## wp18153

## Source details

**Canonical URL:** [wp18153](https://www.imf.org/-/media/files/publications/wp/2018/wp18153.pdf)

## Other formats

- [Markdown version](/-/media/files/publications/wp/2018/wp18153.pdf.md)
- [Structured JSON version](/-/media/files/publications/wp/2018/wp18153.pdf.json)

---

### Data, sample, and primary constructs
- Sample: 60 PRGT-eligible low-income economies, n = 1,550 observations over 1981-2015.
- Crises in sample: 29 economies experienced 35 banking crises in the 1981-2015 period.
- Countries dropped: 13 countries omitted due to insufficient control-variable data; none of the 13 dropped experienced a banking crisis in 1981-2015.
- Crisis and auxiliary datasets:
  - Banking crises start-year data: 2017 update to Laeven and Valencia (2013) covering up to 2015.
  - Commodity prices: IMF Primary Commodity Prices (monthly), fixed weights from Gruss (2014).
  - Controls: World Bank WDI, IMF WEO, IMF FFA, OECD DAC, FRED, UCDP/PRIO Armed Conflict Dataset (version 4-2016), deposit insurance database (Demirguc-Kunt et al. 2008), Gerling et al. (2017) fiscal crises.
- Transformations and constructions:
  - Most explanatory variables transformed into MA(3) processes; dependent variable (banking crisis start year) not MA(3) transformed.
  - Ratio or growth variables expressed in percent values if not already.
  - Credit bonanzas defined as deviations from HP-filtered trend with thresholds at 1 or 2 standard deviations; surges defined by top 30th percentile country and sample thresholds.

### Key descriptive statistics (selected exact values, full sample 1981-2015, N = 60, n = 1,550)
- Banking Crisis dummy (†): mean = 0.02300, median = 0.0, SD = 0.01
- Aggregate Commodity Price Growth: mean = 0.076, median = 0.003, SD = 2.418, Min = -20.194, Max = 22.325
- Price Growth Volatility: mean = 0.305, median = 0.020, SD = 1.086, Min = 0.001, Max = 16.090
- Change in Credit/GDP (in %): mean = -9.144, median = -5.106, SD = 19.366, Min = -113.634, Max = 37.959
- M2/Reserves (in %): mean = 6.289, median = 2.907, SD = 10.686, Min = 0.806, Max = 70.697
- Public Debt/GDP (in %): mean = 77.393, median = 61.137, SD = 55.066, Min = 16.296, Max = 284.253
- Inflation (in %): mean = 11.640, median = 7.902, SD = 13.384, Min = -3.887, Max = 84.504
- Real GDP growth (in %): mean = 3.872, median = 3.976, SD = 3.095, Min = -6.580, Max = 14.074
- Net Capital Inflows/GDP (in %): mean = 3.274, median = 2.099, SD = 4.878, Min = -7.651, Max = 20.489
- Overseas Development Assistance/GNI (in %): mean = 11.556, median = 10.097, SD = 8.019, Min = 0.915, Max = 41.237
- Trade Openness (in %): mean = 55.856, median = 49.671, SD = 26.725, Min = 15.894, Max = 124.744
- US Treasury rate (10-yr) (year-end): mean = 6.019, median = 5.547, SD = 2.827, Min = 2.237, Max = 12.257
- Deposit insurance dummy: mean = 0.13900
- Currency crisis (start year) dummy: mean = 0.03400
- Fiscal crisis (start year) dummy: mean = 0.09400
- Credit/GDP Bonanza (1sd) dummy: mean = 0.00500
- Credit/GDP Surge (at time t) dummy: mean = 0.17800
- Net Capital Inflow/GDP Surge (at time t) dummy: mean = 0.23100

