## wp1848 - 5.1 GDP Growth Forecasts

## Source details

**Canonical URL:** [wp1848 - 5.1 GDP Growth Forecasts](https://www.imf.org/-/media/files/publications/wp/2018/wp1848.pdf)

## Other formats

- [Markdown version](/-/media/files/publications/wp/2018/wp1848.pdf.md)
- [Structured JSON version](/-/media/files/publications/wp/2018/wp1848.pdf.json)

---

### Introduction
- Purpose: assess the accuracy of macroeconomic forecasts for countries that experienced crises severe enough to require IMF financial support.
- Evaluation metrics used:
  - Bias: measures forecast deviations from realizations.
  - Efficiency: tests whether forecast errors were unpredictable and whether forecasts incorporate available information.
  - Information content: measures informational value relative to naive forecast models of directional changes.
- Extensions relative to prior work:
  - Use of Mincer and Zarnowitz (1969) and Holden and Peel (1990) regression approaches to examine forecast bias and efficiency.
  - Evaluation of informational value versus naive directional-change models (Merton, 1981; Henriksson and Merton, 1981; Schnader and Stekler, 1990).
  - Examination of IMF crisis forecasts for a much larger set of macroeconomic variables (29).
  - Decomposition of IMF forecast errors of macroeconomic aggregates (GDP, the balance of payments, and the fiscal accounts) to identify subcomponents driving errors.

### Key findings (accuracy, bias, efficiency, informational value)
- Overall informational value
  - For nearly all variables IMF forecasts contain substantial informational value relative to naive forecasting models.
  - Exceptions: forecasts for inflation, government expenditure growth, and net income growth.
- Bias and efficiency (global sample of program countries)
  - Nominal GDP growth is, on average, forecast too optimistically.
  - Current account growth and government expenditure growth are subject to downward bias.
  - Growth forecasts of real GDP, prices, the financial account, and government revenue are found to be unbiased.
  - Forecasts of key indicators of crisis recovery—growth in government expenditure, government revenue, prices, and reserves—are efficient.
- Heterogeneity by country income group
  - Forecasts for Low-Income Countries (LICs) are substantially more biased and inefficient than for Non-LICs.
  - Possible cause: data and information challenges in LICs during crises.
- Decomposition of forecast errors (GDP, fiscal, BOP)
  - GDP growth forecast errors are significantly affected by forecast errors in all GDP subcomponents (government, consumption, investment, net exports).
  - Private consumption growth is the most important contributor to GDP forecast errors.
  - Fiscal budget forecast errors are driven by forecast errors for non-interest/net-lending expenditures, tax/non-tax revenues, and grants.
  - Balance of payments forecast errors cannot, with one exception (goods imports), be linked to forecast errors in its subcomponents.

### Macroeconomic identities used for decomposition
- Aggregate demand (nominal GDP decomposition) per IMF (2007):
  - Y = Cp + Cg + Ip + Ig + X - M
  - Growth-rate decomposition uses elasticities (e.g., σ_{c,y}^p) so subcomponent growth rates enter with weights.
- Current account growth identity (IMF, 2015):
  - ca decomposed into growth of goods and services exports (g_x, s_x), imports (g_m, s_m), net income (ni), and net transfers (nt) with associated elasticities σ.
- Financial account growth identity (IMF, 2009):
  - fa = contributions from net foreign direct investment (fdi), net portfolio investment (pi), reserve assets (res), and other investment (ot).
- Government budget decompositions (IMF, 2014):
  - Government expenditure growth (gx) = growth of interest expenditures (int), non-interest expenditures (nint), and capital expenditure & net lending (cap).
  - Government revenue growth (gr) = growth of tax revenue (tax), non-tax revenue (ntax), and grants (grt).

### Methodology (tests and decomposition)
- Mincer-Zarnowitz regressions:
  - A_{i,p} = α + β F_{i,p} + θ X_{i,p} + ε_{i,p}; efficient forecast implies α = 0 and β = 1; joint F-test used.
  - Controls (X): continental dummies (America, Africa, Asia) and binary Financial Crisis indicator (Crisis08).
- Holden-Peel test for unbiasedness:
  - Regression of forecast errors on forecasts and controls; t-test for γ = 0.
- Merton-Henriksson timing (directional) tests:
  - 2 x 2 contingency of forecast sign versus actual sign; χ^2 test for independence.
- Decomposition of forecast errors:
  - (ŷ_i - ŷ̂_i) = α + Σ_j β_j (x_{ij} - x̂_{ij}) + ε_i
  - Interpretation: a 1% increase in the average forecast error of subcomponent x_j causes a β_j% change in the average forecast error of aggregate y.

### Data, sample construction, and treatment
- Data source: IMF Monitoring of Fund Arrangements dataset (MONA, IMF 2016b).
- Coverage:
  - Original MONA dataset covers 238 crisis programs since 2002.
  - Data availability for broadest macroeconomic identities limited to 156 programs in 84 countries.
- Timing of forecasts in MONA:
  - For program approved at year t, MONA contains entries for t-3, t-2, t-1, t, t+1, t+2, t+3, t+4.
  - This paper focuses on the IMF forecast for year t (the current, first program year).
- Sample inclusion criteria:
  - Include only programs running longer than 18 months.
  - Multiple programs per country treated as independent observations.
- Outlier treatment:
  - Exclude forecast errors that exceed the respective variable means by four standard deviations.

