## wp18122

## Source details

**Canonical URL:** [wp18122](https://www.imf.org/-/media/files/publications/wp/2018/wp18122.pdf)

## Other formats

- [Markdown version](/-/media/files/publications/wp/2018/wp18122.pdf.md)
- [Structured JSON version](/-/media/files/publications/wp/2018/wp18122.pdf.json)

---

### Stylized facts and measurement of forecast errors
- IMF WEO forecasts display an upward bias in real GDP growth forecasts; the median forecast error for the main variable F3,t equals 0.47.
- Focus on h = 3 (averaging forecasted real GDP growth over years T+1, T+2, and T+3 and subtracting the realized average over those years).
- Forecast error definition (general h): Fh,t ≡ 1/h ∑_{j=1}^{h} g^f_{t+j|t} − 1/h ∑_{j=1}^{h} g_{t+j}.
- Observations more than two standard deviations (8.7 percentage points) from the mean are dropped, shrinking the sample by about 3 percent.
- The ratio of actual recessions to one-year ahead predicted ones stands at 2.4 and nearly 4 in every 5 recessions arrived without being forecasted; IMF sometimes predicted recessions that did not materialize.

### Identification strategy: simultaneity and instrumental variables
- Mechanical simultaneity between forecast errors (partly using contemporaneous outcomes) and recession occurrence addressed with instrumental variables (IV).
- Primary instrument: Mission Chief (MC) fixed effects (μ^(k)), exploiting persistent optimism/caution across MCs who rotate across countries.
  - Assumption 1: No systematic relationship in IMF MC allocation between a MC’s degree of forecast optimism/caution and a country’s economic prospects.
  - Exclusion restriction: MC’s degree of over-optimism affects outcomes only via the forecast channel.
- Estimation of MC fixed effects from regression (2): F1,it = α_t + γ_i + λW_it + μ^(k) + ε_it, where F1,it is one-year ahead forecast error.
- Jackknife approach: MC fixed effects used as instruments are inferred from the MC’s forecasts for countries other than the country being instrumented.
- W covariates when estimating μ: (i) IMF program dummy (1 if country i was under an IMF program in year t) and (ii) PPP-adjusted real GDP per capita.

### Data on Mission Chiefs and first-stage diagnostics
- Dataset: country assignments for IMF MCs since 1990; MC identified for about 80% of country-year pairs since 1990.
- Database contains 705 unique MCs; on average MCs led teams for 2.7 different countries; 475 MCs led teams for multiple countries (average 3.5 different countries among those).
- Average estimated MC fixed effect amount: 0.34; standard deviation: 2.6.
- Joint F-test that all MC fixed effects are equal to 0 is rejected at the 1% significance level; inclusion of MC fixed effects increases adjusted R^2 from 0.06 to 0.13.
- First-stage regressions (Table 2a) show μ is a relevant predictor of F3,t−3; Cragg-Donald statistic values reported (e.g., 13.62, 12.77, 6.65, 7.04) and generally pass common weak-instrument thresholds in the full sample.

### Main IV results: over-optimism increases recession risk
- Second-stage IV estimates (Table 2b) — dependent variable: recession dummy (1 = presence of recession):
  - Coefficients on forecast error F3,t−3:
    - Full sample Column (1): 0.1712131 (z-statistic (3.03)) — ∗∗∗
    - Full sample Column (2): 0.1389873 (z-statistic (2.51)) — ∗∗
    - Emerging/developing Column (3): 0.1754304 (z-statistic (2.24)) — ∗∗
    - Emerging/developing Column (4): 0.131479 (z-statistic (1.91)) — ∗
- Finding: over-optimism increases the probability of a recession occurring 3 years later; result robust to inclusion of covariates (public debt/GDP (t-1), inflation (t-1), average growth in trading partners, percent change in terms-of-trade) and appears in both global and emerging/developing samples.
- The result does not hold when focusing solely on the subsample of advanced countries (z-statistic reported as 0.78).

### Additional instrument: team sentiment and overidentification
- Secondary instrument: one-year ahead forecast error used to instrument longer-horizon forecast error (F(2)3,t ≡ 1/2 ∑_{j=2}^{3} g^f_{t+j|t} − 1/2 ∑_{j=2}^{3} g_{t+j}), exploiting that bullish teams tend to be optimistic across horizons.
- One-year horizon excluded from the instrumented variable to separate instrument and instrumented variables.
- When including both MC fixed effects and team-sentiment instrument, Hansen’s J-statistic suggests instruments are uncorrelated with the error term (overidentified specification).
- Quantitative magnitude: over-estimating average growth over the two/three-year horizon by 1 percentage point increases the probability that a recession arrives 3 years later by about 8 percentage points. The over-estimation of growth is larger than this in 40 percent of all country-year pairs in the sample.