### Main empirical findings (marginal effects reported as 1 sd changes; model and robustness highlights)
- Table 1 — Main results (marginal effects for a 1sd increase; unconditional crisis probability benchmarks):
  - Unconditional Crisis Probability: 2.3% (columns (1)-(5)), 4.3% (column (6)).
  - Commodity Price Growth marginal effects (in percent, winsorized, MA(3) transformed):
    - Column (1): -0.622 (3.05)***
    - Column (2): -1.046 (3.30)***
    - Column (3): -0.836 (2.53)**
    - Column (4): -1.381 (2.28)**
    - Column (5): -1.026 (2.82)***
    - Column (6): -6.859 (0.41)
  - Commodity Price Growth Volatility marginal effects:
    - Column (1): 1.964 (2.37)**
    - Column (2): 1.463 (2.68)***
    - Column (3): 1.577 (1.46)
    - Column (4): 0.710 (0.37)
    - Column (5): -0.624 (1.32)
    - Column (6): 1.734 (0.06)
  - Observations: 1,550 (columns (1)-(5)); 804 (column (6)).
  - Countries: 60 (columns (1)-(5)); 29 (column (6)).
  - Crises: 35 (columns (1)-(5)); 35 (column (6)).
  - AUROC values:
    - Column (1): 0.799, se(AUROC) 0.028
    - Column (2): 0.822, se(AUROC) 0.026
    - Column (3): 0.858, se(AUROC) 0.025
    - Column (4): 0.882, se(AUROC) 0.024
    - Column (5): 0.829, se(AUROC) 0.030
    - Column (6): 0.748, se(AUROC) 0.041
  - ROC comparison p-values reported (examples): (3) vs (1): 0.008; (5) vs (4): 0.007; (6) vs (4): 0.024.
- Robustness (Table 2 — RE-Mundlak Logit and weighting/specification checks):
  - Unconditional Crisis Probability: 2.3% (columns (1),(2),(5)), 2.2% (column (3)), 4.2% (column (4)).
  - Commodity Price Growth marginal effects across specifications:
    - Column (1): -1.381 (2.28)**
    - Column (2): -1.153 (1.88)*
    - Column (3): -1.348 (2.65)***
    - Column (4): -1.987 (1.91)*
    - Column (5): -0.990 (3.13)***
  - AUROC values range: 0.882, 0.877, 0.873, 0.838, 0.834 (se(AUROC) reported).
- Alternative MA-transformations (Table 3):
  - Commodity Price Growth marginal effects become more negative with longer MA windows:
    - MA(3) (column (3)): -1.381 (2.28)**
    - MA(4) (column (4)): -1.705 (4.52)***
    - MA(5) (column (5)): -2.214 (3.43)***
  - AUROC across MA specifications: 0.860 (MA lagged), 0.866 (MA(2)), 0.882 (MA(3)), 0.898 (MA(4)), 0.896 (MA(5)).
- The role of leverage and capital flows (Table 4 and Table 6):
  - Net Capital Inflows/GDP: positive and statistically significant in multiple specifications (examples: 1.112 (2.15)**; 0.992 (2.30)**; Table 6 inclusion: 0.100 (2.45)** for a logit coefficient).
  - M2/Reserves: positive and statistically significant in many specifications (examples: 1.049 (2.47)**; 1.124 (2.63)***).
- Credit bonanzas and surges (Table 5):
  - Credit Bonanza and Surge counts (per columns (1)-(7)): 111, 131, 131, 129, 82, 98.
  - Commodity Price Growth logit coefficients (columns (1)-(7)): -0.269 (2.20)**; -0.284 (1.81)*; -0.302 (1.98)**; -0.295 (2.40)**; -0.308 (2.73)***; -0.289 (2.41)**; -0.302 (2.72)***.
  - AUROC values across columns ≈ 0.882–0.887 (se(AUROC) reported).
- Core patterns retained across commodity-weighting schemes:
  - Commodity price growth generally exhibits a negative association with banking crisis probability across many specifications and weighting schemes.
  - Net capital inflows and M2/reserves typically exhibit positive associations with crisis probability.
  - Commodity price growth volatility shows mixed signs and statistical significance across specifications.

### Model performance, diagnostics, and statistical notes
- AUROC and model fit:
  - AUROC values vary by specification; representative values: 0.799, 0.822, 0.858, 0.882, 0.829, 0.748 (Table B-1 columns (1)–(6)); se(AUROC) values reported per model.
  - Log-likelihood examples (Table B-1): LogL = -148.20, -143.38, -135.21, -127.15, -141.40, -84.05.
  - Wald χ2 statistics and p-values reported for FE and related specifications (examples): Waldχ2 (FE) = 13.55 (p = 0.035), 16.05 (p = 0.042), 42.67 (p = 0.000), 2.98 (p = 0.000) in selected columns.
- Marginal effects:
  - All presented marginal effects are computed as marginal effects multiplied by the covariate standard deviation (1 sd increase).
  - Delta-method standard errors used for logit-based marginal effects where applicable; in parsimonious FE logit models standard errors for some marginal effects could not be obtained.
- Sample composition effects:
  - Some FE Logit specifications have smaller samples (e.g., only countries with at least one banking crisis).
  - MA(3) transformation causes some dummy indicators to show maxima of 0.33 or 0.67 (equal to 1 in one or two of three consecutive years).