### Empirical results — selected numerical findings (preserve exact reported estimates)
- Mincer-Zarnowitz (Full Sample, observations 110 unless noted)
  - Forecast growth rate coefficients (β): real GDP 0.606; avg. prices 1.031; GDP 0.883; Private Cons. 0.801; Public Cons. 0.716; Imports 0.626; Exports 0.723; Public Inv. 0.534; Private Inv. 0.398.
  - F-test examples (α = 0, β = 1): 5.495*** (Real GDP), 0.948 (avg. prices), 3.780** (GDP), 10.170*** (Exports).
  - Holden-Peel t-test examples: -0.847 (Real GDP), -2.106** (GDP), -2.338** (Private Inv.).
- Directional accuracy (Full Sample, Table 2)
  - Real GDP Growth: Correct 89.1% (Forecast > 0), 4.5% (Forecast ≤ 0); Incorrect 4.5% (Forecast > 0), 1.8% (Forecast ≤ 0); Chi Square Value 27.557***.
  - Inflation (avg. prices): Chi Square Value 0.189 (no rejection).
  - GDP Growth: Chi Square Value 51.722***.
  - Private Investment Growth: Chi Square Value 5.607**.
- Contributors to GDP forecast errors (Table 3; coefficients are β_j in decomposition regressions)
  - Private Consumption Growth (FE): All 0.429*** (0.076); LIC 0.484*** (0.071); Non-LIC 0.286** (0.108).
  - Public Consumption Growth (FE): All 0.090** (0.041); LIC 0.112** (0.042); Non-LIC 0.091 (0.068).
  - Import Growth (FE): All -0.200*** (0.066); LIC -0.195*** (0.071); Non-LIC -0.253** (0.108).
  - Export Growth (FE): All 0.171*** (0.049); LIC 0.142** (0.056); Non-LIC 0.329*** (0.058).
  - Public Investment Growth (FE): All 0.047*** (0.014); LIC 0.054*** (0.014); Non-LIC 0.031 (0.028).
  - Private Investment Growth (FE): All 0.092*** (0.016); LIC 0.091*** (0.016); Non-LIC 0.123*** (0.040).
  - Observations: 110 (All), 74 (LICs), 36 (Non-LICs). R-squared: 0.488 (All), 0.526 (LICs), 0.612 (Non-LICs).
- Balance of payments highlights
  - Current account (Full Sample observations 132): Mincer-Zarnowitz β for Current Account 0.164 (0.061); Goods Imports 0.705 (0.123); Goods Exports 0.827 (0.084).
  - Financial account (Full Sample observations 61): Mincer-Zarnowitz β for Financial Account 1.025 (0.363); Net Direct Inv. 0.595 (0.020); Other Inv. 1.167 (0.053).
  - Contributors to Current Account forecast errors (Table 6, Panel A): Goods Import Growth (FE): All 2.178*** (0.829); Non-LICs 3.199*** (0.999). Observations: 132 (All), 86 (LICs), 46 (Non-LICs). R-squared: 0.056 (All), 0.050 (LICs), 0.236 (Non-LICs).
  - Financial account regressions show low explanatory power; R-squared reported as low (examples: 0.030 to 0.079).
- Government budget highlights
  - Government expenditure (Full Sample observations 34): Mincer-Zarnowitz β for Gov. Exp. 0.744 (0.183); Interest Exp. 0.778 (0.008); Cap. Exp. & Net Lending 0.513 (0.227).
  - Government revenue (Full Sample observations 69): Mincer-Zarnowitz β for Gov. Revenue 0.866 (0.083); Grants 0.798 (0.036); Tax Revenue 0.817 (0.120); Non-tax Revenue 0.363 (0.077).
  - Contributors to government expenditure forecast errors (Table 9, Panel A): Non-interest Expenditure Growth (FE) 0.644*** (0.066); Cap. Exp. & Lending Growth (FE) 0.254*** (0.020). Observations 34; R-squared 0.895.
  - Contributors to government revenue forecast errors (Table 9, Panel B): Grants Growth (FE) 0.069*** (0.024); Tax Revenue Growth (FE) 0.529* (0.274); Non-tax Revenue Growth (FE) 0.045* (0.025). Observations 69; R-squared 0.415.

### Synthesis and policy-relevant implications
- Strengths of IMF forecasts in crisis programs
  - IMF forecasts for most aggregates contain significant informational value and often outperform naive directional-change models.
  - Real GDP growth forecasts are unbiased; average price growth forecasts are unbiased and efficient.
  - Government revenue growth forecasts, inflation, and reserve asset growth show no systematic bias and are efficient in many specifications.
- Weaknesses and priorities for improvement
  - Lack of efficiency is the primary weakness, especially in LICs.
  - Nominal GDP growth is biased upward (IMF overestimates nominal GDP growth on average).
  - Balance of payments forecasts and many financial-account subcomponent forecasts perform poorly relative to naive models; goods imports forecast errors are the single subcomponent most linked to aggregate BOP errors.
  - Fiscal aggregate forecast errors are driven by subcomponent forecast errors; tax revenue growth forecast errors have the largest economic impact on aggregate revenue errors.
- Practical recommendations (derived from findings)
  - Improve forecasting of private consumption growth, imports, and exports to reduce GDP forecast errors (private consumption largest contributor: coefficient 0.429*** in full sample).
  - Strengthen methods for forecasting balance of payments subcomponents, particularly goods imports (Goods Import Growth (FE): All 2.178*** (0.829)).
  - Improve tax revenue forecasting, given tax revenue growth forecast errors’ large contribution to aggregate revenue errors (Tax Revenue Growth (FE) 0.529* (0.274) in full sample).
  - Address LIC-specific data and information challenges to reduce biases and inefficiencies concentrated in LICs.
  - Incorporate information from data revisions and adjustments in forecast horizons during crises.

