## 2. Significant Domestic and Global Drivers

## Source details

**Canonical URL:** [2. Significant Domestic and Global Drivers](https://www.imf.org/-/media/files/publications/wp/2021/english/wpiea2021015-print-pdf.pdf)

## Other formats

- [Markdown version](/-/media/files/publications/wp/2021/english/wpiea2021015-print-pdf.pdf.md)
- [Structured JSON version](/-/media/files/publications/wp/2021/english/wpiea2021015-print-pdf.pdf.json)

---

### Purpose and scope
- The IMF can provide financial assistance for eligible low-income countries through the Poverty Reduction and Growth Trust (PRGT), which assists eligible countries in achieving and maintaining a stable and sustainable macroeconomic position consistent with strong and durable poverty reduction and growth (IMF, 2018).
- Given increased demand for IMF concessional financing in recent years prior to the COVID-19 pandemic and the relative scarcity of these resources, it is important to understand the factors behind the demand for concessional financing and to forecast its demand.
- This analysis is based on pre-pandemic factors.
- The study started in summer 2019 and was completed before the COVID-19 pandemic that started in early 2020.
- Findings and conclusions on factors affecting IMF concessional demand are based on historical data as of end-2018 (i.e., on pre-pandemic factors).

### Research gaps and contributions
- Gaps in existing literature:
  - Lack of analysis of country-specific factors that influence concessional borrowing and use of these factors to predict a country’s future concessional financing demand.
  - Lack of use of country-specific factors to predict aggregate demand for concessional financing.
- Contributions of this paper:
  - Applies a factor-augmented probit model to find variables that statistically replicate factors influencing concessional borrowing for a PRGT-eligible country.
  - Uses identified factors to predict a country’s future demand for concessional financing.
  - Uses country-specific predictions and a balanced statistical approach to estimate aggregate demand for currently PRGT-eligible countries.

### Research questions
- (1) What are the main country-specific factors of a country’s demand for concessional financing?
- (2) What is the probability that the country would request concessional financing in the future?
- (3) How much is the predicted annual aggregate demand for concessional financing?

### Executive summary — sample and stylized facts
- Estimation sample: 64 countries, 53 PRGT-eligible and 11 previously PRGT-eligible, covering the period 1986-2018.
- Total arrangements in the sample: 391 arrangements.
- Typical PRGT arrangement: usually more than three years, approved amount about SDR 65 million (equivalent to 55 percent of quota).
- Key finding: country-specific domestic factors dominate the prediction of demand for IMF concessional (PRGT) financing; global factors add limited predictive power for most countries.
- Main country-level predictors (most frequent): external debt, inflation, and the real effective exchange rate (REER).
- Aggregate forecasting: a qualitative classification into four risk regions with calibrated probability thresholds of 0.2252, 0.5383, and 0.7710 is used to forecast annual aggregate demand via a beta regression.

### Stylized facts and simple correlations (selected)
- Data sources: WEO, IFS, FFA, Haver, ICRG, EM-DAT.
- Sample distribution skewed toward African countries.
- Simple OLS correlations linking number of approved financing arrangements to prior economic variables (coefficients and standard errors):
  - Lagged 4-year current account balance/GDP growth: coefficient -0.39*** (standard error 0.12)
  - 4-year lagged external debt/GDP: coefficient 0.07*** (standard error 0.03)
  - 3-year lagged reserves: coefficient -0.00*** (standard error 0.00)
  - Cumulative sum of growth of volatility index (VIX): coefficient 1.68*** (standard error 0.47)
  - Lagged world GDP growth: coefficient -0.63* (standard error 0.31)
  - 3-year lagged commodity price index: coefficient -0.03** (standard error 0.01)
  - Cumulative fiscal balance changes over the last 3 years: coefficient -0.88** (standard error 0.45)
  - 5-year cumulative sum of growth of real GDP per capita: coefficient -0.25*** (standard error 0.11)
  - 4-year lagged consumer price index: coefficient -0.005*** (standard error 0.02)

