## Annex 1. Credit Inclusion and Indicators from the WBES and Findex

## Source details

**Canonical URL:** [Annex 1. Credit Inclusion and Indicators from the WBES and Findex](https://www.imf.org/-/media/files/publications/wp/2026/english/wpiea2026008-source-pdf.pdf)

## Other formats

- [Markdown version](/-/media/files/publications/wp/2026/english/wpiea2026008-source-pdf.pdf.md)
- [Structured JSON version](/-/media/files/publications/wp/2026/english/wpiea2026008-source-pdf.pdf.json)

---

### Data sources and variables
- WBES variables used:
  - fbankinv1: share of firms using banks to finance investments
  - fbankwk1: share of firms using banks to finance working capitals
  - fbankinv2: share of investments financed by banks
  - fbankwk2: share of working capitals financed by banks
  - fcredobst: share of firms identifying access to finance as a non-major constraint
  - fbankln: share of firms with a bank loan or credit line
  - fsmbankln: share of small firms with a bank loan or credit line
- Findex variables used:
  - bor: share of adults who borrowed from a formal financial institution in the past year
  - borany: share of adults who borrowed money from any source in the past year

### Comparison of supplier-side FI measure with WBES/Findex (Annex Table 1)
- Sample sizes and coverage (number of observations; number of countries):
  - fbankinv1: 78; 53
  - fbankwk1: 65; 53
  - fbankinv2: 78; 53
  - fbankwk2: 65; 53
  - fcredobst: 78; 53
  - fbankln: 77; 52
  - fsmbankln: 77; 52
  - bor: 176; 73
  - borany: 116; 69
- Correlations with supplier-side FI measure:
  - fbankinv1: -0.22
  - fbankwk1: -0.08
  - fbankinv2: -0.15
  - fbankwk2: -0.01
  - fcredobst: 0.06
  - fbankln: -0.13
  - fsmbankln: -0.12
  - bor: 0.28
  - borany: 0.12
- Directional concordance (simultaneous changes) — shares in total common observations (%):
  - Simultaneous increases:
    - fbankinv1 61.54, fbankwk1 55.38, fbankinv2 46.15, fbankwk2 52.31, fcredobst 30.77, fbankln 42.86, fsmbankln 44.16, bor 62.50, borany 46.55
  - Simultaneous decreases:
    - fbankinv1 0.00, fbankwk1 0.00, fbankinv2 0.00, fbankwk2 1.54, fcredobst 1.28, fbankln 0.00, fsmbankln 0.00, bor 0.57, borany 0.86
- Share of total # of increases or decreases in ES financial inclusion (%) that are simultaneous:
  - Simultaneous increases:
    - fbankinv1 97.96, fbankwk1 97.30, fbankinv2 97.30, fbankwk2 100.00, fcredobst 100.00, fbankln 97.06, fsmbankln 97.14, bor 99.10, borany 98.18
  - Simultaneous decreases:
    - fbankinv1 0.00, fbankwk1 0.00, fbankinv2 0.00, fbankwk2 3.23, fcredobst 1.85, fbankln 0.00, fsmbankln 0.00, bor 1.54, borany 1.64
- Key finding:
  - Little correlation between the supplier-side FI measure and the WBES/Findex user-side measures, but conditional shares of increases in FI given increases in the other measures are close to 100 percent, indicating directional consistency when tracking increases in credit financial inclusion.

*Source: Authors calculations using IMF FAS, WB ES, and WB Findex*

### Composite FI measure construction (Annex 2)
- Definition:
  - FI_{i,t} = w_{b,i} borrower_{i,t} + w_{l,i} loan_{i,t}
- Weight construction:
  - Weights (w_{b,i}, w_{l,i}) are the eigenvector corresponding to the largest eigenvalue of the correlation matrix computed from borrower_{i,t} and loan_{i,t} for country i’s time horizon.
  - Empirical result: in virtually all countries the two weights are similar, w_{b,i} ∼ w_{l,i} ∼ 0.71 for each country i.
- Data coverage note:
  - borrower_{i,t} and loan_{i,t} have 1,650 and 1,786 observations, respectively, but only 1,222 observations in common; observations where either variable is missing are dropped.
- Cross-check with user-side surveys:
  - Despite low correlations, FI and WBES/Findex measures tend to be directionally consistent: when FI signals an increase in credit inclusion, other indicators signal increases as well.

### Robustness checks (Annex 3)

3.1 Credit inclusion booms with different thresholds (top 15th percentile)
- Definition variant:
  - FIBoom_{i,t} = 1 if Δ^3 FI_{i,t} exceeds the top 15th percentile of its long-run distribution (alternative to top 25th percentile).
- General observation:
  - Estimates become noisier relative to the top 25th percentile due to reduced variation in FIBoom_{i,t}, but results are qualitatively similar.
- Selected predictive regression results (Panel A: FIBoom predictor for Credit Booms within h years):
  - FIBoom coefficients:
    - 1 year: 2.68 (1.78)
    - 2 years: 5.46 (3.35)
    - 3 years: 2.85 (3.63)
    - 4 years: 0.75 (3.79)
  - Observations (1 to 4 years): 699; 633; 503; 322
  - Countries (1 to 4 years): 70; 68; 65; 52
  - R2 (1 to 4 years): 0.037; 0.074; 0.113; 0.293
  - AUC (1 to 4 years): 0.876; 0.878; 0.854; 0.846
- Panel B (FI continuous standardized:
  - FI coefficients (multiply outcomes by 100):
    - 1 year: 1.25* (0.59)
    - 2 years: 2.52* (1.18)
    - 3 years: 4.07* (1.89)
    - 4 years: 3.95* (2.00)
  - R2 (1 to 4 years): 0.038; 0.077; 0.131; 0.309
  - AUC (1 to 4 years): 0.896; 0.899; 0.846; 0.846