### Concluding remarks (overview from section 6)
- IMF forecasts are informative: forecasts of most aggregate macroeconomic variables outperform naive forecasting approaches.
- Lack of efficiency and heterogeneity across LICs versus Non-LICs are key constraints.
- Suggested next steps: incorporate data revisions into forecasting, investigate sources of LIC-specific forecast errors, and pursue further research to refine forecasting approaches for crisis programs.

*Source: wp1848 - 5.1 GDP Growth Forecasts (excerpt from the provided PDF content).*

### 5.1 GDP Growth Forecasts ..............................................................................................1

### wp1848 - 5.1 GDP Growth Forecasts

### Introduction
- Purpose: assess the accuracy of macroeconomic forecasts for countries that experienced crises severe enough to require IMF financial support.
- Evaluation metrics used:
  - Bias: measures forecast deviations from realizations.
  - Efficiency: tests whether forecast errors were unpredictable and whether forecasts incorporate available information.
  - Information content: measures informational value relative to naive forecast models of directional changes.
- Extensions relative to prior work:
  - Use of Mincer and Zarnowitz (1969) and Holden and Peel (1990) regression approaches to examine forecast bias and efficiency.
  - Evaluation of informational value versus naive directional-change models (Merton, 1981; Henriksson and Merton, 1981; Schnader and Stekler, 1990).
  - Examination of IMF crisis forecasts for a much larger set of macroeconomic variables (29).
  - Decomposition of IMF forecast errors of macroeconomic aggregates (GDP, the balance of payments, and the fiscal accounts) to identify subcomponents driving errors.

### Key findings
- Overall informational value:
  - For nearly all variables IMF forecasts contain substantial informational value relative to naive forecasting models.
  - Exceptions: forecasts for inflation, government expenditure growth, and net income growth (likely because these variables tend to trend consistently in one direction for program countries).
- Bias and efficiency (global sample of program countries):
  - Nominal GDP growth is, on average, forecast too optimistically.
  - Current account growth and government expenditure growth are subject to downward bias.
  - Growth forecasts of real GDP, prices, the financial account, and government revenue are found to be unbiased.
  - Forecasts of key indicators of crisis recovery—growth in government expenditure, government revenue, prices, and reserves—are efficient.
- Heterogeneity by country income group:
  - Forecasts for Low-Income Countries (LICs) are substantially more biased and inefficient than for Non-LICs.
  - Possible cause: data and information challenges in LICs during crises.
- Decomposition of forecast errors:
  - GDP growth forecast errors are significantly affected by forecast errors in all GDP subcomponents (government, consumption, investment, net exports).
  - Private consumption growth is the most important contributor to GDP forecast errors.
  - Fiscal budget forecast errors are driven by forecast errors for non-interest/net-lending expenditures, tax/non-tax revenues, and grants.
  - Balance of payments forecast errors cannot, with one exception (goods imports), be linked to forecast errors in its subcomponents.
- Literature context:
  - Prior studies find mixed evidence on systematic biases; some identify IMF forecast model and poor measurement of initial conditions as contributors to bias in crisis countries.

### Macroeconomic identities used for forecasts
- Aggregate demand (nominal GDP decomposition) per IMF (2007):
  - Y = Cp + Cg + Ip + Ig + X - M
  - Growth-rate decomposition (totally differentiated): growth rates of subcomponents enter with elasticities (e.g., σ_{c,y}^p).
- Current account growth identity (IMF, 2015):
  - ca decomposed into growth of goods and services exports (g_x, s_x), imports (g_m, s_m), net income (ni), and net transfers (nt) with associated elasticities σ.
- Financial account growth identity (IMF, 2009):
  - fa = contributions from net foreign direct investment (fdi), net portfolio investment (pi), reserve assets (res), and other investment (ot).
- Government budget decompositions (IMF, 2014):
  - Government expenditure growth (gx) = growth of interest expenditures (int), non-interest expenditures (nint), and capital expenditure & net lending (cap).
  - Government revenue growth (gr) = growth of tax revenue (tax), non-tax revenue (ntax), and grants (grt).

### Methodology: evaluating IMF forecasts
- Mincer-Zarnowitz regressions:
  - Regression linking actual (A_{i}) and forecast (F_{i,p}) values: A_{i,p} = α + β F_{i,p} + θ X_{i,p} + ε_{i,p}.
  - Efficient forecast implies α = 0 and β = 1; joint F-test used to assess efficiency.
  - Controls (X): continental dummies (America, Africa, Asia) and a binary Financial Crisis indicator (Crisis08).
- Holden-Peel test for unbiasedness:
  - Regression of forecast errors on forecasts and controls; t-test for γ = 0 to test unbiasedness.
- Merton-Henriksson timing (directional) tests:
  - 2 x 2 contingency table of forecast sign versus actual sign (Figure 1 in source).
  - χ^2 test used to test independence; rejection implies forecasts have informational value about direction.
- Decomposition of forecast errors:
  - Regress aggregate forecast error on forecast errors of subcomponents (equation (9)):
    - (ŷ_i - ŷ̂_i) = α + Σ_j β_j (x_{ij} - x̂_{ij}) + ε_i
  - Interpretation: a 1% increase in the average forecast error of subcomponent x_j causes a β_j% change in the average forecast error of aggregate y.