### Synthesis of substantive findings (policy-relevant implications implied by empirical results)
- Commodity price growth:
  - A 1 sd increase in aggregate commodity price growth is associated with a reduction in unconditional banking crisis probability in many specifications (examples: -0.622, -1.046, -1.381 marginal-effect estimates reported), robust across MA transformations and weighting schemes.
- Capital inflows and liquidity:
  - Higher Net Capital Inflows/GDP and higher M2/Reserves are associated with higher banking crisis probability in multiple specifications, indicating external flow surges and domestic liquidity expansion can elevate systemic banking risk.
- External conditions and global rates:
  - The 10-yr US Treasury rate (year-end) enters positively and significantly in multiple specifications (examples: 1.318 (3.84)***; 1.596 (3.45)***; 1.463 (2.95)***), linking global financial conditions to crisis risk in low-income economies.
- Heterogeneity and regime considerations:
  - Effects vary by sub-samples and regimes (Table 7): unconditional crisis probabilities and commodity-growth marginal effects differ across regimes and country subsets; AUROC values remain high across many regime specifications.

*Source: wp18153 - References and empirical appendix (extracted content).*

### References

### References

### Literature themes and empirical focus
- Topics covered by cited literature:
  - Commodity prices and their macroeconomic and financial effects (e.g., commodity prices and bank lending; commodity prices and civil war; commodity prices and growth).
  - Banking crises: frequency, determinants, systemic crises, booms and busts, and early warning systems.
  - Capital flows: surges, bonanzas, FDI vs non-FDI inflows, and their relationship to crises.
  - Credit booms, leverage, and financial cycles across long historical samples.
  - Methodological contributions on panel models, rare events logit, discrete-choice models, and MA-transformations.
- Representative methodological and empirical references (authors and themes preserved as in source):
  - Agarwal, I., Duttagupta, R., and Presbitero, A. F. (2017) — Commodity prices and bank lending in low-income countries. International Monetary Fund, Working Paper WP/17/279.
  - Aizenman, J. and Noy, I. (2013) — Macroeconomic adjustment and the history of crises in open economies.
  - Anundsen, A. K., Gerdrup, K., Hansen, F., and Kragh-Sørensen, K. (2016) — Bubbles and crises: The role of house prices and credit.
  - Caballero, J. A. (2016) — Do surges in international capital inflows influence the likelihood of banking crises?
  - Laeven, L. and Valencia, F. (2013) — Systemic banking crises database.
  - Jordà, Ò., Schularick, M., and Taylor, A. M. (2011, 2013, 2015, 2016) — Financial crises, credit booms, leveraged bubbles, sovereigns versus banks.
  - Ilzetzki, E., Reinhart, C. M., and Rogoff, K. S. (2017) — Exchange arrangements entering the 21st century.
  - King, G. and Zeng, L. (2001) — Logistic regression in rare events data.
  - Park, J. Y. and Phillips, P. C. B. (2000) — Nonstationary binary choice.

### Figures and key descriptive notes (preserved values and captions)
- Figure 1: Banking Crises – Frequency Analysis (1981-2015)
  - N= 60 PRGT-eligible economies.
  - Sample period: 1981-2015; 29 countries in sample experienced crises.
  - Count (out of 60 Countries) axis and Commodity price index (2010=100) shown in original figure.
- Figure 2: Aggregate Commodity Price Growth
  - Annualised aggregate commodity price growth (median and various percentiles; multiplied by 100) for N= 60 PRGT-eligible economies.
  - Monthly sample covers the period 1981-2016.
  - Fixed weights used to aggregate individual commodity price series are the net export/GDP means over the entire time horizon.
- Figure 3: Banking Crises – Event Analysis
  - Selected event analysis plots for period 1981-2015.
  - Estimates derived from crisis dummy lags and leads in a pooled robust regression with country fixed effects (equation (2)), 90% confidence intervals.
  - Samples analysed exclude ongoing crisis years. Variables are winsorized; no MA(3) transformation in these plots.
- Figure 4: ROC Plots
  - Panel (a): compares model with commodity price variables (black line) and without (orange line) for full sample (specification in column (4) of Table 1).
  - Panel (b): compares three estimators for full model: (i) pooled logit model (orange line), (ii) FE Logit model (dashed grey line), (iii) RE-Mundlak model (black line).
  - FE Logit model has smaller sample (only countries with at least one banking crisis).
  - Reported p-values indicate statistically significant increase in predictive power between RE-Mundlak benchmark and alternative models.