3.2 Local projections with continuous regressor (standardized Δ^3 FI)
- Rationale:
  - Replace FIBoom_{i,t} with standardized Δ^3 FI_{i,t} to increase variation and improve precision.
- Summary selected local projection estimates (Panel A: Model with Interaction; dependent = Non-Performing Loan Rates and Bank Z-Score; t+1 to t+4):
  - Boom (NPL rates):
    - t+1: 0.81* (0.37)
    - t+2: 1.15+ (0.66)
    - t+3: 1.86+ (1.09)
    - t+4: 2.81 (2.08)
  - FIBoom (NPL rates):
    - t+1: -0.13 (0.25)
    - t+2: -0.78+ (0.46)
    - t+3: -1.03 (0.79)
    - t+4: -1.11 (0.85)
  - Boom*FIBoom (NPL rates):
    - t+1: 0.31 (0.43)
    - t+2: 1.94+ (1.04)
    - t+3: 2.05 (1.28)
    - t+4: 0.99 (0.96)
  - Boom*FIBoom (Bank Z-Score):
    - t+1: -1.26** (0.52)
    - t+2: -0.47 (0.53)
    - t+3: -0.56 (0.72)
    - t+4: 0.35 (0.54)
  - Observations (NPL t+1 to t+4): 349; 299; 250; 203
  - Countries (NPL t+1 to t+4): 52; 50; 47; 42
  - R2 (NPL t+1 to t+4): 0.025; 0.06; 0.08; 0.1
  - Observations (Z-Score t+1 to t+4): 445; 390; 338; 292
  - Countries (Z-Score t+1 to t+4): 59; 55; 50; 49
  - R2 (Z-Score t+1 to t+4): 0.153; 0.23; 0.197; 0.341
- Panel B (without interaction) yields broadly similar patterns with some coefficients larger and precision varying.

3.3 Alternative variables for FI measure (borrower or loan separately)
- Approach:
  - Replace composite FI_{i,t} with borrower_{i,t} or loan_{i,t}, restricting to subsamples where both are not missing.
- Main findings:
  - Both borrower and loan yield similar patterns in estimated effects compared to composite FI.
  - Estimates using loan are more precise than borrower, likely due to larger variance of loan variable.
  - No significant information loss observed using composite FI.
- Selected results (borrower-based, Panel A Model with Interaction; NPL):
  - Boom (NPL):
    - t+1: 0.83* (0.41)
    - t+2: 1.16 (0.74)
    - t+3: 1.92 (1.26)
    - t+4: 3.09 (2.43)
  - Boom*Borrower (NPL):
    - t+1: 0.11 (0.42)
    - t+2: 1.30+ (0.74)
    - t+3: 1.22 (0.92)
    - t+4: -0.22 (1.62)
  - Observations (NPL t+1 to t+4): 349; 299; 250; 203
  - Countries: 52; 50; 47; 42
  - R2 (NPL): 0.024; 0.046; 0.067; 0.086

3.4 Credit booms identified with different threshold (1.5× SD)
- Robustness check:
  - Credit boom threshold relaxed from 1.75× SD to 1.5× SD of cyclical component.
- Result:
  - Local projections with relaxed credit boom threshold produce results consistent with main text.
- Selected results (Panel A Model with Interaction; loan variable; NPL t+1 to t+4):
  - Boom (NPL):
    - t+1: 0.43 (0.35)
    - t+2: 0.67 (0.66)
    - t+3: 1.52 (1.16)
    - t+4: 2.62 (2.25)
  - Boom*Loan (NPL):
    - t+1: 1.03* (0.53)
    - t+2: 2.31** (0.73)
    - t+3: 2.07** (0.85)
    - t+4: 1.28 (0.94)
  - Observations (NPL t+1 to t+4): 349; 299; 250; 203
  - Countries (NPL t+1 to t+4): 52; 50; 47; 42
  - R2 (NPL t+1 to t+4): 0.031; 0.059; 0.069; 0.086

### Riskiness of Credit Allocation (RCA) — construction and data scope (Annex 4)
- RCA constructed for four firm-level vulnerability indicators:
  - rca_lev: leverage (total debt to total assets)
  - rca_debt: debt overhang (total debt to EBITDA)
  - rca_icr: interest coverage ratio (EBITDA to interest expenses)
  - rca_edf: expected default frequency (EDF)
- Construction steps:
  - Each year, each firm assigned a decile (1 to 10) of the indicator distribution in its country; higher decile = larger vulnerability.
  - Firms sorted by changes in net debt to lagged total assets into five equal-sized bins; top issuers = largest increases in debt; bottom issuers = largest decreases.
  - RCA measure = average vulnerability decile of top issuers minus average vulnerability decile of bottom issuers.
- Sample considerations:
  - RCA series use listed firms mainly from middle-to-high income countries; fewer observations and countries than npl and zscore when matched to credit inclusion measure.

### RCA summary statistics (when matched to main data)
- rca_lev: N 252; Countries 21; Mean -0.11; Std. Dev. 0.96; Min -2.7; Pctl. 25 -0.75; Pctl. 75 0.43; Max 3.3
- rca_debt: N 250; Countries 21; Mean -0.18; Std. Dev. 0.87; Min -3.1; Pctl. 25 -0.68; Pctl. 75 0.33; Max 2.3
- rca_icr: N 251; Countries 21; Mean -0.06; Std. Dev. 0.81; Min -2.2; Pctl. 25 -0.57; Pctl. 75 0.47; Max 3.4
- rca_edf: N 244; Countries 21; Mean 0.0021; Std. Dev. 0.96; Min -3.2; Pctl. 25 -0.62; Pctl. 75 0.63; Max 3.2

### Findings from RCA impulse response analysis (Annex Figures 11–12 and Annex Table 9)
- General:
  - Impulse responses from credit inclusion boom shocks for the four RCA series are generally noisy; possibly due to weak association between RCA measures and credit inclusion measure.
- Leverage-based RCA (strongest signal):
  - Evidence that credit inclusion booms during credit booms increase RCA (i.e., leverage of top issuers expands relative to bottom issuers), interpreted as mounting financial strains.
- EDF-based RCA:
  - Observed weak and opposite effects: credit inclusion booms during credit booms lead to decreases in the RCA measure; rapid decreases in RCA measures are often accompanied by sharp economic downturns, which can be interpreted as evidence of negative effects on financial stability.
- Effects of credit booms:
  - Credit booms during credit inclusion booms lead to significant reductions across all RCA measures, possibly indicating immediate financial instability and economic downturns.