### Data: MONA dataset and sample construction
- Data source: IMF’s Monitoring of Fund Arrangements dataset (MONA, IMF 2016b).
- Coverage notes from source:
  - Original MONA dataset covers 238 crisis programs since 2002.
  - Data availability for the broadest macroeconomic identities is limited to 156 programs in 84 countries (see Table A1).
  - Observations are lost due to reporting, measurement and validation discrepancies.
- Outlier treatment:
  - Exclude forecast errors that exceed the respective variable means by four standard deviations.
- Forecasts and realizations timing in MONA:
  - For program approved at year t, MONA contains entries for t-3, t-2, t-1, t, t+1, t+2, t+3, t+4.
  - t is the forecast for the current, first program year; t+1 … t+4 are 1- to 4-year ahead forecasts.
  - This paper focuses on the accuracy of the IMF forecast for year t to maximize sample size.
- Sample inclusion criterion:
  - Include only programs running longer than 18  months to ensure last-review data represent realized values and to allow for data revisions.
- Multiple programs per country:
  - Several countries have more than one IMF program; these are treated as independent observations because programs occur in separate years and under different conditions.
- Timing of program initiation:
  - Programs are initiated at different points throughout the year, but no evidence was found that forecast errors systematically differ across programs due to timing differences.

*Source: wp1848 - 5.1 GDP Growth Forecasts (excerpt from the provided PDF content)*

### 5. IMF Forecast Errors: Decomposition and Determinants

### 5. IMF Forecast Errors: Decomposition and Determinants

### 5.1 GDP Growth Forecasts
- Methodology
  - Employed Mincer-Zarnowitz and Holden-Peel regressions (equations (7) and (8)), Merton-Henriksson timing tests, and an empirical forecast error model (equation (9)).
  - Reported separate results for low-income countries (LICs) and non-LICs (Non-LICs).

- Efficiency and unbiasedness (Mincer-Zarnowitz; Holden-Peel)
  - Full sample:
    - Real GDP forecasts: unbiased but inefficient (joint F-test rejected at the 1 percent level; Holden-Peel cannot reject unbiasedness).
    - Nominal GDP growth and average prices examined separately:
      - Average price growth: unbiased and efficient irrespective of sample.
      - Nominal GDP growth: inefficient and biased in full sample (both F-test and Holden-Peel rejected at the 5 percent level). Intercept estimate and negative Holden-Peel t-statistic indicate IMF overestimates nominal GDP growth on average.
  - Subsamples:
    - Inefficiency of real GDP forecasts is completely driven by the LIC sample.
    - Nominal GDP bias mainly driven by Non-LIC sample; LIC nominal GDP growth forecasts are unbiased and efficient.
    - GDP subcomponents: except for private consumption, forecasts are significantly biased and/or inefficient in the full sample.
      - Biased/inefficient subcomponent forecasts (except public consumption and public investment growth) are driven entirely by the LIC sample.
      - In the Non-LIC sample, only public consumption and public investment components reject efficiency; public investment also shows bias.

- Informational value (Merton-Henriksson)
  - Full sample:
    - χ2-statistics reject the null of independent forecasts and actuals at least at the 5 percent level for all variables except average prices — IMF forecasts generally contain significant informational value.
  - Non-LIC sample:
    - Pattern broadly mirrored; IMF forecasts generally add informational value.
  - LIC sample:
    - Mixed results: IMF forecasts for real GDP, nominal GDP, imports, and public investment contain statistically significant informational value; many other subcomponents and average price growth do not outperform a naive forecast.

- Forecast error contributors (forecast error regression, equation (9); Table 3)
  - Full sample (column 3a):
    - Forecast errors in every GDP subcomponent significantly predict IMF forecast errors in GDP growth.
    - Largest contributors (statistical and economic significance): private consumption, imports, and exports.
  - LIC sample (column 3b):
    - Mirrors full-sample estimates closely.
  - Non-LIC sample (column 3c):
    - Private consumption's role is somewhat subdued relative to LICs.
    - Imports and exports forecast errors take a more prominent role.
    - Forecast errors in public consumption and investment play no role in explaining Non-LIC GDP growth forecast errors.

- Synthesis
  - Confirms earlier literature on bias and/or inefficiency of GDP forecasts; inflation forecasts generally not subject to same caveat.
  - Bias/inefficiency in growth forecasts of real GDP and most subcomponents are driven by the LIC sample.
  - IMF forecasts generally possess significant informational value, though evidence is mixed for LICs.
  - Improving accuracy in private consumption, imports, and exports would most improve aggregate GDP growth forecasts.

### 5.2 Balance of Payments Growth Forecasts
- Context
  - BOP forecasts are key for financial assistance, program design, and assessing progress in closing BOP gaps and rebuilding buffers.
  - Current account growth decomposed into six subcomponents: goods import and export growth, services import and export growth, net transfers growth, net income growth.
  - Financial account growth decomposed into net direct investment, reserve assets, net portfolio investment, and “other investments.”