### Core empirical findings on forecast errors and macro outcomes
- Past positive forecast errors (over-optimism) are associated with an increased probability of recessions, fiscal crises, and Balance of Payments (BoP) crises in the future.
- Key IV estimates (Table 3b, dependent variable: recession dummy):
  - forecast error on growth “F(2)3,t−3”: 0.0989018 (∗∗∗) (3.20) in Column (1: full)
  - forecast error on growth “F(2)3,t−3”: 0.0872071 (∗∗∗) (2.86) in Column (2: full)
  - forecast error on growth “F(2)3,t−3”: 0.0823811 (∗∗) (2.44) in Column (3: eme/dev)
  - forecast error on growth “F(2)3,t−3”: 0.0734848 (∗∗) (2.20) in Column (4: eme/dev)
  - public debt/GDP (t-1) appears positive and sometimes significant (e.g., 0.0004917 ∗∗ (2.27) in Column (1))
  - Hansen J-statistics (p-values): 4.349 (0.0370), 1.846 (0.1742), 3.424 (0.0643), 1.509 (0.2194) across columns respectively.
- Sustained drops in real income (depression dummy) (Table 4, IV regressions):
  - forecast error on growth “F(2)3,t−3”: 0.2693736 (∗∗∗) (4.29) in Column (1: full)
  - Column (2: full): 0.2335627 (∗∗∗) (4.39)
  - Column (3: eme/dev): 0.2733438 (∗∗∗) (3.66)
  - Column (4: eme/dev): 0.230579 (∗∗∗) (3.81)
  - public debt/GDP (t-1) significant and positive: 0.0017053 (∗∗∗) (4.64) in Column (1)
- Crisis incidence (Table 5, IV regressions):
  - fiscal crisis: forecast error on growth “F(2)3,t−3”: 0.0890415 (∗∗) (2.36) in Column (1)
  - fiscal crisis with covariates: 0.0806261 (∗∗) (2.36) in Column (2)
  - BoP crisis: 0.0904298 (∗∗∗) (2.69) in Column (3)
  - BoP crisis with covariates: 0.0852909 (∗∗) (2.53) in Column (4)
  - public debt/GDP (t-1) large and highly significant for fiscal crises (e.g., 0.0019071 (∗∗∗) (6.19) in Column (1))

### Robustness, alternative estimators, and sample variation
- Results robust to:
  - inclusion of controls,
  - different estimation methods: IV, OLS, probit, logit, Local Projections,
  - different dependent variables: growth rates, recession dummies, crisis dummies,
  - subsamples: results hold for emerging and developing economies; no detectable causal impact when restricted to advanced economies only.
- OLS/logit/lagging strategy (Appendix D):
  - Logit with lagged forecast error (k = 1, 2) shows positive and significant coefficients:
    - Table D2 (logit, global sample): forecast error on growth “F3,t−3−k”: 0.0508912 (∗∗∗) (2.68) for k=2; 0.1014209 (∗∗∗) (5.48) for k=1.
  - Bayesian Model Averaging (Table D8) yields averaged estimate for forecast error coefficient 0.0183287 (∗∗∗) (3.84) when F3,t−4 always present; 0.0177432 (∗∗∗) (3.13) when treated as auxiliary.

### Local Projections and dynamic mechanism
- Local Projection IRFs (LP-IRFs) for a forecast shock (three-year-ahead average forecast ̄gfit) indicate:
  - Real GDP growth: a short-run boost in economic growth that lasts for only two years; from four years after the shock onwards the forecast shock tends to reduce growth (pattern holds in both global and emerging/developing samples).
  - Government debt-to-GDP ratio: no significant impact on impact, but from the second year onwards a forecast shock tends to lead to faster accumulation of public debt (observed in both samples).
  - Private credit-to-GDP ratio: private credit accumulates in response to a forecast shock; the response is quicker than for government debt and is much weaker in the emerging/developing sample.
  - Investment-to-GDP ratio: falls after a forecast shock in both samples — implying investment falls relative to GDP (which tends to rise following the forecast shock), consistent with consumption increasing relative to GDP.
- Narrative from dynamics:
  - Rosier forecasts induce both public and private sectors to accumulate more debt and favor consumption over investment, building financial and fiscal fragilities that raise the likelihood of later recessions and crises.

