structural Residual Corr. −0.0522; Trough IRF −0.3202
  - Mean Group (MG) estimator: β GDP→π 0.0714; Structural Residual Corr. −0.0534
  - Within forecasts (conditional): β GDP→π −0.0164; Structural Residual Corr. −0.1099
  - Note: The data carry a positive slope (GDP→π); the forecasts do not.

### A.IV Error decomposition
- Anatomy of optimism: level vs. shape (values in percentage points)
  - Level (trend) error, mean absolute: 1.054 pp
  - Shape (cycle) error, mean absolute: 1.700 pp
  - Signed level bias (the optimism): −0.56 pp
  - Signed shape bias (zero by construction): 0.00 pp
  - Over-prediction by cyclical starting point:
    - economy above trend (boom): −0.53 pp
    - economy below trend (downturn): −0.48 pp
    - slope on real-time HP gap: −0.08 (*** t-stat = −6.2)
  - Note: Signed optimism is entirely in the level; the shape component sums to zero by construction. Over-prediction is similar from booms and slumps, rejecting a resilience/recovery bias.

### A.V The reality check
- Historical correlation check (1951–2023):
  - Observations: 1,730
  - Average Correlation (GDP, Inflation): −0.102
  - Note: This check uses the full realized history available in the January 2024 vintage (1951–2023), unlike the rest of the paper which covers 1999–2023.

### A.VI Counterfactual benchmark — Forecast accuracy (GDP growth RMSE)
- Table 9: RMSE (GDP growth)
  - 3-Year (Recovery Included): WEO 2.469, BVAR 2.922, Naïve MA 2.921; WEO vs. BVAR: 15.5% reduction; WEO vs. MA: 15.5% reduction
  - 5-Year (Historical Standard): WEO 2.456, BVAR 2.791, Naïve MA 2.746; WEO vs. BVAR: 12.0% reduction; WEO vs. MA: 10.5% reduction
- Structural extraction (Latest Vintage History):
  - Phillips Curve Beta (β GDP→π): 0.1348
  - Structural Residual Corr.: −0.0578
  - Trough IRF (Inflation to Shock): −0.3138
- Note: Naïve MA is a flat 5-year moving average of the realized history available at each vintage. Last two columns give percentage RMSE reduction of the WEO forecast relative to each benchmark.

### A.VII Forecast bias statistics and long-run anchors
- Aggregate forecast error bias metrics (realized minus forecast):
  - Mean Error (Bias): GDP growth −0.504 pp; Inflation +0.426 pp
- Long-run steady state bias (medium-term forecast v+5 minus 10-year historical average):
  - GDP growth forecasts: median 0.21 pp (mild positive bias)
  - Inflation forecasts: median −0.47 pp; mean −1.18 pp (modest bias below recent history with a long left tail)
  - Left tail (deviations below −10 pp, about 3 percent of observations) reflects projected disinflation in economies emerging from high-inflation episodes.

### A.VIII Evolution of forecast performance (sub-periods)
- Evolution of forecast bias (early vs. late periods)
  - GDP Bias (3Y): Early (1999–2010) −0.249 pp; Pre-COVID (2011–19) −0.597 pp; Post-COVID (2020–23) +0.044 pp
  - GDP Bias (5Y): Early −0.388 pp; Pre-COVID −0.681 pp; Post-COVID −0.051 pp
  - Inflation Bias (3Y): Early +0.670 pp; Pre-COVID −0.441 pp; Post-COVID +1.785 pp
  - Inflation Bias (5Y): Early +0.862 pp; Pre-COVID −0.178 pp; Post-COVID +1.917 pp
- Scaled forecast performance by sub-period (5-year window, h≤4)
  - GDP growth:
    - 1999–2010: Bias −0.39, RMSE 2.79, Realized SD 3.54, Theil’s U 0.66
    - 2011–2019: Bias −0.68, RMSE 3.25, Realized SD 3.58, Theil’s U 0.93
    - 2020–2023: Bias −0.05, RMSE 2.72, Realized SD 3.80, Theil’s U 0.38
  - Inflation:
    - 1999–2010: Bias +0.86, RMSE 2.91, Realized SD 3.83, Theil’s U 0.63
    - 2011–2019: Bias −0.18, RMSE 2.33, Realized SD 3.25, Theil’s U 0.73
    - 2020–2023: Bias +1.92, RMSE 3.85, Realized SD 4.47, Theil’s U 0.83
- Interpretation:
  - The low-inflation decade 2011–2019 shows small absolute errors but worse relative skill (Theil’s U rose from 0.63 to 0.73), indicating apparent improvements were regime-driven rather than forecasting skill gains.
  - Forecasts anchor to steady states and adapt slowly to prolonged regime shifts; anchoring explains small errors in calm periods and large misses during regime reversals (e.g., post-2020 inflation surge).