- Efficiency and unbiasedness (Mincer-Zarnowitz; Holden-Peel)
  - Current account (Table 4a):
    - Full sample (upper panel): joint F-test rejects efficient forecasts at least at the 10 percent level for all current account variables except goods export growth.
    - Holden-Peel: unbiasedness rejected only for current account and net transfers.
    - LIC sample:
      - All current account forecast variables suffer from inefficiency (at least at the 5 percent level).
      - Goods imports and services imports forecasts show significant bias.
    - Non-LIC sample:
      - Forecast efficiency rejected for overall current account and two subcomponents (net transfers and net income).
      - Only net transfers growth estimated with significant bias.
  - Financial account (Table 4b):
    - Full sample:
      - Reject efficient forecast (1 percent significance) for growth in net direct investment, other investment, and net portfolio investment.
      - No evidence of biased forecasts for financial account or subcomponents (Holden-Peel).
    - LIC sample:
      - Inefficient and/or biased forecasts for all four financial-account subcomponents; joint F-test and Holden-Peel reject efficient and unbiased forecasts at least at the 5 percent level, except “other investments” where no bias detected.
      - Aggregate financial account growth forecasts are neither biased nor inefficient (suggesting subcomponent errors may cancel).
    - Non-LIC sample:
      - Similar to full sample, but financial account forecasts now inefficient while net direct investment is not.

- Informational value (Merton-Henriksson)
  - Current account (Table 5a):
    - Full sample: all forecasts except net income contain statistically significant informational value.
    - LIC sample: similar pattern; net income forecasts outperform naive model; net transfers forecasts become statistically indistinguishable from naive.
    - Non-LIC sample: net transfers and net income growth forecasts do not provide statistically significant informational value relative to naive.
  - Financial account (Table 5b):
    - Full sample: only forecasts of financial account growth itself and net direct investment growth add significant informational value.
    - Across subsamples: except for financial account growth in the Non-LIC group, most financial-account subcomponent forecasts do not significantly outperform naive models.
    - Implication: IMF forecasting approach for financial account and subcomponents in crisis countries needs improvement.

- Forecast error contributors (regression approach, Table 6)
  - Full sample:
    - Only one regressor — forecast error in goods import growth — linked to aggregate BOP forecast errors.
    - Large variances in growth rate forecasts for current/financial accounts and subcomponents likely drive limited explanatory power.
  - Conclusion: reinforces need to adjust forecasting approach for balance of payments subcomponents in program countries.

### 5.3 Government Revenue and Expenditure Growth Forecasts
- Structure
  - Government expenditure growth decomposed into growth of interest, non-interest, and capital expenditures.
  - Government revenue growth decomposed into grants, tax revenues, and non-tax revenues.
  - Same battery of tests applied: Mincer-Zarnowitz, Holden-Peel, Merton-Henriksson, and forecast-error regression (equation (9)).

- Efficiency and unbiasedness (Tables 7a and 7b)
  - Government expenditure (Table 7a):
    - Full sample:
      - Joint F-test cannot reject efficient forecasts for aggregate government expenditure growth.
      - Only interest expenditure growth forecasts found inefficient.
      - Holden-Peel indicates significant bias in forecasts for aggregate government expenditure growth.
    - LIC sample:
      - Aggregate government expenditure growth found to be unbiased.
      - Capital expenditure category forecasts are inefficient.
    - Non-LIC sample not reported for expenditure due to only 5 countries (insufficient for reliable inference).
  - Government revenue (Table 7b):
    - Full sample:
      - Joint F-test indicates inefficiency for grants and non-tax revenue growth forecasts.
      - Holden-Peel shows no evidence of bias for aggregate government revenue or subcomponents.
    - LIC sample:
      - Biased and inefficient throughout for aggregate government revenue growth and all subcomponents.
    - Non-LIC sample not reported for revenue due to only 12 observations (insufficient).

- Informational value (Merton-Henriksson; Tables 8a and 8b)
  - Government expenditure (Table 8a):
    - Full sample:
      - IMF forecasts for aggregate government expenditure growth do not contain statistically significant informational value.
      - IMF forecasts for all government expenditure subcomponents significantly outperform naive models.
    - Results are driven by LICs; considerable overlap between full and LIC panels.
    - Non-LIC results omitted due to limited observations.
  - Government revenue (Table 8b):
    - Full and LIC samples:
      - Forecasts for aggregate government revenue growth and all subcomponents significantly outperform naive models.
    - Non-LIC results omitted due to only 12 observations.

- Forecast error contributors (Table 9; equation (9))
  - Aggregate government expenditure growth (panel A):
    - Forecast errors in non-interest expenditures, capital expenditures, and net lending are significant drivers of aggregate expenditure forecast errors.
    - LIC results similar.
  - Aggregate government revenue growth (panel B):
    - Forecast errors in all subcomponents significantly contribute to aggregate revenue forecast errors in the full sample.
    - Tax revenue growth forecast errors have by far the greatest economic impact.
    - In LICs, forecast errors in non-tax revenue growth do not matter for aggregate revenue forecast errors; otherwise LIC results resemble full sample.

- Synthesis
  - Several fiscal aggregate and subcomponent forecasts, particularly on the revenue side, are biased and/or inefficient, but IMF forecasts still add informational value.
  - Forecast errors in most fiscal subcomponents feed into aggregate fiscal forecast errors in crisis countries.
  - Improving accuracy in fiscal subcomponent forecasts, especially tax revenue growth, would help improve aggregate fiscal balance forecasts.