### Mechan identification and first-stage evidence
- First-stage evidence (Table 3a) shows Mission Chief fixed effect and one-year ahead forecast error significantly predict the three-year-ahead forecast error variable.
- Cragg-Donald statistics reported: 13.42, 13.86, 9.46, 9.67 across first-stage specifications.
- Identification logic frames optimism as a noise shock; negative IV estimates for the effect of forecast errors on growth are interpreted as evidence that over-optimism causes delayed recessions rather than merely temporary booms.

### Policy implications and recommendations
- Basing policy on realistic or cautious medium-term macroeconomic forecasts can mitigate the build-up of debt-financed vulnerabilities that follow overly optimistic forecasts.
- Cautious forecasts can help prevent recessions; the Chile example (Frankel (2011)) is cited where conservative copper-price budgeting enabled countercyclical fiscal policy during the Global Financial Crisis.
- Given the mechanism runs through higher debt accumulation (public and private) and relative declines in investment, recommended policy actions include:
  - Treat overly-rosy medium-term growth projections with caution when setting fiscal policy and borrowing plans.
  - Strengthen fiscal buffers and monitoring of private credit expansion during periods of elevated optimism.
  - Favor policies that limit procyclical borrowing against anticipated income streams that may not materialize.

*Source: wp18122 (PDF chapter/section).*

### 0.58 percentage points higher than the subsequent realization.  Such an upward bias is

### wp18122 - 0.58 percentage points higher than the subsequent realization.  Such an upward bias is

### Stylized facts and measurement of forecast errors
- IMF WEO forecasts display an upward bias in real GDP growth forecasts; the median forecast error for the main variable F3,t equals 0.47.
- Many macro decisions are multi-year; the paper focuses on h = 3 (averaging forecasted real GDP growth over years T+1, T+2, and T+3 and subtracting the realized average over those years).
- Forecast error definition (general h): Fh,t ≡ 1/h ∑_{j=1}^{h} g^f_{t+j|t} − 1/h ∑_{j=1}^{h} g_{t+j}.
- Observations more than two standard deviations (8.7 percentage points) from the mean are dropped, shrinking the sample by about 3 percent.
- The ratio of actual recessions to one-year ahead predicted ones stands at 2.4 and nearly 4 in every 5 recessions arrived without being forecasted; IMF sometimes predicted recessions that did not materialize.

### Identification strategy: simultaneity and instrumental variables
- Problem: mechanical simultaneity between forecast errors (partly using contemporaneous outcomes) and recession occurrence. Solution: instrumental variables (IV).
- Primary instrument: Mission Chief (MC) fixed effects (μ^(k)), exploiting persistent optimism/caution across MCs who rotate across countries.
  - Assumption 1: No systematic relationship in IMF MC allocation between a MC’s degree of forecast optimism/caution and a country’s economic prospects.
  - Exclusion restriction: MC’s degree of over-optimism affects outcomes only via the forecast channel.
- Estimation of MC fixed effects from regression (2): F1,it = α_t + γ_i + λW_it + μ^(k) + ε_it, where F1,it is one-year ahead forecast error.
- Jackknife approach: MC fixed effects used as instruments are inferred from the MC’s forecasts for countries other than the country being instrumented (preventing use of a country’s own history in its instrument).
- W covariates when estimating μ: (i) IMF program dummy (1 if country i was under an IMF program in year t) and (ii) PPP-adjusted real GDP per capita.

### Data on Mission Chiefs and first-stage diagnostics
- Dataset: country assignments for IMF MCs since 1990; MC identified for about 80% of country-year pairs since 1990.
- Database contains 705 unique MCs; on average MCs led teams for 2.7 different countries; 475 MCs led teams for multiple countries (average 3.5 different countries among those).
- Average estimated MC fixed effect amount: 0.34; standard deviation: 2.6.
- Joint F-test that all MC fixed effects are equal to 0 is rejected at the 1% significance level; inclusion of MC fixed effects increases adjusted R^2 from 0.06 to 0.13.
- First-stage regressions (Table 2a) show μ is a relevant predictor of F3,t−3; Cragg-Donald statistic values reported (e.g., 13.62, 12.77, 6.65, 7.04) and generally pass common weak-instrument thresholds in the full sample.