Selected coefficients from Annex Table 9 (Panel A and Panel B; dependent = RCA measures t+1 to t+4)
- Panel A: Leverage-Based RCA (t+1 to t+4)
  - Boom: -0.80** (0.20); -1.14* (0.50); -1.57** (0.46); -1.44** (0.42)
  - FIBoom: -0.08 (0.21); -0.04 (0.21); 0.32 (0.22); -0.08 (0.28)
  - Boom*FIBoom: 0.66* (0.29); 0.08 (0.64); 0.85 (0.65); 0.45 (0.36)
  - Observations: 152; 135; 118; 101
  - Countries: 17; 17; 17; 16
  - R2: 0.432; 0.5; 0.455; 0.496
- Panel A: Debt-Overhang-Based RCA (t+1 to t+4)
  - Boom: 0.12 (0.39); 0.11 (0.47); 0.13 (0.76); 0.08 (0.98)
  - FIBoom: -0.04 (0.18); 0.23 (0.21); 0.36 (0.30); 0.06 (0.31)
  - Boom*FIBoom: -0.29 (0.42); -0.95 (0.63); -0.57 (0.64); -0.15 (0.55)
  - Observations: 150; 133; 116; 99
  - Countries: 17; 17; 17; 16
  - R2: 0.32; 0.391; 0.408; 0.253
- Panel B: ICR-Based RCA (t+1 to t+4)
  - Boom: -0.12 (0.11); 0.17 (0.23); -0.01 (0.46); 0.09 (0.64)
  - FIBoom: -0.04 (0.16); 0.34 (0.25); 0.36+ (0.21); 0.03 (0.25)
  - Boom*FIBoom: 0.05 (0.25); -0.85+ (0.48); -0.43 (0.52); -0.10 (0.59)
  - Observations: 152; 135; 118; 101
  - R2: 0.373; 0.444; 0.456; 0.409
- Panel B: EDF-Based RCA (t+1 to t+4)
  - Boom: -0.43 (0.27); 0.17 (0.44); 0.62 (0.80); -0.28 (0.65)
  - FIBoom: -0.22 (0.19); 0.06 (0.26); 0.18 (0.33); -0.50 (0.30)
  - Boom*FIBoom: -0.72 (0.82); -1.12** (0.37); -0.74 (0.63); -0.29 (1.01)
  - Observations: 142; 127; 112; 97
  - R2: 0.488; 0.475; 0.473; 0.643

_Italic: Source — Authors’ calculations using IMF FAS, WB ES, WB Findex, and firm-level data as presented in wpiea2026008-source-pdf._

### Annex 1. Credit Inclusion and Indicators from the WBES and Findex ......................................................

### Annex 1. Credit Inclusion and Indicators from the WBES and Findex ......................................................

### Contained annexes and subsections
- Annex 1. Credit Inclusion and Indicators from the WBES and Findex ......................................................... 26
- Annex 2. Construction of the Composite FI Measure ................................................................................... 27
- Annex 3. Robustness Checks ...................................................................................................................... 27
  - 3.1 Credit inclusion booms with different thresholds ............................................................................. 27
  - 3.2 Local Projections regression with continuous Regressor ................................................................ 30
  - 3.3 Alternative variables for FI measure ................................................................................................ 32
  - 3.4 Credit with a different threshold ....................................................................................................... 36
- Annex 4. Riskiness of Credit Allocation ........................................................................................................ 38

*Source: wpiea2026008-source-pdf - Annex 1. Credit Inclusion and Indicators from the WBES and Findex*

### References .............................................................................................................

### wpiea2026008-source-pdf - References .............................................................................................................

### I. Introduction — research questions and overview
- Research focus: interaction between credit inclusion (financial inclusion on the credit side) and credit booms, and implications for subsequent financial instability.
- Two main questions:
  - To what extent does rapid growth in financial inclusion coincide with credit booms?
  - Is a rapid increase in financial inclusion more likely to precede subsequent financial instability when accompanied by a credit boom?
- Dataset: novel country-panel covering 96 countries from 2004 to 2021, matching supply-side credit inclusion data (IMF Financial Access Survey) with identified credit boom periods (IMF GDD) and measures of financial instability.
- Main conceptual distinction: intensive margin (more funds for same borrowers) vs extensive margin (expanding pool of borrowers = credit inclusion).

### Key high-level findings (from the Introduction)
- Rapid growth in credit inclusion is more common during credit booms than outside them.
- Rapid growth in credit inclusion is more frequent during “bad” credit booms (those followed by a crisis within three years) than during “good” booms.
- Within credit booms, rapid growth in credit inclusion signals a subsequent crisis that is more than twice as severe as one associated with slower growth.
- Credit inclusion booms during credit booms predict increases in non-performing loan rates (NPL) by about 40 percent at their peak (statistically significant).
- Credit inclusion booms not accompanied by credit booms tend to improve financial stability.
- Financial stability effects of credit booms alone are smaller than those of credit inclusion booms during credit booms.
- Significant negative effects of credit inclusion booms on median GDP growth rates reported.

### Policy implications (from the Introduction)
- Monitoring increases in credit inclusion alongside credit growth can complement early warning toolkits.
- High-quality, real-time credit inclusion data are essential.
- When promoting financial inclusion, monitor current levels of private credit; promoting inclusion outside of credit booms appears to improve financial stability.

---

### II. Data and methods — variables, identification, and sample
- Financial inclusion data:
  - Source: IMF Financial Access Survey (FAS), years 2004 to 2023.
  - Two FAS variables used: number of borrowers from commercial banks per 1,000 adults (borrower) and number of loan accounts with commercial banks per 1,000 adults (loan).
  - Correlation between the two variables: about 0.84.
  - Composite indicator constructed via principal component analysis (PCA) — denoted FI.
- Credit boom identification:
  - Private credit from IMF Global Debt Database (GDD): private credit to nonfinancial sectors (bank loans + domestic and external securities).
  - Decompose private credit-to-GDP ratio using the Hamilton filter (Hamilton, 2018).
  - Boom defined when cyclical component exceeds boom threshold multiple of long-run standard deviation; start/end determined by lower threshold multiple.
  - Thresholds: boom threshold multiple set at 1.75; start/end-year multiples at 1.0 (following Mendoza and Terrones (2008)).
- Financial stability measures:
  - Primary distress measure: banking crisis data from Laeven and Valencia (2020); crisis dummy (crises) = 1 if banking crisis in a given year.
  - Other measures: bank Z-score (zscore) and nonperforming loans to gross loans (npl) from World Bank Global Financial Development Database.
  - Bank Z-score formula referenced as (ROA + equity/assets) / sd(ROA) aggregated across commercial banks.
- Sample and matching:
  - Dataset covers 2004 to 2021 and 96 countries after matching and dropping observations with missing borrower or loan_accounts in the same year.