### PRGT arrangements summary (Table 1 highlights)
- Number of programs (by facility): PRGT Arrangements Only 62; ECF 86; ESAF 11; ESF 71; PRGF 23; SAF 5; Total/Median 260
- Original duration (months): most around 36; SAF 18
- Actual duration (months): median 38; some facilities show actual longer than original (e.g., ECF actual 41 vs original 36)
- Approved amount (percent of quota): PRGT 66; ECF 67; ESAF 67; ESF 45; PRGF 26; SAF 144; Total/Median 55
- Approved amount (SDR million): PRGT 80; ECF 72; ESAF 114; ESF 51; PRGF 26; SAF 144; Total/Median 65

### Model description
- Two-stage estimation:
  1. Estimate a single dynamic common factor from a candidate set of variables for each country using a dynamic factor model.
  2. Use the estimated factor as the primary explanatory variable in a country-specific probit model predicting the binary outcome of initiating a PRGT arrangement.
- Candidate domestic factors: GDP per capita, inflation, current account balance, fiscal balance, government public debt, reserves, terms-of-trade, capital flows, REER.
- Candidate global factors: commodity price index, non-fuel and fuel price indices, oil price, world GDP.
- Estimation choices and robustness:
  - Single dynamic common factor per country; independent AR(2) processes for dynamic factor and idiosyncratic shocks.
  - Variable selection balances estimation precision, predictive power, information (lowest information criteria), and statistical significance (p ≤ 0.1).
  - Robustness checks: three factor-extraction methods (principal component; Kim and Nelson (1999) Kalman/Carter-Kohn; Otrok and Whiteman (1998) Gibbs sampling) and alternative demand definitions (demand dummy = 1 at arrangement start vs 1 for entire arrangement duration).
  - Baseline calibration controls for years when a country participated in HIPC or MDRI by setting program dummy = 1.

### Results — country-level determinants
- Domestic factors dominate country-specific forecasts:
  - Most frequent significant domestic factors (appearing in >30 percent of country-specific forecasts): external debt, inflation, REER.
    - Caveat: 16 of 64 countries lack a REER series; importance of REER could be understated.
  - Next-most frequent (~25 percent of sample): capital flows, terms-of-trade, international reserves.
  - Least significant in country-specific models: current account balance, fiscal balance, GDP per capita growth (despite their importance in simple correlations at the aggregate level).
- Global factors:
  - Mostly statistically insignificant across baseline and robustness checks for over 60 percent of countries.
  - World GDP growth and the U.S. financial market volatility index (VIX) are the most influential among global factors where significance is found.
  - Commodity prices (fuel/nonfuel) generally insignificant in country-specific forecasting—possibly because their informative content is captured by domestic variables such as terms-of-trade or external debt.
- Institutional variables:
  - Some country-level institutional indicators are statistically significant predictors when included (examples with sign and significance):
    - Benin: Natural disaster economic impact score, sign Negative, significance *
    - Ghana: Risk of democratic accountability, sign Positive, significance *
    - Guinea: Risk for foreign debt service, sign Positive, significance *
    - Côte d'Ivoire: Risk of law and order, sign Positive, significance *
    - Kenya: Risk of annual inflation rate, sign Positive, significance **
    - Madagascar: Risk of ethnic tensions, sign Negative, significance **
    - Mali: Risk of net international liquidity, sign Positive, significance *
    - Niger: Risk of exchange rate stability, sign Positive, significance *
    - Senegal: Risk of investment profile, sign Positive, significance **
    - Sierra Leone: Risk of exchange rate stability, sign Negative, significance *
  - Interpretation caveat: institutional indices are constructed so that higher values denote lower risk; data availability for institutional variables is limited and short in time coverage.

### Regional patterns
- External debt is frequently ranked as the most important factor across regions, except in Latin America where fiscal balance is more prominent.
- Inflation is among the top three factors in most regions.
- REER appears among the top five factors in most regions (data permitting).
- Global factors remain generally less important at regional levels, although world GDP growth and VIX have regional relevance (e.g., VIX is significant for Asia and Pacific).