### Selected empirical magnitudes and diagnostics (preserved exactly from tables)
- Table 1 — Main Results (economic magnitudes for a 1sd increase in covariate; unconditional crisis probability benchmarks):
  - Unconditional Crisis Probability: 2.3% (columns (1)-(5)), 4.3% (column (6)).
  - Commodity Price Growth marginal effects (in percent, winsorized, MA(3) transformed):
    - Column (1): -0.622 (3.05)***
    - Column (2): -1.046 (3.30)***
    - Column (3): -0.836 (2.53)**
    - Column (4): -1.381 (2.28)**
    - Column (5): -1.026 (2.82)***
    - Column (6): -6.859 (0.41)
  - Commodity Price Growth Volatility marginal effects:
    - Column (1): 1.964 (2.37)**
    - Column (2): 1.463 (2.68)***
    - Column (3): 1.577 (1.46)
    - Column (4): 0.710 (0.37)
    - Column (5): -0.624 (1.32)
    - Column (6): 1.734 (0.06)
  - Observations: 1,550 (columns (1)-(5)); 804 (column (6)).
  - Countries: 60 (columns (1)-(5)); 29 (column (6)).
  - Crises: 35 (columns (1)-(5)); 35 (column (6)).
  - AUROC values:
    - Column (1): 0.799, se(AUROC) 0.028
    - Column (2): 0.822, se(AUROC) 0.026
    - Column (3): 0.858, se(AUROC) 0.025
    - Column (4): 0.882, se(AUROC) 0.024
    - Column (5): 0.829, se(AUROC) 0.030
    - Column (6): 0.748, se(AUROC) 0.041
  - ROC comparison p-values shown (e.g., (3) vs (1): 0.008; (5) vs (4): 0.007; (6) vs (4): 0.024).
- Table 2 — Robustness checks (RE-Mundlak Logit):
  - Unconditional Crisis Probability: 2.3% (columns (1),(2),(5)), 2.2% (column (3)), 4.2% (column (4)).
  - Commodity Price Growth marginal effects across specifications:
    - Column (1): -1.381 (2.28)**
    - Column (2): -1.153 (1.88)*
    - Column (3): -1.348 (2.65)***
    - Column (4): -1.987 (1.91)*
    - Column (5): -0.990 (3.13)***
  - Observations range: 1,550; 1,455; 1,619; 759; 1,550.
  - Countries: 60; 52; 60; 46; 60.
  - Crises: 35; 33; 35; 32; 35.
  - AUROC values: 0.882, 0.877, 0.873, 0.838, 0.834 with se(AUROC) reported per column.
- Table 3 — Alternative MA-transformation (RE-Mundlak Logit):
  - MA(3) (column (3)) commodity price growth: -1.381 (2.28)**
  - MA(4) (column (4)): -1.705 (4.52)***
  - MA(5) (column (5)): -2.214 (3.43)***
  - AUROC across MA specifications: 0.860 (MA lagged), 0.866 (MA(2)), 0.882 (MA(3)), 0.898 (MA(4)), 0.896 (MA(5)).
- Table 4 — The Role of Leverage:
  - Commodity Price Growth (column (1)): -1.381 (2.28)**
  - Net Capital Inflows/GDP marginal effects in columns shown (e.g., column (1): 1.112 (2.15)**; column (2): 0.992 (2.30)**).
  - AUROC values: 0.774, 0.801, 0.842, 0.872, 0.882 with se(AUROC) reported.
- Table 5 — Credit Bonanzas and Surges (raw logit coefficients):
  - Credit Bonanza and Surge counts across definitions: 111, 131, 131, 129, 82, 98 (per columns (1)-(7) respectively).
  - Commodity Price Growth logit coefficients (columns (1)-(7)): -0.269 (2.20)**; -0.284 (1.81)*; -0.302 (1.98)**; -0.295 (2.40)**; -0.308 (2.73)***; -0.289 (2.41)**; -0.302 (2.72)***.
  - AUROC values across columns ~0.882–0.887 with se(AUROC) reported.
- Table 6 — Capital Inflow Bonanzas and Surges (raw logit coefficients):
  - Panel A bonanza/surge counts and Panel B FDI/non-FDI counts preserved as in table.
  - Commodity Price Growth coefficients across columns range roughly -0.287 to -0.244 with associated t-statistics in parentheses (e.g., (2.47)**).
  - Including net capital inflows/GDP coefficient in (2): 0.100 (2.45)**.
  - AUROC values across panels ~0.874–0.899 with se(AUROC) reported.
- Table 7 — Focus on Alternative Regimes (RE-Mundlak Logit):
  - Banking Crises in Regime: 352, 314, 351, 2 (per columns).
  - Countries in Regime: 60, 38, 31, 59, 14.
  - Observations in Regime: 1,550; 1,157; 826; 1,516; 349.
  - Unconditional Crisis Prob in Regime: 2.3%; 2.0%; 1.7%; 2.3%; 3.4%.
  - Unconditional Crisis Prob in Base: 3.1%; 2.9%; 2.0%.
  - Commodity Price Growth marginal effects presented (e.g., column (1): -1.062 (2.10)**; column (2): -1.099 (2.11)**).
  - AUROC values around 0.872–0.878 with se(AUROC) reported.