### Descriptive statistics (Table 1 highlights)
- FI: N = 1222, Countries = 96, Mean = 348, Std. Dev. = 396, Min = -272, 25th Pctl. = 46, 75th Pctl. = 546, Max = 3835.
- FI: 3-year changes: N = 936, Countries = 93, Mean = 31, Std. Dev. = 120, Min = -415, 25th Pctl. = -5.7, 75th Pctl. = 51, Max = 1291.
- boom: N = 942, Countries = 76, Mean = 0.1, Std. Dev. = 0.31, Min = 0, Pctl. 25 = 0, Pctl. 75 = 0, Max = 1.
- bad: N = 466, Countries = 50, Mean = 0.026, Std. Dev. = 0.16, Min = 0, Pctl. 25 = 0, Pctl. 75 = 0, Max = 1.
- crises: N = 521, Countries = 54, Mean = 0.052, Std. Dev. = 0.22, Min = 0, Pctl. 25 = 0, Pctl. 75 = 0, Max = 1.
- npl: N = 773, Countries = 74, Mean = 6.9, Std. Dev. = 6.1, Min = 0.35, 25th Pctl. = 2.9, 75th Pctl. = 9.5, Max = 49.
- zscore: N = 939, Countries = 82, Mean = 14, Std. Dev. = 7.5, Min = 0.22, 25th Pctl. = 8.8, 75th Pctl. = 19, Max = 67.
- Frequencies: credit booms ≈ 10 percent; crises ≈ 5 percent.
- Crisis data available only up to 2017; booms identified up to 2022.

### Banking crises by income class (Table 2 highlights)
- Observations: Low = 141, Middle = 624, High = 231.
- # of countries: Low = 12, Middle = 52, High = 21.
- Years of crises: Low = 0, Middle = 10, High = 41.
- Frequency of crises (%): Low = 0, Middle = 1.6, High = 17.75.
- Fraction of crises from 2007 to 2011 relative to all crises (%): Middle = 50.0, High = 80.49.

---

### III. Stylized facts — co-movement and classification of booms
- Distributional patterns:
  - Histograms of 3-year changes in FI show thicker right tails during booms vs outside booms (Figure 2): increases in credit inclusion are more intensive during booms.
- Classification of booms:
  - “Bad booms” = booms followed by a crisis within a three-year window; “good booms” = not followed by a crisis within three years.
- Conditional patterns (Figure 3 / Figure 4 summary):
  - More than 80 percent of bad booms are characterized by rapid credit inclusion; less than 20 percent exhibit slow credit inclusion.
  - Experiencing high growth in credit inclusion and a credit boom simultaneously makes a subsequent crisis more than twice as likely compared to slow growth in credit inclusion during a boom.
- Signal vs state matrices (Table 4 summary):
  - Evaluated true/false positive/negative rates for predicting crises using credit booms alone vs concurrent booms (credit boom + high FI).
  - Both indicators achieve true positive rate 0.667 (> 0.5).
  - Concurrent booms reduce false positives (0.075) compared to credit booms alone (0.112).
  - AUC: concurrent booms = 0.80; credit booms = 0.78 (benchmark AUC = 0.5 for coin toss).

---

### IV. Empirical strategy — local projections and predictive regressions
- Local projections specification (Jorda, 2005) used to estimate cumulative changes in outcomes y over horizons h = 1,...,4:
  - Δh yi,t+h = αi,h + βh Boomi,t + γh FIBoomi,t + δh Boomi,t × FIBoomi,t + Σj=0^3 θj,h Δy i,t−j + εi,t+h
  - FIBoomi,t = 1 if three-year change in financial inclusion, Δ3FIi,t, exceeds top 25th percentile of country-specific long-run distribution (alternative percentiles tested in Annex).
  - Parameter of interest for inclusion booms during credit booms: γh + δh (expected difference in future distress between boom with FIBoom =1 vs boom with FIBoom =0).
  - Standard errors: Driscoll and Kraay (1998).

- Predictive regression for credit booms:
  - Boomi,t→t+h = αi,h + βh FIBoomi,t + Σj=1^3 θj,h Boomi,t−j + εi,t+h
  - Outcome Boomi,t→t+h = max{Boomi,t, ..., Boomi,t+h} equals 1 if a credit boom begins between t and t+h.
  - Primary interest: βh (added probability of credit boom following a FIBoom).

- Reverse predictive regression (whether credit booms predict FIBoom):
  - FIBoomi,t→t+h = αi,h + βh Boomi,t + Σj=1^3 θj,h FIBoomi,t−j + εi,t+h

---

### V. Financial inclusion, credit booms, and financial stability — empirical results
- Impulse response findings (Figures 5 and 6 summary):
  - Credit inclusion booms during credit booms predict higher financial distress:
    - Simultaneous credit inclusion and credit boom predicts nonperforming loan rates to increase by as much as 2 percent compared to baseline (Figure 5 narrative).
  - γh estimates (inclusion booms outside of credit booms) are negative, suggesting credit inclusion booms outside booms can be beneficial for financial stability.
  - Credit booms during inclusion booms also predict financial instability; baseline credit boom effects (βh) are smaller than simultaneous boom effects (βh + δh).
  - Interaction term Boom × FIBoom dominates individual effects; excluding interaction can mask predictive power of inclusion booms.

- Table 5 (Panel A: model with interaction; Panel B: model without interaction) — selected coefficient patterns:
  - Panel A (with interaction), dependent variable: Non-Performing Loan Rates:
    - Boom coefficients: t+1 = 0.43 (0.35), t+2 = 0.69 (0.66), t+3 = 1.48 (1.18), t+4 = 2.58 (2.26).
    - FIBoom coefficients: t+1 = 0.06 (0.25), t+2 = -0.46 (0.37), t+3 = -0.76 (0.67), t+4 = -0.83 (0.73).
    - Boom*FIBoom interaction: t+1 = 1.01* (0.51), t+2 = 2.29** (0.71), t+3 = 2.28** (0.80), t+4 = 1.46 (0.91).
    - Observations t+1 = 349, Countries = 52; R2 t+1 = 0.032 (varies by horizon).
  - Panel A, dependent variable: Bank Z-Score:
    - Boom*FIBoom interaction: t+1 = -1.32** (0.50), t+2 = -1.37** (0.49), t+3 = -0.96 (0.66), t+4 = 0.11 (0.37).
    - Observations t+1 = 445, Countries = 59; R2 t+1 = 0.156.
  - Panel B (without interaction) shows larger coefficients on Boom alone for NPLs but interaction omission reduces interpretability; R2 values are lower compared to models with interaction in some horizons.