### Forecasting capability and aggregate forecasts
- Country-specific forecasts:
  - The model produces probabilistic forecasts where higher point estimates align with historical approval periods and lower probabilities align with non-arrangement periods for illustrated country cases.
- Qualitative classification framework:
  - Constructed J = 4 risk regions with calibrated probability thresholds: 0.2252, 0.5383, 0.7710 that minimize RMSE and maximize Pseudo R2 in a beta regression linking the share of countries in each risk class to the annual share of countries in IMF concessional arrangements.
- Beta regression for aggregate demand (Table 4 results):
  - Dependent variable: annual share of countries in IMF concessional financing arrangements.
  - Coefficients (standard errors):
    - Constant, β0: -1.83*** (0.28)
    - Severe risk group (p̂ ≥ 0.7710), β1: 11.39*** (2.18)
    - High risk (0.7710 > p̂ ≥ 0.5383), β2: -5.07 (3.34)
    - Moderate risk (0.5383 > p̂ ≥ 0.2252), β3: 1.72 (1.73)
    - Low risk (0.2252 > p̂), β4: -1.13*** (0.45)
    - Precision parameter, ϕ: 107.06*** (27.22)
  - Pseudo R2: 0.5580
  - RMSE: 0.0297
  - Number of observations (years): 31
  - Significance notation: *** p ≤ 0.01, ** p ≤ 0.05, * p ≤ 0.1
- Model performance:
  - Aggregate forecasts (within-sample up to 2018 and out-of-period for 2019) capture actual annual aggregate demand within 95 percent confidence intervals for each year in the sample used.
  - Note: only PRGT-eligible countries with at least one significant predictor (p-value ≤ 0.1) are included each year.

### Model selection and performance metrics (Appendix 3 highlights)
- Variable selection criteria:
  - Predictive power: Brier’s quadratic probability score (QPS) — desirable to be close to zero.
  - Model precision: confidence score (CS) from 95 percent confidence bands — lower CS desirable.
  - Informational quality: Akaike Information Criterion (AIC) normalized by number of periods T.
- Weights set for selection: ω_QPS = ω_CS = ω_AIC = 1/3.
- Inclusion threshold: Only models with p-value at most 0.1 were included in final analyses.
- Performance and summary statistics for selected models (across all countries, regardless of significance):
  - Average QPS = 0.064
  - Average CI = 0.099
  - Average AIC = 0.626
  - Average Pseudo R^2 = 0.443

### Conclusions and policy implications
- Main empirical conclusions:
  - Country-specific domestic variables—especially external debt, inflation, and REER—are the most informative predictors of demand for PRGT concessional financing.
  - Global variables generally do not materially improve country-level predictive power for most countries; where they matter, world GDP growth and VIX are the most relevant.
  - Institutional variables can be significant predictors in some country cases but are constrained by data availability and potential informational redundancy with domestic variables.
  - The two-step factor-augmented probit framework provides significant capability to predict historical initiation of PRGT arrangements at the country level and supports construction of aggregate demand forecasts via a qualitative-to-beta-regression approach.
- Policy recommendations implied by findings:
  - Maintaining sustainable external debt and reducing inflation should help reduce demand for concessional lending.
  - Strengthening fiscal positions, improving GDP growth, and enhancing governance could potentially lower the need for PRGT financing.
  - Country-specific diagnostics and forecasting can inform allocation and management of the IMF’s limited concessional resources.

### Suggested extensions for future research
- Apply the approach to IMF non-concessional financing demand.
- Examine the impact of the COVID-19 pandemic on demand (current study uses pre-pandemic data as of end-2018).
- Simplify estimation by exploring dynamic factor models with Markov-switching.