*Source: 5. IMF Forecast Errors: Decomposition and Determinants (wp1848).*

### 6. Concluding Remarks

### 6. Concluding Remarks

### Overview
- Examination of IMF forecasts for countries experiencing severe economic crises and requiring access to IMF lending facilities.
- Evaluation focuses on bias, efficiency, and information content of IMF forecasts.
- Uses accounting identities to disentangle contributions of subcomponents’ forecast errors to forecast errors of key macroeconomic aggregates.

### Key findings on forecast performance
- IMF forecasts are informative: forecasts of most aggregate macroeconomic variables outperform naive forecasting approaches.
- IMF forecasts are unbiased (no significant deviations of forecasts from realizations) and efficient (all information available at the time of forecasts is used) for:
  - government revenue growth
  - inflation
  - reserve asset growth
- Real GDP growth forecasts are unbiased.
- Government expenditure growth forecasts are efficient.
- Nominal GDP and balance of payments forecasts show scope for improvements.

### Heterogeneity and weaknesses
- Lack of efficiency is identified as the weakest link in IMF forecasts.
- Significant heterogeneity in forecast accuracy between low-income countries (LICs) and Non-LICs.
- In most cases, biases and inefficiencies in the global sample can be traced directly to forecast errors originating in LICs.

### Recommendations and avenues for improvement
- Incorporate information from data revisions and adjustments in forecast horizons to help increase forecast accuracy during times of crises.
- Further inquiry and research are encouraged to build on these findings.

*Source: wp1848 - 6. Concluding Remarks*

### References

### References

### Key bibliographic sources
- Atoyan, Rouben, Patrick Conway, Marcelo Selowsky and Tsidi Tsikata, 2004. “Macroeconomic Adjustments in IMF-supported Programs: Projections and Reality,” IEO Background Paper BP/04/2, International Monetary Fund.
- Atoyan, Rouben and Patrick Conway, 2011. “Projecting Macroeconomic Outcomes: Evidence from the IMF,” Review of International Organizations, 6 (September), 415-441.
- Baqir, Reza, Rodney Ramcharan, and Ratna Sahay, 2005. “IMF Programs and Growth: Is Optimism Defensible?” IMF Staff Papers, 52 (2), 260-286.
- Blanchard, Olivier and Daniel Leigh, 2013. “Growth Forecast Errors and Fiscal Multipliers,” IMF Working Paper WP/13/1, International Monetary Fund.
- Castle, Jennifer L., David F. Hendry and Oleg I. Kitov, forthcoming. “Forecasting and Nowcasting Macroeconomic Variables: A Methodological Overview,” Handbook on Rapid Estimates.
- Ghosh, Atish, Charis Christofides, Jun Kim, Laura Papi, Uma Ramakrishnan, Alun Thomas and Juan Zalduendo, 2005. “The Design of IMF-Supported Programs,” Occasional Paper 241, International Monetary Fund.
- IMF manuals and datasets cited: “The System of Macroeconomic Accounts Statistics: An Overview,” Pamphlet Series, No. 56 (2007); “Balance of Payments and International Investment Position Manual,” Sixth Edition (BPM6) (2009); “Government Finance Statistics Manual,” Sixth Edition (2014); “Balance of Payment and International Investment Position Statistics” (2015); “IMF Crisis Lending Fact Sheet” (2016a); “Monitoring of Fund Arrangements (MONA) Dataset” (2016b).
- Other methodological and evaluation sources include works by Henriksson and Merton (1981), Holden and Peel (1990), Mincer and Zarnowitz (1969), Sinclair et al., Joutz and Stekler, Lahiri and Sheng, Musso and Phillips, Palm and Zellner, Silver (2012), Swanson and van Dijk (2006), and others listed.

### Empirical findings — Forecast evaluation (Mincer-Zarnowitz and related tests)
- Full Sample Mincer-Zarnowitz regression results (dependent variable: Actual growth rate):
  - Forecast growth rate, β: real GDP 0.606; avg. prices 1.031; GDP 0.883; Private Cons. 0.801; Public Cons. 0.716; Imports 0.626; Exports 0.723; Public Inv. 0.534; Private Inv. 0.398.
  - Constants, α: range from -0.023 to 0.036 (specific entries provided).
  - Observations: 110 for each column reported.
  - R-squared: 0.463 (Real GDP) up to 0.785 (GDP).
  - F-test (α = 0, β = 1): e.g., 5.495*** (Real GDP), 0.948 (avg. prices), 3.780** (GDP), 10.170*** (Exports).
  - Holden-Peel t-test: e.g., -0.847 (Real GDP), -2.106** (GDP), -2.338** (Private Inv.).

- LIC Sample (observations 74) and Non-LIC Sample (observations 36) Mincer-Zarnowitz coefficients and diagnostics are reported separately with:
  - LIC: β for Real GDP 0.482; GDP 0.914; Private Cons. 0.523; R-squared up to 0.782; F-test examples: 7.600*** (Real GDP), 11.060*** (Private Cons.).
  - Non-LIC: β for Real GDP 1.069; GDP 0.951; Private Cons. 0.895; R-squared up to 0.934 (avg. prices); F-test examples: 4.141** (Imports), 4.882** (Public Inv.).

- Notes: Robust standard errors in parentheses. ***, ** and * indicate rejection of the null hypothesis of the Mincer-Zarnowitz F-test and the Holden-Peel t-test at the 10, 5 and 1 percent level of statistical significance, respectively.