- Real GDP growth effects (Figures 7 and 8; Table 6 highlights):
  - Concurrent inclusion and credit booms have short-term negative effects on economic growth, reaching as much as -2% in impulse responses.
  - Table 6 Panel A (Model with FIBoom), dependent variable: Real GDP Growth Rates:
    - Boom coefficients: t+1 = 0.42 (0.82), t+2 = 0.74 (1.42), t+3 = -1.56 (1.94), t+4 = -3.52 (2.47).
    - FIBoom coefficients: t+1 = -0.31 (0.42), t+2 = -0.19 (0.74), t+3 = 0.02 (0.96), t+4 = 0.98 (1.31).
    - Boom*FIBoom interaction: t+1 = -0.57 (1.24), t+2 = -1.57 (1.62), t+3 = -0.61 (1.72), t+4 = 1.28 (2.12).
    - Observations t+1 = 592, Countries = 69; R2 t+1 = 0.004.
  - Table 6 Panel B (Model with FI continuous), Boom*FI significant at t+2 = -1.78** (0.64).

---

### VI. Predictive relations: inclusion booms preceding credit booms
- Predictive regression results (Table 7):
  - Panel A (FIBoom as predictor; dependent = Credit Booms within...):
    - FIBoom coefficients: 1 year = 2.93* (1.48); 2 years = 5.47+ (2.90); 3 years = 2.37 (3.45); 4 years = 0.22 (3.48).
    - Observations for 1 year = 699, Countries = 70, R2 = 0.039.
    - AUCs: 1 year = 0.873, 2 years = 0.884, 3 years = 0.854, 4 years = 0.84.
    - Interpretation: credit inclusion booms can increase probability of entering a credit boom by up to 5 percent (peak).
  - Panel B (FI continuous standardized as predictor):
    - FI coefficients: 1 year = 1.25* (0.59); 2 years = 2.52* (1.18); 3 years = 4.07* (1.89); 4 years = 3.95* (2.00).
    - Observations and countries same as Panel A.
    - AUCs: 1 year = 0.896, 2 years = 0.899, 3 years = 0.846, 4 years = 0.846.
    - Interpretation: a one standard deviation increase in Δ3FI raises probability of a credit boom by the indicated percentage points.
- Reverse predictive regression (Table 8):
  - Dependent = Credit Inclusion Booms within...; predictor = Boom:
    - Boom coefficients: 1 year = 6.41 (5.04); 2 years = 8.19 (8.52); 3 years = 1.58 (9.73); 4 years = 8.35 (12.21).
    - Observations 1 year = 490, Countries = 64.
    - AUCs: 1 year = 0.746, 2 years = 0.769, 3 years = 0.726, 4 years = 0.751.
    - Interpretation: larger but noisier estimates; credit booms weakly signal onset of inclusion booms compared to the stronger signaling of inclusion booms for credit booms.

---

### VII. Synthesis of empirical conclusions
- Complementarity: credit inclusion booms and credit booms together predict higher subsequent financial distress (NPL increases and Z-score declines) and short-term negative GDP growth effects.
- Directionality: rapid increases in financial inclusion tend to precede credit booms; credit booms do not strongly predict subsequent rapid increases in inclusion.
- Heterogeneous effects:
  - Inclusion booms during non-boom periods can be associated with improved financial stability (negative γh estimates).
  - Inclusion booms during credit booms predict materially worse financial outcomes (positive γh + δh).
- Importance of interaction: models that omit Boom × FIBoom interaction may understate the predictive power of inclusion booms for instability.

### VIII. Targeted policy recommendations (from Conclusion and policy discussion)
- Monitor credit inclusion alongside aggregate private credit:
  - If a credit boom is ongoing, promoting financial inclusion may heighten financial stability risks.
  - Promoting financial inclusion outside of credit booms tends to improve financial stability.
- Data priorities:
  - Timely, high-frequency, comprehensive data on financial inclusion (credit inclusion in particular) are valuable for policymaking and early warning.
  - FAS annual frequency is useful but limited; higher-frequency monitoring is desirable.
- Macroprudential application:
  - Use credit inclusion indicators to help distinguish “bad” from “good” credit booms and to enhance early warning toolkits.

### IX. Open questions and avenues for future research
- Mechanisms underlying concurrent booms:
  - Is the concurrency driven by lax lending standards or weak regulatory oversight?
  - Do concurrent booms generate widespread asset price inflation that precipitates crises?
  - Are there particular sectors where concurrent booms are more dangerous?
- These channel-identification questions are flagged for future work.

_Italic: Source — Authors’ calculations using IMF FAS, IMF GDD, Laeven and Valencia (2020), and WB GFD as presented in wpiea2026008-source-pdf._

### Annex 1. Credit Inclusion and Indicators from the WBES and Findex

### Annex 1. Credit Inclusion and Indicators from the WBES and Findex

### Data sources and variables
- WBES variables used:
  - fbankinv1: share of firms using banks to finance investments
  - fbankwk1: share of firms using banks to finance working capitals
  - fbankinv2: share of investments financed by banks
  - fbankwk2: share of working capitals financed by banks
  - fcredobst: share of firms identifying access to finance as a non-major constraint
  - fbankln: share of firms with a bank loan or credit line
  - fsmbankln: share of small firms with a bank loan or credit line
- Findex variables used:
  - bor: share of adults who borrowed from a formal financial institution in the past year
  - borany: share of adults who borrowed money from any source in the past year