*Italic: Source — IMF staff paper (estimation sample and results based on data through end-2018).*

### 2. Significant Domestic and Global Drivers ______________________________________12

### 2. Significant Domestic and Global Drivers

### Purpose and scope
- The IMF can provide financial assistance for eligible low-income countries through the Poverty Reduction and Growth Trust (PRGT), which assists eligible countries in achieving and maintaining a stable and sustainable macroeconomic position consistent with strong and durable poverty reduction and growth (IMF, 2018).
- Given increased demand for IMF concessional financing in recent years prior to the COVID-19 pandemic and the relative scarcity of these resources, it is important to understand the factors behind the demand for concessional financing and to forecast its demand.
- This analysis is based on pre-pandemic factors.
- The study started in summer 2019 and was completed before the COVID-19 pandemic that started in early 2

### Research gaps and contributions
- Existing literature has examined links between domestic conditions and the IMF’s financing for a typical developing country, but has not focused on:
  - analyzing country-specific factors that influence concessional borrowing and using these factors to predict a country’s future concessional financing demand;
  - using country-specific factors to predict aggregate demand for concessional financing.
- This paper makes three contributions:
  1. Applies a factor-augmented probit model to find the set of variables that are statistically accurate in replicating the factors that have influenced concessional borrowing for a PRGT-eligible country.
  2. Uses the identified set of factors to predict the country’s future demand for concessional financing.
  3. Uses the country-specific prediction and a balanced statistical approach to estimate aggregate demand for the currently PRGT-eligible countries.

### Research questions
- The paper aims to address the following questions:
  (1) What are the main country-specific factors of a country’s demand for concessional financing?
  (2) What is the probability that the country would request concessional financing in the future?
  (3) How much is the predicted annual aggregate demand for concessional financing?

### Main findings (as reported)
- The paper reports three main findings. The first finding (partially quoted in the source) is:
  - The main country-specific factors for demand vary among PRGT-eligible countries. The external debt level, inflation, and real

*Source: wpiea2021015-print-pdf - 2. Significant Domestic and Global Drivers ______________________________________12*

### 2020. Its findings and conclusions on factors affecting IMF concessional demand are based on historical data as

### wpiea2021015-print-pdf - 2020. Its findings and conclusions on factors affecting IMF concessional demand are based on historical data as of end-2018 (i.e., on pre-pandemic factors)

### Executive summary
- Estimation sample: 64 countries, 53 PRGT-eligible and 11 previously PRGT-eligible, covering the period 1986-2018.
- Total arrangements in the sample: 391 arrangements.
- Typical PRGT arrangement: usually more than three years, approved amount about SDR 65 million (equivalent to 55 percent of quota).
- Key finding: country-specific domestic factors dominate the prediction of demand for IMF concessional (PRGT) financing; global factors add limited predictive power for most countries.
- Main country-level predictors (most frequent): external debt, inflation, and the real effective exchange rate (REER).
- Aggregate forecasting: a qualitative classification into four risk regions with calibrated probability thresholds of 0.2252, 0.5383, and 0.7710 is used to forecast annual aggregate demand via a beta regression.

### Literature and motivation
- Earlier cross-section and panel studies link IMF financing demand to variables such as current account balance, inflation, GNP/GDP per capita, imports, reserves, external debt service, and capital flows.
- Recent low-income–country–focused studies (Bird and Rowlands (2009); Bal Gündüz (2009)) find domestic and some global variables (commodity prices, world trade) significant.
- Two gaps highlighted by the paper:
  - Variables significant on average across developing countries do not necessarily hold for a specific country.
  - Variables with historical correlations do not necessarily have the highest predictive power for forecasting a specific country’s future concessional financing demand.

### Stylized facts and data
- Data sources: World Economic Outlook (WEO), International Financial Statistics (IFS), Financial Flow Analysis (FFA), Haver, International Country Risk Guide (ICRG), EM-DAT.
- Sample distribution skewed toward African countries.
- Simple OLS correlations (Table 2) linking the number of approved financing arrangements to prior economic variables:
  - Lagged 4-year current account balance/GDP growth: coefficient -0.39*** (standard error 0.12)
  - 4-year lagged external debt/GDP: coefficient 0.07*** (standard error 0.03)
  - 3-year lagged reserves: coefficient -0.00*** (standard error 0.00)
  - Cumulative sum of growth of volatility index (VIX): coefficient 1.68*** (standard error 0.47)
  - Lagged world GDP growth: coefficient -0.63* (standard error 0.31)
  - 3-year lagged commodity price index: coefficient -0.03** (standard error 0.01)
  - Cumulative fiscal balance changes over the last 3 years: coefficient -0.88** (standard error 0.45)
  - 5-year cumulative sum of growth of real GDP per capita: coefficient -0.25*** (standard error 0.11)
  - 4-year lagged consumer price index: coefficient -0.005*** (standard error 0.02)
- Table 1 summary for PRGT arrangements in sample:
  - Number of programs (by facility): PRGT Arrangements Only 62; ECF 86; ESAF 11; ESF 71; PRGF 23; SAF 5; Total/Median 260
  - Original duration (months): most around 36; SAF 18
  - Actual duration (months): median 38; some facilities show actual longer than original (e.g., ECF actual 41 vs original 36)
  - Approved amount (percent of quota): PRGT 66; ECF 67; ESAF 67; ESF 45; PRGF 26; SAF 144; Total/Median 55
  - Approved amount (SDR million): PRGT 80; ECF 72; ESAF 114; ESF 51; PRGF 26; SAF 144; Total/Median 65