### Main IV results: over-optimism increases recession risk
- Second-stage IV estimates (Table 2b) — dependent variable: recession dummy (1 = presence of recession):
  - Coefficients on forecast error F3,t−3:
    - Full sample Column (1): 0.1712131 (z-statistic (3.03)) — ∗∗∗
    - Full sample Column (2): 0.1389873 (z-statistic (2.51)) — ∗∗
    - Emerging/developing Column (3): 0.1754304 (z-statistic (2.24)) — ∗∗
    - Emerging/developing Column (4): 0.131479 (z-statistic (1.91)) — ∗
- Finding: over-optimism increases the probability of a recession occurring 3 years later; result robust to inclusion of covariates (public debt/GDP (t-1), inflation (t-1), average growth in trading partners, percent change in terms-of-trade) and appears in both global and emerging/developing samples.
- The result does not hold when focusing solely on the subsample of advanced countries (z-statistic reported as 0.78), consistent with IMF forecasts playing a lesser role where alternative forecasts are readily available.

### Additional instrument: team sentiment and overidentification
- Secondary instrument: one-year ahead forecast error used to instrument longer-horizon forecast error (F(2)3,t ≡ 1/2 ∑_{j=2}^{3} g^f_{t+j|t} − 1/2 ∑_{j=2}^{3} g_{t+j}), exploiting that bullish teams tend to be optimistic across horizons.
- One-year horizon excluded from the instrumented variable to separate instrument and instrumented variables.
- When including both MC fixed effects and team-sentiment instrument, Hansen’s J-statistic suggests instruments are uncorrelated with the error term (overidentified specification) — supporting instrument validity.
- Quantitative magnitude: over-estimating average growth over the two/three-year horizon by 1 percentage point increases the probability that a recession arrives 3 years later by about 8 percentage points. The over-estimation of growth is larger than this in 40 percent of all country-year pairs in the sample.

### Robustness and auxiliary findings
- Results robust to alternative horizons (within IMF WEO limits) and to using growth rate of real GDP as dependent variable (Appendix C).
- When country fixed effects are dropped, the IMF program dummy becomes significantly positive (0.43, p-value 0.03) and real GDP per capita coefficient becomes significantly negative (−0.000015, p-value 0.003), indicating larger forecast errors in less-developed countries (footnote discussion).
- Stock-Yogo (2005) critical values cited: 20% (15%) maximal bias values for Cragg-Donald statistic equal 6.66 (8.96); heteroskedasticity-robust Kleibergen-Paap statistics noted as slightly lower.

*Source: wp18122 - 0.58 percentage points higher than the subsequent realization.  Such an upward bias is (IMF working paper content provided).*

### 25.2 percent, this reflects a 30 percent increase from the mean.

### wp18122 - 25.2 percent, this reflects a 30 percent increase from the mean.

### Core empirical findings on forecast errors and macro outcomes
- Past positive forecast errors (over-optimism) are associated with an increased probability of recessions, fiscal crises, and Balance of Payments (BoP) crises in the future.
- Key IV estimates (Table 3b, dependent variable: recession dummy):
  - forecast error on growth “F(2)3,t−3”: 0.0989018 (∗∗∗) (3.20) in Column (1: full)
  - forecast error on growth “F(2)3,t−3”: 0.0872071 (∗∗∗) (2.86) in Column (2: full)
  - forecast error on growth “F(2)3,t−3”: 0.0823811 (∗∗) (2.44) in Column (3: eme/dev)
  - forecast error on growth “F(2)3,t−3”: 0.0734848 (∗∗) (2.20) in Column (4: eme/dev)
  - public debt/GDP (t-1) appears positive and sometimes significant (e.g., 0.0004917 ∗∗ (2.27) in Column (1))
  - Hansen J-statistics (p-values) reported: 4.349 (0.0370), 1.846 (0.1742), 3.424 (0.0643), 1.509 (0.2194) across columns respectively.
- Sustained drops in real income (depression dummy) (Table 4, IV regressions):
  - forecast error on growth “F(2)3,t−3”: 0.2693736 (∗∗∗) (4.29) in Column (1: full)
  - Column (2: full): 0.2335627 (∗∗∗) (4.39)
  - Column (3: eme/dev): 0.2733438 (∗∗∗) (3.66)
  - Column (4: eme/dev): 0.230579 (∗∗∗) (3.81)
  - public debt/GDP (t-1) significant and positive: 0.0017053 (∗∗∗) (4.64) in Column (1)
- Crisis incidence (Table 5, IV regressions):
  - fiscal crisis: forecast error on growth “F(2)3,t−3”: 0.0890415 (∗∗) (2.36) in Column (1)
  - fiscal crisis with covariates: 0.0806261 (∗∗) (2.36) in Column (2)
  - BoP crisis: 0.0904298 (∗∗∗) (2.69) in Column (3)
  - BoP crisis with covariates: 0.0852909 (∗∗) (2.53) in Column (4)
  - public debt/GDP (t-1) large and highly significant for fiscal crises (e.g., 0.0019071 (∗∗∗) (6.19) in Column (1))