### Comparison of supplier-side FI measure with WBES/Findex (Annex Table 1)
- Sample sizes and coverage:
  - # of observations: fbankinv1 78, fbankwk1 65, fbankinv2 78, fbankwk2 65, fcredobst 78, fbankln 77, fsmbankln 77, bor 176, borany 116
  - # of countries: fbankinv1 53, fbankwk1 53, fbankinv2 53, fbankwk2 53, fcredobst 53, fbankln 52, fsmbankln 52, bor 73, borany 69
  - # of common observations (with FI measure): identical to # of observations above for each variable
- Correlations with supplier-side FI measure:
  - fbankinv1: -0.22
  - fbankwk1: -0.08
  - fbankinv2: -0.15
  - fbankwk2: -0.01
  - fcredobst: 0.06
  - fbankln: -0.13
  - fsmbankln: -0.12
  - bor: 0.28
  - borany: 0.12
- Directional concordance (simultaneous changes) — shares in total common observations (%):
  - Simultaneous increases:
    - fbankinv1 61.54, fbankwk1 55.38, fbankinv2 46.15, fbankwk2 52.31, fcredobst 30.77, fbankln 42.86, fsmbankln 44.16, bor 62.50, borany 46.55
  - Simultaneous decreases:
    - fbankinv1 0.00, fbankwk1 0.00, fbankinv2 0.00, fbankwk2 1.54, fcredobst 1.28, fbankln 0.00, fsmbankln 0.00, bor 0.57, borany 0.86
- Share of total # of increases or decreases in ES financial inclusion (%) that are simultaneous:
  - Simultaneous increases:
    - fbankinv1 97.96, fbankwk1 97.30, fbankinv2 97.30, fbankwk2 100.00, fcredobst 100.00, fbankln 97.06, fsmbankln 97.14, bor 99.10, borany 98.18
  - Simultaneous decreases:
    - fbankinv1 0.00, fbankwk1 0.00, fbankinv2 0.00, fbankwk2 3.23, fcredobst 1.85, fbankln 0.00, fsmbankln 0.00, bor 1.54, borany 1.64
- Key finding from the comparison:
  - Little correlation between the supplier-side FI measure and the WBES/Findex user-side measures, but conditional shares of increases in FI given increases in the other measures are close to 100 percent, indicating directional consistency when tracking increases in credit financial inclusion.

*Source: Authors calculations using IMF FAS, WB ES, and WB Findex*

### Annex 2. Construction of the Composite FI Measure
- Composite FI definition:
  - FI_{i,t} = w_{b,i} borrower_{i,t} + w_{l,i} loan_{i,t}
- Weight construction:
  - Weights (w_{b,i}, w_{l,i}) are the eigenvector corresponding to the largest eigenvalue of the correlation matrix computed from borrower_{i,t} and loan_{i,t} for country i’s time horizon.
  - Empirical result: in virtually all countries the two weights are similar, w_{b,i} ∼ w_{l,i} ∼ 0.71 for each country i, reflecting strong correlations and similar importance of borrower and loan variables.
- Data coverage note:
  - Borrower and loan variables have 1,650 and 1,786 observations, respectively, but only 1,222 observations in common; observations where either variable is missing are dropped in the composite construction.
- Cross-check with user-side surveys:
  - Although low correlation exists between FI and WBES/Findex measures, they tend to be directionally consistent: when FI signals an increase in credit inclusion, other indicators signal increases as well.

### Annex 3. Robustness Checks

### 3.1 Credit inclusion booms with different thresholds (top 15th percentile)
- Definition variant:
  - FIBoom_{i,t} redefined as dummy = 1 if Δ^3 FI_{i,t} exceeds the top 15th percentile of its long-run distribution (alternative to top 25th percentile used in main analysis).
- General observation:
  - Estimates become noisier relative to the top 25th percentile due to reduced variation in FIBoom_{i,t}, but results are qualitatively similar.
- Predictive regression results (Annex Table 2) — Panel A: Model with FIBoom (dependent variable: Credit Booms within...)
  - FIBoom coefficients (standard errors in parentheses):
    - 1 year: 2.68 (1.78)
    - 2 years: 5.46 (3.35)
    - 3 years: 2.85 (3.63)
    - 4 years: 0.75 (3.79)
  - Observations: 1 year 699, 2 years 633, 3 years 503, 4 years 322
  - Countries: 1 year 70, 2 years 68, 3 years 65, 4 years 52
  - R2: 0.037, 0.074, 0.113, 0.293 (1 to 4 years)
  - AUC: 0.876, 0.878, 0.854, 0.846 (1 to 4 years)
- Predictive regression results — Panel B: Model with FI (coefficients multiply outcomes by 100; standard errors in parentheses)
  - FI coefficients:
    - 1 year: 1.25* (0.59)
    - 2 years: 2.52* (1.18)
    - 3 years: 4.07* (1.89)
    - 4 years: 3.95* (2.00)
  - Observations, Countries, R2, AUC: same counts as Panel A; R2: 0.038, 0.077, 0.131, 0.309; AUC: 0.896, 0.899, 0.846, 0.846

### 3.2 Local projections with continuous regressor (standardized Δ^3 FI)
- Rationale:
  - Replace FIBoom_{i,t} with standardized Δ^3 FI_{i,t} to increase variation and potentially improve precision.
- Summary result:
  - Panel B of Annex Table 4 and associated figures show similar financial inclusion boom effects and credit boom effects as main text, confirming robustness to choice of regressor.
- Selected local projection estimates (Annex Table 4 summarized):
  - Panel A: Model with Interaction (dependent variables: Non-Performing Loan Rates and Bank Z-Score at t+1 to t+4)
    - Boom coefficients (NPL rates):
      - t+1: 0.81* (0.37)
      - t+2: 1.15+ (0.66)
      - t+3: 1.86+ (1.09)
      - t+4: 2.81 (2.08)
    - FIBoom coefficients (NPL rates):
      - t+1: -0.13 (0.25)
      - t+2: -0.78+ (0.46)
      - t+3: -1.03 (0.79)
      - t+4: -1.11 (0.85)
    - Boom*FIBoom interaction (NPL rates):
      - t+1: 0.31 (0.43)
      - t+2: 1.94+ (1.04)
      - t+3: 2.05 (1.28)
      - t+4: 0.99 (0.96)
    - Observations (NPL): t+1 349, t+2 299, t+3 250, t+4 203
    - Countries (NPL): 52, 50, 47, 42
    - R2 (NPL): 0.025, 0.06, 0.08, 0.1
    - Boom coefficients (Bank Z-Score):
      - t+1: 0.35 (0.31)
      - t+2: 0.49 (0.39)
      - t+3: 0.22 (0.52)
      - t+4: 0.25 (0.71)
    - Boom*FIBoom interaction (Bank Z-Score):
      - t+1: -1.26** (0.52)
      - t+2: -0.47 (0.53)
      - t+3: -0.56 (0.72)
      - t+4: 0.35 (0.54)
    - Observations (Z-Score): t+1 445, t+2 390, t+3 338, t+4 292
    - Countries (Z-Score): 59, 55, 50, 49
    - R2 (Z-Score): 0.153, 0.23, 0.197, 0.341
  - Panel B: Model without Interaction (selected coefficients)
    - Boom (NPL rates):
      - t+1: 0.86** (0.33)
      - t+2: 1.48* (0.65)
      - t+3: 2.12+ (1.09)
      - t+4: 2.96 (1.99)
    - FIBoom (Bank Z-Score) at t+4: 0.48* (0.24)
    - Observations and countries identical to Panel A; R2 similar.