### Appendix — Data sources and sample construction (preserved terminology and datasets)
- Crisis Data:
  - Banking crises start-year data taken from a 2017 update to Laeven and Valencia (2013) covering banking crises up to 2015.
- Commodity Price Data:
  - Aggregate commodity price growth and volatility constructed from IMF Primary Commodity Prices (monthly data) using fixed weights from Gruss (2014).
  - Weighting, data filtering and transformation details are described in the main text.
- Controls and macrofinancial variables (data sources and variable definitions):
  - World Bank World Development Indicators (WDI): Real GDP growth (in %), Inflation (GDP deflator, in %), M2/reserves (in %), short-term debt (in % of total external debt), size (broad money in % of GDP), overseas development assistance (foreign aid) as a share of GNI, depreciation (growth rate of the annual LCU-US$ exchange rate), total debt service (in % of exports of goods, services and primary income), and trade openness (Merchandise trade in % of GDP).
  - Domestic credit to the private sector: change of domestic credit-to-GDP ratio (in %), with credit/GDP from WDI integrated with FinStats and Global Financial Development Database (GFDD).
  - Public debt to GDP: IMF World Economic Outlook database.
  - Net capital inflows/GDP: IMF Financial Flows Analytics (FFA) Database — specifically “Total Net Nonofficial Inflows, in percent of GDP in U.S. Dollars.” FDI and non-FDI net flows constructed from same dataset.
  - Herfindahl-Hirschman index for ODA donor share: OECD DAC data.
  - IMF International Financial Statistics lines 22d, 24, and 25 used to compute liquidity = claims (22d) divided by demand deposits (24) and other deposits (25).
  - Conflict: UCDP/PRIO Armed Conflict Dataset (version 4-2016) covering 1946-.