### Forecast correctness (directional accuracy)
- Table 2 — Full Sample (Real GDP Growth and subcomponents): proportions of correct forecasts when Forecast > 0 / Forecast ≤ 0 and incorrect splits:
  - Real GDP Growth: Correct 89.1% (Forecast > 0), 4.5% (Forecast ≤ 0); Incorrect 4.5% (Forecast > 0), 1.8% (Forecast ≤ 0); Chi Square Value 27.557***.
  - Inflation (avg. prices): Correct 89.6% (Forecast > 0), 0.0% (Forecast ≤ 0); Incorrect 8.5%, 1.9%; Chi Square Value 0.189.
  - GDP Growth: Correct 94.5% / 3.6%; Incorrect 0.9% / 0.9%; Chi Square Value 51.722***.
  - Public Investment Growth: Correct 67.3% / 8.2%; Incorrect 19.1% / 5.5%; Chi Square Value 7.566***.
  - Private Investment Growth: Correct 70.0% / 7.3%; Incorrect 13.6% / 9.1%; Chi Square Value 5.607**.
- LIC and Non-LIC samples reported analogous directional accuracy tables with specific percentages and Chi Square values (examples: LIC Real GDP Growth Correct 93.2% / 2.7% with Chi Square 12.161***; Non-LIC Real GDP Growth Correct 80.6% / 8.3% with Chi Square 8.528***).

### Contributors to GDP forecast errors (Table 3)
- Regression of GDP growth forecast errors on subcomponent forecast errors (All, LICs, Non-LICs):
  - Private Consumption Growth (FE): All 0.429*** (0.076); LIC 0.484*** (0.071); Non-LIC 0.286** (0.108).
  - Public Consumption Growth (FE): All 0.090** (0.041); LIC 0.112** (0.042); Non-LIC 0.091 (0.068).
  - Import Growth (FE): All -0.200*** (0.066); LIC -0.195*** (0.071); Non-LIC -0.253** (0.108).
  - Export Growth (FE): All 0.171*** (0.049); LIC 0.142** (0.056); Non-LIC 0.329*** (0.058).
  - Public Investment Growth (FE): All 0.047*** (0.014); LIC 0.054*** (0.014); Non-LIC 0.031 (0.028).
  - Private Investment Growth (FE): All 0.092*** (0.016); LIC 0.091*** (0.016); Non-LIC 0.123*** (0.040).
  - Constant: All -0.006* (0.003); LIC -0.010*** (0.004); Non-LIC 0.012* (0.006).
  - Observations: 110 (All), 74 (LICs), 36 (Non-LICs).
  - R-squared: 0.488 (All), 0.526 (LICs), 0.612 (Non-LICs).
- Notes: All variables are forecast errors of growth rates. Robust standard errors in parenthesis. ***, ** and * indicate 1, 5 and 10 percent level of statistical significance, respectively.

### Balance of payments: Mincer-Zarnowitz and correctness
- Table 4a (Current Account and subcomponents, Full Sample observations 132):
  - Forecast growth rate, β for Current Account 0.164 (0.061); Goods Imports 0.705 (0.123); Goods Exports 0.827 (0.084); Services Imports 0.576 (0.105); Services Exports 0.276 (0.192); Net Transfers 0.083 (0.107); Net Income 0.247 (0.098).
  - R-squared ranges from 0.023 (Net Transfers) to 0.432 (Goods Exports).
  - F-test (α = 0, β = 1) examples: 95.560*** (Current Account), 2.870* (Goods Imports), 9.125*** (Services Imports).
  - Holden-Peel t-test examples: 0.421** (Current Account), 2.018 (Services Imports), 2.215** (Net Transfers).
- Table 4b (Financial Account and subcomponents, Full Sample observations 61):
  - Forecast growth rate, β: Financial Account 1.025 (0.363); Net Direct Inv. 0.595 (0.020); Reserve Assets 0.741 (0.475); Other Inv. 1.167 (0.053); Net Portfolio Inv. -0.144 (0.238).
  - R-squared examples: 0.599 (Financial Account), 0.782 (Net Direct Inv.), 0.743 (Other Inv.).
  - F-test examples: 0.469 (Financial Account), 243.500*** (Net Direct Inv.), 5.211*** (Other Inv.).
- Directional correctness (Table 5a and 5b):
  - Full Sample Current Account Growth: Correct 34.1% (Forecast > 0) and 34.1% (Forecast ≤ 0); Incorrect 19.7% and 12.1%; Chi Square Value 16.754***.
  - Goods Import Growth: Correct 66.7% / 19.7%; Incorrect 9.8% / 3.8%; Chi Square 54.081***.
  - Financial Account Growth (Table 5b Full Sample): Correct 23.0% / 42.6%; Incorrect 13.1% / 21.3%; Chi Square 4.079**.
  - Subsample splits (LIC / Non-LIC) reported with specific percentages and Chi Square values (examples: LIC Current Account Growth Chi Square 5.544**; Non-LIC Current Account Growth Chi Square 7.998***).