### 3.3 Alternative variables for FI measure (borrower or loan separately)
- Approach:
  - Replace composite FI_{i,t} with borrower_{i,t} or loan_{i,t}, restricting to subsamples where both are not missing.
- Main findings:
  - Both borrower and loan yield similar patterns in estimated effects of financial inclusion and credit booms compared to the composite FI measure.
  - Estimates using loan are more precise than borrower, likely due to larger variance of loan variable.
  - No significant information loss observed using the composite FI measure.
- Selected results (Annex Table 5: number of borrowers)
  - Panel A: Model with Interaction (NPL rates)
    - Boom coefficients (NPL):
      - t+1: 0.83* (0.41)
      - t+2: 1.16 (0.74)
      - t+3: 1.92 (1.26)
      - t+4: 3.09 (2.43)
    - Borrower coefficients (NPL):
      - t+1: -0.00 (0.22)
      - t+2: -0.35 (0.33)
      - t+3: -0.69 (0.54)
      - t+4: -0.61 (0.53)
    - Boom*Borrower interaction (NPL):
      - t+1: 0.11 (0.42)
      - t+2: 1.30+ (0.74)
      - t+3: 1.22 (0.92)
      - t+4: -0.22 (1.62)
    - Observations (NPL): 349, 299, 250, 203; Countries: 52, 50, 47, 42; R2: 0.024, 0.046, 0.067, 0.086
  - Panel B: Model without Interaction (NPL)
    - Boom coefficients (NPL):
      - t+1: 0.86** (0.33)
      - t+2: 1.50** (0.64)
      - t+3: 2.20+ (1.13)
      - t+4: 3.02 (2.02)
    - Borrower coefficients are small and imprecise; borrower at Bank Z-Score t+4: 0.49+ (0.28)
- Annex Tables 6 and Figures 5–8 report analogous results for loan_{i,t} and show higher precision for loan-based estimates.

### 3.4 Credit booms identified with different threshold (1.5× SD)
- Main text threshold: 1.75 times the standard deviation of the cyclical component of credit growth.
- Robustness check: threshold set to 1.5 times the standard deviation.
- Result:
  - Local projections with relaxed credit boom threshold produce results consistent with the main text, indicating robustness to the choice of credit boom threshold.
- Selected results (Annex Table 7 summary with relaxed threshold) — Panel A: Model with Interaction (Loan variable presented)
  - Boom coefficients (NPL rates):
    - t+1: 0.43 (0.35)
    - t+2: 0.67 (0.66)
    - t+3: 1.52 (1.16)
    - t+4: 2.62 (2.25)
  - Loan coefficients (NPL rates): small and imprecise across horizons
  - Boom*Loan interaction (NPL rates):
    - t+1: 1.03* (0.53)
    - t+2: 2.31** (0.73)
    - t+3: 2.07** (0.85)
    - t+4: 1.28 (0.94)
  - Observations and countries: same counts as earlier local projections (NPL observations 349, 299, 250, 203; countries 52, 50, 47, 42)
  - R2 (NPL): 0.031, 0.059, 0.069, 0.086
  - Panel B: Model without Interaction yields Boom (NPL) t+1 0.84** (0.33), t+2 1.53* (0.67), t+3 2.25+ (1.21), t+4 3.11 (2.18)

*IMF WORKING PAPERS Financial Inclusion, Credit Booms, and Financial Stability Risk — INTERNATIONAL MONETARY FUND*

### Annex Figure 10. Impulse response functions from credit boom shocks with a relaxed credit boom threshold

### Annex Figure 10. Impulse response functions from credit boom shocks with a relaxed credit boom threshold

### Impulse response visualization and estimation note
- The red real line depicts the estimated 훾_h + 훿_h, and the blue dashed line depicts the estimated 훾_h.
- The shaded area represents 95% confidence interval for 훾_h + 훿_h, computed from Driscoll and Kraay (1998)’s robust standard error.

### Key results reported in Annex Table 7 (Effects of credit inclusion booms and credit booms on financial instability with a relaxed credit boom threshold)
- Note: +, *, and ** denote significance at the 10%, 5%, 1% level, respectively. The outcomes are transformed so the presented estimates are read in %. The standard errors are based on Driscoll and Kraay (1998).

Panel A: Model with Interaction — Dependent variables: Non-Performing Loan Rates (t+1 to t+4) and Bank Z-Score (t+1 to t+4)
- Boom (NPL t+1 to t+4): 0.36 (0.23); 0.48 (0.43); 0.90 (0.68); 1.46 (1.26)
- FIBoom (NPL t+1 to t+4): 0.07 (0.26); -0.47 (0.40); -0.76 (0.72); -0.75 (0.77)
- Boom*FIBoom (NPL t+1 to t+4): 0.65 (0.47); 1.56* (0.74); 1.56 (0.99); 1.04 (1.15)
- Boom (Z-Score t+1 to t+4): 0.43 (0.36); 0.65 (0.43); 0.47 (0.37); 0.23 (0.55)
- FIBoom (Z-Score t+1 to t+4): 0.10 (0.20); 0.09 (0.27); 0.27 (0.31); 0.36 (0.29)
- Boom*FIBoom (Z-Score t+1 to t+4): -0.79 (0.51); -0.69 (0.55); -0.54 (0.56); 0.48 (0.41)
- Observations (NPL): 349; 299; 250; 203
- Countries (NPL): 52; 50; 47; 42
- Observations (Z-Score): 445; 390; 338; 292
- Countries (Z-Score): 59; 55; 50; 49
- R2 (NPL): 0.026; 0.045; 0.057; 0.061
- R2 (Z-Score): 0.149; 0.233; 0.2; 0.345