*Source: wp18153 - References (extracted content).*

### 2015.   We  code  countries  as  being  in  conflict  if  they  have  an  intensity  score  of  2.   The  deposit

### wp18153 - 2015

### Data, sample, and primary constructs
- Sample: 60 PRGT-eligible low-income economies, n = 1,550 observations over 1981-2015.
- Crises in sample: 29 economies experienced 35 banking crises in the 1981-2015 period.
- Countries dropped: 13 countries omitted due to insufficient control-variable data (primarily fragile states and small island states); none of the 13 dropped experienced a banking crisis in 1981-2015.
- Crisis coding and related datasets:
  - Countries coded as being in conflict if intensity score = 2.
  - Deposit insurance dummy based on July 2015 update of the ‘Deposit Insurance Database’ by Demirguc-Kunt et al. (2008).
  - Fiscal crisis events from Gerling et al. (2017) database.
  - Currency crises from update of Laeven and Valencia (2013) dataset (crisis start year dummies constructed).
  - 10-year Treasury constant maturity date taken from FRED; year-end value on final trading day used.
  - Net foreign assets/GDP (separating FDI from non-FDI) computed from updated External Wealth of Nations database (Lane and Milesi-Ferretti, 2007).
- Transformations: all ratio or growth variables transformed into percent values if not already; most variables transformed into MA(3) processes (except banking crisis start year dependent variable).
- Credit/capital inflow episodes:
  - Bonanzas: deviations from long-run HP-filtered trend of credit volumes or capital inflows; thresholds at 1 or 2 standard deviations (2 sd threshold for credit bonanzas yields only a single event and is not analysed).
  - Surges: observations both in top 30th percentile of an individual country and top 30th percentile of the full 60-country sample; indicators constructed from annual flow data (not winsorized, MA-transformed data).

### Sample composition (regression sample makeup)
- Full regression sample: 60 countries, 1,550 observations over 1981-2015.
- Observations per country vary; Table A-2 lists ISO codes and crisis counts for all 60 countries.
- The MA(3) transformation causes some dummy indicators to show maxima of 0.33 or 0.67 (equal to 1 in one or two of three consecutive years).

### Descriptive statistics (selected values from Table A-1; full sample 1981-2015, N = 60, n = 1,550)
- Banking Crisis dummy (†): mean = 0.02300, median = 0.0, SD = 0.01
- Aggregate Commodity Price Growth: mean = 0.076, median = 0.003, SD = 2.418, Min = -20.194, Max = 22.325
- Price Growth Volatility: mean = 0.305, median = 0.020, SD = 1.086, Min = 0.001, Max = 16.090
- Change in Credit/GDP (in %): mean = -9.144, median = -5.106, SD = 19.366, Min = -113.634, Max = 37.959
- M2/Reserves (in %): mean = 6.289, median = 2.907, SD = 10.686, Min = 0.806, Max = 70.697
- Public Debt/GDP (in %): mean = 77.393, median = 61.137, SD = 55.066, Min = 16.296, Max = 284.253
- Inflation (in %): mean = 11.640, median = 7.902, SD = 13.384, Min = -3.887, Max = 84.504
- Real GDP growth (in %): mean = 3.872, median = 3.976, SD = 3.095, Min = -6.580, Max = 14.074
- Net Capital Inflows/GDP (in %): mean = 3.274, median = 2.099, SD = 4.878, Min = -7.651, Max = 20.489
- Overseas Development Assistance/GNI (in %): mean = 11.556, median = 10.097, SD = 8.019, Min = 0.915, Max = 41.237
- Trade Openness (in %): mean = 55.856, median = 49.671, SD = 26.725, Min = 15.894, Max = 124.744
- Debt service (in % of exports): mean = 13.677, median = 10.431, SD = 10.916, Min = 0.814, Max = 49.937
- Short-term debt share of total external debt (in %): mean = 21.075, median = 16.149, SD = 15.318, Min = 2.893, Max = 61.756
- Exchange Rate Depreciation (in %): mean = 7.180, median = 3.314, SD = 12.658, Min = -9.972, Max = 67.450
- Liquidity (in %): mean = 82.438, median = 74.568, SD = 42.321, Min = 20.763, Max = 231.220
- Size (in %): mean = 34.704, median = 27.562, SD = 21.142, Min = 9.907, Max = 94.461
- Deposit insurance dummy: mean = 0.13900
- Currency crisis (start year) dummy: mean = 0.03400
- Fiscal crisis (start year) dummy: mean = 0.09400
- US Treasury rate (10-yr) (year-end): mean = 6.019, median = 5.547, SD = 2.827, Min = 2.237, Max = 12.257
- Net foreign FDI assets/GDP (in %): mean = -23.852, median = -12.771, SD = 28.210, Min = -119.946, Max = 0
- Net foreign non-FDI assets/GDP (in %): mean = -44.375, median = -31.909, SD = 51.813, Min = -206.331, Max = 41.967
- Credit/GDP Bonanza (1sd) dummy: mean = 0.00500
- Credit/GDP Surge (at time t) dummy: mean = 0.17800
- Net Capital Inflow/GDP Surge (at time t) dummy: mean = 0.23100
- Net FDI Inflow/GDP Surge (at time t) dummy: mean = 0.22200
- Net non-FDI Inflow/GDP Surge (at time t) dummy: mean = 0.23700
- Notes: dummy/indicator variables marked with D; MA(3) transformation explains 0.33 and 0.67 maxima for some dummies.