### Model description
- Two-stage estimation: (i) estimate a single dynamic common factor from a candidate set of variables for each country using a dynamic factor model; (ii) use the estimated factor as the primary explanatory variable in a country-specific probit model predicting the binary outcome of initiating a PRGT arrangement.
- Candidate domestic factors: GDP per capita, inflation, current account balance, fiscal balance, government public debt, reserves, terms-of-trade, capital flows, real effective exchange rate (REER).
- Candidate global factors: commodity price index, non-fuel and fuel price indices, oil price, world GDP.
- Estimation choices and robustness:
  - Single dynamic common factor per country (to remain consistent across varying country variable sets and limited sample length).
  - Independent AR(2) processes for dynamic factor and idiosyncratic shocks.
  - Selection of variable sets balances four criteria: estimation precision, predictive power, information (lowest information criteria), and statistical significance (p ≤ 0.1).
  - Robustness checks include three estimation approaches for factor extraction (principal component; Kim and Nelson (1999) Kalman/Carter-Kohn; Otrok and Whiteman (1998) Gibbs sampling) and alternative definitions of demand (demand dummy = 1 at arrangement start vs 1 for entire arrangement duration).
  - Baseline calibration controls for years when a country participated in HIPC or MDRI by setting program dummy = 1.

### Results — country-level determinants
- Domestic factors dominate country-specific forecasts:
  - Most frequent significant domestic factors (appearing in >30 percent of country-specific forecasts): external debt, inflation, REER.
    - Caveats: 16 of 64 countries lack a REER series; importance of REER could be understated.
  - Next-most frequent (~25 percent of sample): capital flows, terms-of-trade, international reserves.
  - Least significant in country-specific models: current account balance, fiscal balance, GDP per capita growth (despite their importance in simple correlations at the aggregate level).
- Global factors:
  - Mostly statistically insignificant across the baseline and robustness checks for over 60 percent of countries.
  - World GDP growth and the U.S. financial market volatility index (VIX) are the most influential among global factors where significance is found.
  - Commodity prices (fuel/nonfuel) generally insignificant in country-specific forecasting—possibly because their informative content is captured by domestic variables such as terms-of-trade or external debt.
- Institutional variables:
  - Some country-level institutional indicators are statistically significant predictors when included (Table 3 entries):
    - Benin: Natural disaster economic impact score, sign Negative, significance *
    - Ghana: Risk of democratic accountability, sign Positive, significance *
    - Guinea: Risk for foreign debt service, sign Positive, significance *
    - Côte d'Ivoire: Risk of law and order, sign Positive, significance *
    - Kenya: Risk of annual inflation rate, sign Positive, significance **
    - Madagascar: Risk of ethnic tensions, sign Negative, significance **
    - Mali: Risk of net international liquidity, sign Positive, significance *
    - Niger: Risk of exchange rate stability, sign Positive, significance *
    - Senegal: Risk of investment profile, sign Positive, significance **
    - Sierra Leone: Risk of exchange rate stability, sign Negative, significance *
  - Interpretation caveat: institutional indices are constructed so that higher values denote lower risk; data availability for institutional variables is limited and short in time coverage.