Panel B: Model without Interaction — Dependent variables: Non-Performing Loan Rates (t+1 to t+4) and Bank Z-Score (t+1 to t+4)
- Boom (NPL t+1 to t+4): 0.60** (0.22); 1.06** (0.45); 1.46+ (0.80); 1.85 (1.36)
- FIBoom (NPL t+1 to t+4): 0.16 (0.23); -0.23 (0.39); -0.50 (0.64); -0.57 (0.66)
- Boom (Z-Score t+1 to t+4): 0.15 (0.24); 0.40 (0.32); 0.27 (0.38); 0.39 (0.52)
- FIBoom (Z-Score t+1 to t+4): -0.02 (0.18); -0.02 (0.25); 0.18 (0.27); 0.44+ (0.24)
- Observations (NPL): 349; 299; 250; 203
- Countries (NPL): 52; 50; 47; 42
- Observations (Z-Score): 445; 390; 338; 292
- Countries (Z-Score): 59; 55; 50; 49
- R2 (NPL): 0.023; 0.034; 0.048; 0.057
- R2 (Z-Score): 0.144; 0.229; 0.198; 0.343

### Annex 4 — Riskiness of Credit Allocation (RCA): construction and data scope
- RCA constructed for four firm-level vulnerability indicators:
  - rca_lev: leverage (total debt to total assets)
  - rca_debt: debt overhang (total debt to EBITDA)
  - rca_icr: interest coverage ratio (EBITDA to interest expenses)
  - rca_edf: expected default frequency (EDF)
- Construction steps:
  - Each year, each firm assigned a decile (1 to 10) of the indicator distribution in its country; higher decile = larger vulnerability.
  - Firms sorted by changes in net debt to lagged total assets into five equal-sized bins; top issuers = largest increases in debt; bottom issuers = largest decreases.
  - RCA measure = average vulnerability decile of top issuers minus average vulnerability decile of bottom issuers.
- Sample considerations:
  - RCA series use listed firms mainly from middle-to-high income countries; fewer observations and countries than npl and zscore when matched to credit inclusion measure.
  - Relationship between credit inclusion measure and RCA series is subtle; unclear whether listed firms in RCA sample contain new borrowers captured by FI.

### Summary statistics for RCA series when matched to main data (Annex Table reported values)
- rca_lev: N 252; Countries 21; Mean -0.11; Std. Dev. 0.96; Min -2.7; Pctl. 25 -0.75; Pctl. 75 0.43; Max 3.3
- rca_debt: N 250; Countries 21; Mean -0.18; Std. Dev. 0.87; Min -3.1; Pctl. 25 -0.68; Pctl. 75 0.33; Max 2.3
- rca_icr: N 251; Countries 21; Mean -0.06; Std. Dev. 0.81; Min -2.2; Pctl. 25 -0.57; Pctl. 75 0.47; Max 3.4
- rca_edf: N 244; Countries 21; Mean 0.0021; Std. Dev. 0.96; Min -3.2; Pctl. 25 -0.62; Pctl. 75 0.63; Max 3.2

### Findings from RCA impulse response analysis (Annex Figures 11–12 and Annex Table 9)
- General:
  - Impulse responses from credit inclusion boom shocks for the four RCA series are generally noisy; possibly due to weak association between RCA measures and credit inclusion measure.
- Leverage-based RCA (strongest signal):
  - Evidence that credit inclusion booms during credit booms increase RCA (i.e., leverage of top issuers expands relative to bottom issuers), interpreted as mounting financial strains.
- EDF-based RCA:
  - Observed weak and opposite effects: credit inclusion booms during credit booms lead to decreases in the RCA measure; given that rapid decreases in RCA measures are often accompanied by sharp economic downturns, this can be interpreted as evidence of negative effects on financial stability.
- Effects of credit booms (Annex Table 9 and Annex Figure 12):
  - Credit booms during credit inclusion booms lead to significant reductions across all RCA measures, possibly indicating immediate financial instability and economic downturns; consistent with Section 3.1 results showing negative effects of credit booms on subsequent GDP growth rates.

Annex Table 9 — Selected coefficients and model fit (Panels A and B; dependent variables are RCA measures across t+1 to t+4)
- Panel A: Leverage-Based RCA (t+1 to t+4)
  - Boom: -0.80** (0.20); -1.14* (0.50); -1.57** (0.46); -1.44** (0.42)
  - FIBoom: -0.08 (0.21); -0.04 (0.21); 0.32 (0.22); -0.08 (0.28)
  - Boom*FIBoom: 0.66* (0.29); 0.08 (0.64); 0.85 (0.65); 0.45 (0.36)
  - Observations: 152; 135; 118; 101
  - Countries: 17; 17; 17; 16
  - R2: 0.432; 0.5; 0.455; 0.496
- Panel A: Debt-Overhang-Based RCA (t+1 to t+4)
  - Boom: 0.12 (0.39); 0.11 (0.47); 0.13 (0.76); 0.08 (0.98)
  - FIBoom: -0.04 (0.18); 0.23 (0.21); 0.36 (0.30); 0.06 (0.31)
  - Boom*FIBoom: -0.29 (0.42); -0.95 (0.63); -0.57 (0.64); -0.15 (0.55)
  - Observations: 150; 133; 116; 99
  - Countries: 17; 17; 17; 16
  - R2: 0.32; 0.391; 0.408; 0.253
- Panel B: ICR-Based RCA (t+1 to t+4)
  - Boom: -0.12 (0.11); 0.17 (0.23); -0.01 (0.46); 0.09 (0.64)
  - FIBoom: -0.04 (0.16); 0.34 (0.25); 0.36+ (0.21); 0.03 (0.25)
  - Boom*FIBoom: 0.05 (0.25); -0.85+ (0.48); -0.43 (0.52); -0.10 (0.59)
  - Observations: 152; 135; 118; 101
  - Countries: 17; 17; 17; 16
  - R2: 0.373; 0.444; 0.456; 0.409
- Panel B: EDF-Based RCA (t+1 to t+4)
  - Boom: -0.43 (0.27); 0.17 (0.44); 0.62 (0.80); -0.28 (0.65)
  - FIBoom: -0.22 (0.19); 0.06 (0.26); 0.18 (0.33); -0.50 (0.30)
  - Boom*FIBoom: -0.72 (0.82); -1.12** (0.37); -0.74 (0.63); -0.29 (1.01)
  - Observations: 142; 127; 112; 97
  - Countries: 16; 16; 16; 16
  - R2: 0.488; 0.475; 0.473; 0.643

*Financial Inclusion, Credit Booms, and Financial Stability Risk — Working Paper No. WP/2026/008*

---


_Source: https://www.imf.org/-/media/files/publications/wp/2026/english/wpiea2026008-source-pdf.pdf_