### A.IX Robustness checks
- Inertia and forecast efficiency:
  - Baseline deviation β (Table 1): GDP growth +0.148 (t 1.59); Inflation +0.004 (t 5.51)
  - + year (time) effects: GDP +0.120 (t 1.79); Inflation +0.034 (t 0.81)
  - Advanced economies only: GDP +0.107 (t 1.03); Inflation +0.651 (t 2.73)
  - Emerging/developing only: GDP +0.162 (t 1.70); Inflation +0.055 (t 1.17)
  - Direct anchoring (ȳF on ȳH slope): GDP 0.262 (t 5.26); Inflation 0.172 (t 4.54)
  - Coibion–Gorodnichenko (error on revision): GDP −0.098 (t −1.53); Inflation +0.052 (t 0.58)
  - Conclusion: Forecasts partially anchor to recent history (carry roughly a quarter of recent growth into medium-term path) but show no significant under-reaction in predictable errors.
- Structural content of the common factor:
  - First principal-component score regressed on US and China growth explains about 80% of factor variation in realized data but about 50% in forecasts; forecast factor’s loading on US growth is under a third of the realized one (1.21 vs. 4.98).
  - Using trade-weighted partner growth explains more of country growth (54% vs. 40% for US factor), and forecasts capture less of this trade channel too (44% vs. 54%), so synchronization gap is not due to summarizing by US and China alone.
- Coherence under cleaner identification:
  - Following McLeay and Tenreyro (2020), restricting to credible inflation targeters preserves the contrast: the data slope stays near previous estimates (text truncated in source).

*Source: wpiea2026164-source-pdf - References*

### 0.13 and the within-forecast slope near zero (Table 14).

### Looking for Underlying Structure in WEO Forecasts — Working Paper No. WP/2026/164

### Spillover content and coherence robustness (Table 14)
- Panel A: first principal component regressed on US and China growth (Realized / Forecast)
  - R2 (US, China explain the factor): 0.80 / 0.50
  - Loading on US growth: 4.98 / 1.21
  - Loading on China growth: 1.93 / 1.75
- Panel B: coherence test, inflation-targeting subsample (n=14)
  - IT sample / Full sample
  - Data BVAR slope β (GDP→π): 0.126 / 0.135
  - Within-forecast conditional slope: 0.009 / −0.016
  - Note: the IT subsample includes 14 credible inflation-targeting economies: Australia, Brazil, Canada, Chile, the United Kingdom, Indonesia, India, Mexico, New Zealand, Peru, the Philippines, Sweden, the United States, and South Africa.
- Panel C: mean variance share explained (US factor vs. trade-weighted partners) (Realized / Forecast)
  - Mean R2 from US growth: 0.40 / 0.24
  - Mean R2 from trade-weighted partner growth: 0.54 / 0.44
  - Note: trade-weighted partner growth uses export-share weights from IMF IMTS bilateral trade, averaged 2000–2019 over the 29-economy set; higher for trade-weighted partners in both the data and the forecasts, and lower in the forecasts for both regressors.
- Additional note on Panel A: regresses the first principal-component score of growth levels (target year × country, h ≤ 4 for forecasts) on US and China growth.

### Data and sample
- The analysis uses WEO vintages and core macro variables across a 29-economy sample chosen for broad geographic and income-level coverage, spanning advanced, emerging, and developing countries.
- The sample focuses on economies that are intensively covered in the WEO process and typically benefit from dedicated country teams; findings are presented as a conservative assessment with potential implications for the broader membership.

### Variable definitions (B.I)
- GDP Growth (NGDP_R): Gross domestic product, constant prices (Percent change). Sourced from WEO Database variable NGDP_R.
- Inflation (PCPI): Inflation, average consumer prices (Percent change). Sourced from WEO Database variable PCPI.
- WEO Vintages: All semi-annual (Spring/Fall) and available quarterly updates from 1999 to 2023.

### Sample composition (B.II)
- The sample covers 29 economies:
  - ARM: Armenia — Middle East & Central Asia
  - AUS: Australia — Advanced Economies
  - BRA: Brazil — Latin America & Caribbean
  - CAN: Canada — Advanced Economies
  - CHL: Chile — Latin America & Caribbean
  - CHN: China — Emerging & Developing Asia
  - DEU: Germany — Advanced Economies
  - DOM: Dominican Republic — Latin America & Caribbean
  - EGY: Egypt — Middle East & Central Asia
  - FRA: France — Advanced Economies
  - GBR: United Kingdom — Advanced Economies
  - HND: Honduras — Latin America & Caribbean
  - IDN: Indonesia — Emerging & Developing Asia
  - IND: India — Emerging & Developing Asia
  - ITA: Italy — Advanced Economies
  - JPN: Japan — Advanced Economies
  - MEX: Mexico — Latin America & Caribbean
  - MYS: Malaysia — Emerging & Developing Asia
  - NZL: New Zealand — Advanced Economies
  - OMN: Oman — Middle East & Central Asia
  - PAK: Pakistan — Middle East & Central Asia
  - PER: Peru — Latin America & Caribbean
  - PHL: Philippines — Emerging & Developing Asia
  - SGP: Singapore — Advanced Economies
  - SWE: Sweden — Advanced Economies
  - TZA: Tanzania — Sub-Saharan Africa
  - USA: United States — Advanced Economies
  - VNM: Vietnam — Emerging & Developing Asia
  - ZAF: South Africa — Sub-Saharan Africa

*Source: Looking for Underlying Structure in WEO Forecasts — Working Paper No. WP/2026/164*

---


_Source: https://www.imf.org/-/media/files/publications/wp/2026/english/wpiea2026164-source-pdf.pdf_