### Key empirical findings (selected from regression tables B-1 through B-5)
- Commodity price growth:
  - Negative and statistically significant marginal effects in many specifications.
  - Examples from Table B-1 (marginal effects × covariate SD): column (1): -0.622 (t = 3.05)***; column (2): -1.046 (t = 3.30)***; column (4): -1.381 (t = 2.28)**.
  - Effect sizes reported in robustness tables vary with weighting scheme (e.g., Table B-3 effect sizes: Commodity Price Growth: -0.39 to -1.83 across models; Table B-4 effect sizes: -0.27 to -0.90).
- Commodity price growth volatility:
  - Mixed signs; in Table B-1 column (2) Commodity Price Growth Volatility = 1.964 (t = 2.37)**; in other specs volatility often smaller or insignificant.
- Net Capital Inflows/GDP:
  - Positive and statistically significant in multiple specifications.
  - Table B-1 column (1): 0.992 (t = 2.30)**; Table B-2 several columns also show positive significance (e.g., 1.159 (t = 2.58)**, 1.249 (t = 2.73)***).
- M2/Reserves:
  - Positive and statistically significant in many specifications.
  - Table B-1 columns: 1.049 (t = 2.47)**; 1.124 (t = 2.63)***.
- Official development assistance / Foreign Aid (ODA/GNI):
  - Positive and statistically significant in multiple specifications.
  - Table B-1 column (1): 2.122 (t = 3.25)***.
- 10-yr US Treasury Rate (year-end):
  - Positive and statistically significant across many specifications.
  - Table B-1 columns show values: 1.318 (t = 3.84)***; 1.596 (t = 3.45)***; 1.463 (t = 2.95)***; 1.872 (t = 3.70)***; 1.026 (t = 2.75)***; 9.785 (t = 2.03)** (column (6), FE).
- Trade Openness:
  - Negative and statistically significant in several specifications (e.g., Table B-1 column (1): -2.592 (t = 2.60)***).
- Inflation, Public Debt/GDP, Credit/GDP Growth:
  - Inflation: positive, occasionally significant at 10% (e.g., Table B-1 column (2): 1.498 (t = 1.71)*).
  - Public Debt/GDP: generally positive but not consistently statistically significant.
  - Credit/GDP Growth: positive but not consistently statistically significant in many specifications.
- Model performance and diagnostics:
  - AUROC values vary by specification; examples from Table B-1: 0.799, 0.822, 0.858, 0.882, 0.829, 0.748 (columns (1)–(6)); se(AUROC) values reported (e.g., 0.028, 0.026, 0.025, 0.024, 0.030, 0.041).
  - Log-likelihoods reported (examples): LogL = -148.20, -143.38, -135.21, -127.15, -141.40, -84.05 (Table B-1 columns (1)–(6)).
  - Wald χ2 statistics reported for FE models (e.g., Waldχ2 (FE) = 13.55, 16.05, 42.67, 2.98 in selected columns) with corresponding Wald p-values (e.g., 0.035, 0.042, 0.000, 0.000).
- Robustness to commodity-weighting schemes:
  - Multiple tables present results using base year commodity net exports/GDP weights (Table B-3), mean commodity exports/GDP weights (Table B-4), and annual net commodity exports/GDP weights (Table B-5).
  - Core qualitative findings (commodity price growth negative association with crisis probability; net capital inflows and M2/reserves positive association) are present across many weighting schemes and specifications, though magnitudes and statistical significance vary by specification.

### Additional analytic notes and transformations
- All presented marginal effects are computed as marginal effects multiplied by the covariate standard deviation (1 sd increase).
- The MA(3) transformation applied to nearly all explanatory variables; dependent variable (banking crisis start year) is not MA(3) transformed.
- Global variables (in some robustness checks) represent cross-section averages for the full sample excluding country i.
- Delta-method standard errors used for logit-based marginal effects where applicable; in parsimonious FE logit models standard errors for some marginal effects could not be obtained.

*Source: wp18153 - 2015 (IMF working paper content provided).*

---


_Source: https://www.imf.org/-/media/files/publications/wp/2018/wp18153.pdf_