### Contributors to Balance of Payments forecast errors (Table 6)
- Panel A: Current Account contributors (All, LICs, Non-LICs):
  - Goods Import Growth (FE): All 2.178*** (0.829); LIC 1.555 (0.992); Non-LICs 3.199*** (0.999).
  - Goods Export Growth (FE): All -0.664 (0.542); LIC -0.373 (0.773); Non-LIC -1.342 (0.952).
  - Services Import Growth (FE): All -0.254 (0.748); LIC -0.709 (1.254); Non-LIC 0.928 (0.953).
  - Services Export Growth (FE): All -0.657 (0.527); LIC -0.656 (0.661); Non-LIC -0.366 (0.620).
  - Net Transfers Growth (FE): All 0.133 (0.491); LIC 0.412 (1.027); Non-LIC 0.181 (0.167).
  - Net Income Growth (FE): All 0.005 (0.082); Non-LICs -0.337** (0.151).
  - Constant: All 0.066 (0.155); Non-LIC 0.243** (0.118).
  - Observations: 132 (All), 86 (LICs), 46 (Non-LICs). R-squared: 0.056, 0.050, 0.236 respectively.
- Panel B: Financial Account contributors (All, LICs, Non-LICs):
  - Net Direct Investment Growth (FE): All 0.299 (0.386); LIC 0.316 (0.505); Non-LIC 0.202 (0.570).
  - Reserve Assets Growth (FE): All 0.001 (0.046); LIC -0.094 (0.061); Non-LIC 0.069 (0.050).
  - Other Investment Growth (FE): All -0.011 (0.008); LIC -0.014 (0.014); Non-LIC -0.005 (0.011).
  - Net Portfolio Investment Growth (FE): All 0.008 (0.062); LIC -0.026 (0.095); Non-LIC 0.027 (0.105).
  - Constant: All -0.298 (0.311).
  - Observations: 61 (All), 34 (LICs), 27 (Non-LICs). R-squared low (0.030 to 0.079 to 0.061 reported in table context).

### Government budget forecasts: Mincer-Zarnowitz, correctness, and contributors to errors
- Government expenditure growth (Table 7a, Full Sample observations 34):
  - Forecast growth rate, β: Gov. Exp. 0.744 (0.183); Interest Exp. 0.778 (0.008); Non-interest Exp. 0.922 (0.176); Cap. Exp. & Net Lending 0.513 (0.227).
  - R-squared: 0.451 (Gov. Exp.) to 0.979 (Interest Exp.).
  - F-test examples: 3.094 (Gov. Exp.), 405.400*** (Interest Exp.).
  - Holden-Peel t-test: 2.078** (Gov. Exp.).
- Government revenue growth (Table 7b, Full Sample observations 69):
  - Forecast growth rate, β: Gov. Revenue 0.866 (0.083); Grants 0.798 (0.036); Tax Revenue 0.817 (0.120); Non-tax Revenue 0.363 (0.077).
  - R-squared: 0.775 (Gov. Revenue) to 0.899 (Grants).
  - F-test examples: 1.345 (Gov. Revenue), 35.610*** (Non-tax Revenue).
- Correct/Incorrect directional forecasts (Table 8a and 8b, Full Sample):
  - Gov. Expenditure Growth: Correct 79.4% (Forecast > 0) / 5.9% (Forecast ≤ 0); Incorrect 11.8% / 2.9%; Chi Square 2.370.
  - Tax Revenue Growth: Correct 85.5% / 8.7%; Incorrect 1.4% / 4.3%; Chi Square 29.493***.
  - Grants Growth: Correct 52.2% / 15.9%; Incorrect 24.6% / 7.2%; Chi Square 5.419**.
- Contributors to government budget forecast errors (Table 9):
  - Panel A (Gov. expenditure FE regressors, All LICs):
    - Interest Expenditure Growth (FE): 0.003 (0.008).
    - Non-interest Expenditure Growth (FE): 0.644*** (0.066).
    - Cap. Exp. & Lending Growth (FE): 0.254*** (0.020).
    - Observations 34; R-squared 0.895.
  - Panel B (Gov. revenue FE regressors, All LICs):
    - Grants Growth (FE): 0.069*** (0.024) for All; 0.082** (0.033) for LICs.
    - Tax Revenue Growth (FE): 0.529* (0.274) for All; 0.513* (0.288) for LICs.
    - Non-tax Revenue Growth (FE): 0.045* (0.025) for All; 0.043 (0.028) for LICs.
    - Observations 69 (All), 57 (LICs); R-squared 0.415 (All), 0.422 (LICs).

### Appendix — Programs in broadest global sample (selected entries)
- The broadest global sample lists country/program entries with dates; examples include:
  - Afghanistan June 2006; Afghanistan November 2011.
  - Albania June 2002; Albania January 2006.
  - Argentina September 2003; Armenia May 2005; Armenia March 2009; Armenia June 2010; Armenia March 2014.
  - Bangladesh June 2003; Bangladesh April 2012.
  - Burkina Faso June 2003; Burkina Faso April 2007; Burkina Faso June 2010; Burkina Faso December 2013.
  - Ghana May 2003; Ghana July 2009; Ghana April 2015.
  - Greece May 2010; Greece March 2012.
  - Mozambique July 2004; Mozambique June 2007; Mozambique June 2010; Mozambique June 2013.
  - Tanzania March 2000; Tanzania August 2003; Tanzania February 2007; Tanzania June 2010; Tanzania July 2012; Tanzania July 2014.
  - Uganda September 2002; Uganda December 2006; Uganda May 2010; Uganda June 2013.
  - Additional country/program date pairs are listed comprehensively in the appendix table.

*This content unit corresponds to wp1848 - References (pdf).*

---


_Source: https://www.imf.org/-/media/files/publications/wp/2018/wp1848.pdf_