### Regional patterns
- Across regions external debt is frequently ranked as the most important factor, except in Latin America where fiscal balance is more prominent.
- Inflation is among the top three factors in most regions.
- REER appears among the top five factors in most regions (data permitting).
- Global factors remain generally less important at regional levels, although world GDP growth and VIX have regional relevance (e.g., VIX is significant for Asia and Pacific).

### Forecasting capability and aggregate forecasts
- Country-specific forecasts:
  - The model produces probabilistic forecasts where higher point estimates align with historical approval periods and lower probabilities align with non-arrangement periods for illustrated country cases.
- Qualitative classification framework:
  - Constructed J = 4 risk regions with calibrated probability thresholds: 0.2252, 0.5383, 0.7710 that minimize RMSE and maximize Pseudo R2 in a beta regression linking the share of countries in each risk class to the annual share of countries in IMF concessional arrangements.
- Beta regression for aggregate demand (Table 4 results):
  - Dependent variable: annual share of countries in IMF concessional financing arrangements.
  - Coefficients (standard errors):
    - Constant, β0: -1.83*** (0.28)
    - Severe risk group (p̂ ≥ 0.7710), β1: 11.39*** (2.18)
    - High risk (0.7710 > p̂ ≥ 0.5383), β2: -5.07 (3.34)
    - Moderate risk (0.5383 > p̂ ≥ 0.2252), β3: 1.72 (1.73)
    - Low risk (0.2252 > p̂), β4: -1.13*** (0.45)
    - Precision parameter, ϕ: 107.06*** (27.22)
  - Pseudo R2: 0.5580
  - RMSE: 0.0297
  - Number of observations (years): 31
  - Significance notation: *** p ≤ 0.01, ** p ≤ 0.05, * p ≤ 0.1
- Model performance:
  - Aggregate forecasts (within-sample up to 2018 and out-of-period for 2019) capture actual annual aggregate demand within 95 percent confidence intervals for each year in the sample used.
  - Note: only PRGT-eligible countries with at least one significant predictor (p-value ≤ 0.1) are included each year.

### Conclusions and policy implications
- Main empirical conclusions:
  - Country-specific domestic variables—especially external debt, inflation, and REER—are the most informative predictors of demand for PRGT concessional financing.
  - Global variables generally do not materially improve country-level predictive power for most countries; where they matter, world GDP growth and VIX are the most relevant.
  - Institutional variables can be significant predictors in some country cases but are constrained by data availability and potential informational redundancy with domestic variables.
  - The two-step factor-augmented probit framework provides significant capability to predict historical initiation of PRGT arrangements at the country level and supports construction of aggregate demand forecasts via a qualitative-to-beta-regression approach.
- Policy recommendations implied by findings:
  - Maintaining sustainable external debt and reducing inflation should help reduce demand for concessional lending.
  - Strengthening fiscal positions, improving GDP growth, and enhancing governance could potentially lower the need for PRGT financing.
  - Country-specific diagnostics and forecasting can inform allocation and management of the IMF’s limited concessional resources.
- Suggested extensions for future research:
  - Apply the approach to IMF non-concessional financing demand.
  - Examine the impact of the COVID-19 pandemic on demand (current study uses pre-pandemic data as of end-2018).
  - Simplify estimation by exploring dynamic factor models with Markov-switching.

*Italic: Source — IMF staff paper (estimation sample and results based on data through end-2018).*

### References

### wpiea2021015-print-pdf - References

### Country sample (Appendix 1)
- Table 5 lists the country sample used in the analysis. 
- The sample includes low-income and developing countries across regions. 
- "* Denotes HIPC countries." (Source: IMF’s Finance Department.)