### Robustness, alternative estimators, and sample variation
- Results are robust to:
  - inclusion of controls,
  - different estimation methods: IV, OLS, probit, logit, Local Projections,
  - different dependent variables: growth rates, recession dummies, crisis dummies,
  - subsamples: results hold for emerging and developing economies (where IMF forecasts carry largest weight); no detectable causal impact when restricted to advanced economies only.
- OLS/logit/lagging strategy (Appendix D):
  - Logit with lagged forecast error (k = 1, 2) shows positive and significant coefficients:
    - Table D2 (logit, global sample): forecast error on growth “F3,t−3−k”: 0.0508912 (∗∗∗) (2.68) for k=2; 0.1014209 (∗∗∗) (5.48) for k=1.
  - OLS with lagged forecast error also yields similar patterns (Tables D5–D7).
  - Bayesian Model Averaging (Table D8) yields averaged estimate for forecast error coefficient 0.0183287 (∗∗∗) (3.84) when F3,t−4 always present; 0.0177432 (∗∗∗) (3.13) when treated as auxiliary — forecast error stands out as the robust predictor of future recessions.

### Local Projections and dynamic mechanism
- Local Projection IRFs (LP-IRFs) for a forecast shock (three-year-ahead average forecast ̄gfit) indicate:
  - Real GDP growth: a short-run boost in economic growth that lasts for only two years; from four years after the shock onwards the forecast shock tends to reduce growth (pattern holds in both global and emerging/developing samples). This reversal corroborates earlier IV findings that past over-optimism precedes later recessions.
  - Government debt-to-GDP ratio: no significant impact on impact, but from the second year onwards a forecast shock tends to lead to faster accumulation of public debt (observed in both samples).
  - Private credit-to-GDP ratio: private credit accumulates in response to a forecast shock; the response is quicker than for government debt and is much weaker in the emerging/developing sample.
  - Investment-to-GDP ratio: falls after a forecast shock in both samples — implying investment falls relative to GDP (which tends to rise following the forecast shock), consistent with consumption increasing relative to GDP.
- Narrative from dynamics:
  - Rosier forecasts induce both public and private sectors to accumulate more debt and favor consumption over investment, building financial and fiscal fragilities that raise the likelihood of later recessions and crises.

### Mechan identification and instrument strategy
- Instrumental variables approach:
  - Exploits variation in IMF Mission Chief fixed effects (μ) and one-year ahead forecast errors as instruments for forecast errors, leveraging quasi-random re-assignment of Mission Chiefs across countries.
  - First-stage evidence (Table 3a) shows Mission Chief fixed effect and one-year ahead forecast error significantly predict the three-year-ahead forecast error variable.
  - Cragg-Donald statistics reported: 13.42, 13.86, 9.46, 9.67 across first-stage specifications; countries and observations listed per specification.
- Identification logic (Appendix B):
  - The paper frames optimism as a noise shock and shows that under plausible correlations between instruments and future noise shocks, obtaining a negative IV estimate for the effect of forecast errors on growth is strong evidence that over-optimism causes delayed recessions rather than merely temporary booms.

### Policy implications and recommendations
- Basing policy on realistic or cautious medium-term macroeconomic forecasts can mitigate the build-up of debt-financed vulnerabilities that follow overly optimistic forecasts.
- Cautious forecasts can help prevent recessions; the Chile example (Frankel (2011)) is cited where conservative copper-price budgeting enabled countercyclical fiscal policy during the Global Financial Crisis.
- Given the mechanism runs through higher debt accumulation (public and private) and relative declines in investment, policy should:
  - Treat overly-rosy medium-term growth projections with caution when setting fiscal policy and borrowing plans.
  - Strengthen fiscal buffers and monitoring of private credit expansion during periods of elevated optimism.
  - Favor policies that limit procyclical borrowing against anticipated income streams that may not materialize.

*Italic: Source — wp18122 (PDF chapter/section) as provided.*

---


_Source: https://www.imf.org/-/media/files/publications/wp/2018/wp18122.pdf_