### Data description and source (Appendix 2)
- Gross domestic product — Percent change of gross domestic product per capita at constant prices in national currency; Source: WEO
- Inflation — Period average of percent change of consumer prices; Source: WEO
- Current account balance — Net total current account as percent of GDP; Source: WEO
- Fiscal balance — Budget balance as percent of GDP; Source: WEO
- External debt — Total external debt (U.S. dollars) as percent of GDP; Source: WEO
- Reserves — Total reserves (excluding gold) in US dollars; Source: WEO
- Terms of trade — Growth of terms of trade; Source: WEO
- Capital inflows — Total net inflows (in U.S. dollars) as percent of GDP; Source: Financial Flow Analysis (Internal database)
- VIX — CBOE new volatility index; Source: Haver
- U.S. federal fund rate — End-of-period U.S. Federal Funds rate; Source: Haver
- Commodity price index — Commodity price index; Source: WEO
- Euro rate — End-of-period 1-month Euro deposit; Source: Haver
- Real effective exchange rate — Real effective exchange rate; Source: IFS
- Risk of external conflict — Assessment of risk to incumbent government from foreign action ranging from non-violent external pressure to violent external pressure; Source: ICRG
- Risk of ethnic tensions — Assessment of degree of tension within a country attributable to racial, nationality, or language divisions; Source: ICRG
- Risk of law and order — Assessment combining strength/impartiality of legal system and popular observance of the law; Source: ICRG
- Risk of annual inflation rate — Estimated annual inflation rate (unweighted average of the consumer price index) with risk points assigned according to a specified scale; Source: ICRG
- Risk of net international liquidity — Comparative liquidity risk ratio (months of imports financeable with reserves) with risk points assigned according to a specified scale; Source: ICRG
- Risk of exchange rate stability — Appreciation/depreciation of a currency against the U.S. dollar over a calendar year with risk points assigned according to the specified scale; Source: ICRG
- Risk of investment profile — Assessment of factors affecting investment risk, sum of three subcomponents each with max 4 points and min 0; Source: ICRG
- Risk of foreign debt service, percent of exports of goods and services — Foreign debt service expressed as percent of exports with risk points assigned according to a specified scale; Source: ICRG
- Risk of democratic accountability — Measure of how responsive government is to its people with points awarded on the basis of the type of governance; Source: ICRG

### Model selection (Appendix 3)
- Objective: Determine country-specific factors for demand of IMF concessional financing by balancing predictive power, model precision, informational quality, and statistical significance.
- Three criteria used for initial variable selection:
  - Predictive power measured by Brier’s quadratic probability score (QPS); desirable for QPS to be close to zero.
  - Model precision measured by a confidence score (CS) derived from the 95 percent confidence bands of within-period probability forecasts; lower (closer to zero) CS is desirable.
  - Informational quality measured by Akaike Information Criterion (AIC), normalized by number of periods T.
- Confidence score definition details:
  - CS_i = δ (1/T) Σ range(CI_it+1(E(y_it+1 | ... ) | α = 0.05)) over t = 0,...,T−1
  - δ is a scaling parameter; the authors set δ = 1/4 to be comparable in magnitude with QPS.
- AIC formula used:
  - AIC_i = 2k − 2 log(L̂), where k is number of estimated parameters and L̂ is maximized likelihood. AIC normalized by T for comparisons.
- Variable selection procedure:
  - The three measures QPS, CS, and AIC were weighted equally in initial selection. 
  - The objective function minimized per country i is ω_QPS QPS_i + ω_CS CS_i + ω_AIC AIC_i, where ω_QPS + ω_CS + ω_AIC = 1.
  - The authors set ω_QPS = ω_CS = ω_AIC = 1/3.
  - Estimates were checked to ensure non-explosive behavior (i.e., Pseudo R^2 < 1).
  - Country-specific factors determined in three stages: domestic factors first, then up to one global factor (inside the dynamic factor model or as a separate explanatory variable), then up to one other/institutional variable outside the dynamic factor model.
- Statistical significance threshold for inclusion in final analyses:
  - Only models with p-value at most 0.1 were included.
- Performance and summary statistics for selected models (across all countries, regardless of significance):
  - Average QPS = 0.064
  - Average CI = 0.099
  - Average AIC = 0.626
  - Average Pseudo R^2 = 0.443
- Additional note:
  - The paper reports that the majority of common factors proved to be significant in forecasting demand for PRGT resources (frequency of significance shown in a figure).

*Source: wpiea2021015-print-pdf - References*

---


_Source: https://www.imf.org/-/media/files/publications/wp/2021/english/wpiea2021015-print-pdf.pdf_