Global Financial Cycle Transmission to Indonesian Bank Balance Sheets v2.7 — FULL DRAFT end-to-end

BUK panel evidence, 2002–2025. Seven figures + ten tables + three LaTeX-rendered equations. Full draft complete: Abstract → Conclusion. Last build: 2026-05-27.

📄 BMEB 2026 Submission Package — v2.7 (FULL DRAFT COMPLETE end-to-end)

📖 Read HTML (MDPI-style + KaTeX math) ⬇ MSWord DOCX (4.1 MB) ⬇ PDF (LaTeX-rendered, 28 pp) ⬇ Markdown source

Submit to https://submit.confbay.com/conf/bmeb2026 or email [email protected]. Manuscript: 11 sections + references + Appendix A/B/C · ~10,900 words · 7 embedded figures · 10 tables · 3 LaTeX-rendered equations · 37 verified references · JEL codes G21, F32, E58, G28.

v2.7 milestone (2026-05-27) — FULL DRAFT COMPLETE end-to-end: All 11 sections drafted and tightened through section-by-section vibe-check cycles. Manuscript covers Abstract → Introduction → Related literature (5 strands) → Data (7 sub-sections + Table 1 Data Scope) → Empirical strategy (4 sub-sections + 3 LaTeX equations) → Main results (4 sub-sections, slow-build + accommodation-phase reading) → Heterogeneity (2 sub-sections, recapitalisation signature + FX channel) → Episode event-study (4 sub-sections, three-regime taxonomy) → Robustness (3 sub-sections, BAA-spread shock decomposition) → Asymmetric-macroprudential interpretation (3 sub-sections, "asymmetry by design, not by failure") → Discussion & policy implications (4 sub-sections, exposure-cap toolkit + forbearance budget + credit-spread monitoring) → Conclusion (3 paragraphs, framework-mature-but-calibration-frontier-open closing).

Analytical contributions baked in: (1) slow-build transmission profile with accommodation-phase reading of the GFC factor; (2) asymmetric capital-management triad at bank level (credit ↑ NPL ↓ CAR ↓); (3) recapitalisation signature uniformly across all three domestic ownership classes (not state-specific) → regulator-uniform forbearance hypothesis; (4) trading book retreats before loan book in FX channel; (5) three-regime episode taxonomy with 2008-vs-2022 regime-shift (20-log-point swing); (6) BAA corporate-credit spread is the operative shock, not VIX (R4b/R4c decomposition); (7) post-2016 sensitivity 3-4× pre-2014; (8) framework-level asymmetry by instrument design.

Policy frontier identified: foreign-currency reserve requirements (GWM Valas), surge-cyclical CCB calibration triggered by BAA spread (not credit-to-GDP), and binding FX-exposure caps on Q3/Q4 banks. Plus monitoring-side reform: replace VIX with BAA spread as the cross-border-cycle indicator. Plus forbearance-budget rule for the regulator-uniform recapitalisation pattern.

What remains before submission: (a) ✅ citation-integrity audit COMPLETE 2026-05-27: all 37 references verified via WebSearch / primary-source spot-check; zero phantom citations; zero dead refs; body↔refs cross-check 100%; (b) visual rendering polish in DOCX (table column widths, equation alignment fine-tuning); (c) LaTeX → BMEB-compliant PDF conversion (Task #21, deferred); (d) bos's deep-read pass + Q&A iteration on detailed content.

What this bundle shows — Four stylised-facts charts built from public OJK BUK data plus FRED, covering Indonesia's 97 commercial banks across 24 years. The bundle establishes that (i) global risk-off conditions are well-summarised by a single GFC factor explaining 48 % of cross-indicator variance; (ii) bank-level foreign-currency exposure is heterogeneous and persistent; (iii) capital and credit-quality ratios cluster around healthy levels with episodic stress; (iv) credit and deposits have grown 15× since 2002 with visible deceleration around each major global episode.

Coverage97 BUK banks · 2002–2025 · 24 years

GFC factor (PC1)48.2 % variance · 6 episodes coded

FX exposure proxy2,289 bank-period obs · r = 0.925 cross-validate

Outcome panel2,152 bank-years · assets / credit / DPK

1. Research design — one paragraph

Local-projection (Jordà 2005) bank-level panel regressions of log-credit, NPL, and CAR on (a) the composite GFC factor and (b) episode-dummy event studies, with bank fixed effects and time controls. Heterogeneous effects identified through interactions with bank size (asset quintile), funding structure (DPK share of liabilities), and a foreign-currency exposure proxy built from the off-balance-sheet komitmen-kontinjensi data. The 2002–2025 panel covers six major risk-off windows: GFC 2008-09, Taper tantrum 2013, CNY / commodity bust 2015-16, EM contagion 2018, COVID 2020, Fed hike cycle 2022-23.

2. Figure 1 — Global risk-off indicators and the composite GFC factor

Figure 1. Global risk-off indicators and the composite GFC factor, 2002Q1–2025Q4

Four panels assemble the global financial-cycle inputs: VIX, broad USD index (spliced DTWEXBGS + DTWEXM), Fed Funds rate, US 10-year yield, and the BAA – 10Y corporate-credit spread. Panel (d) plots the first principal component of all five z-scored indicators (48.2 % of variance, sign-aligned so positive = risk-off). Brick shading marks six pre-specified episode windows.

File: figure_1_gfc_factor.png · 13.8 × 8.6 in · 320 DPI · also available as PDF

Table 1 — Summary statistics, global risk-off indicators

Variable	Units	FRED ID	N	Mean	SD	p10	p90
VIX	Index points	VIXCLS	96	19.9	7.4	12.7	30.5
Broad USD index	Index, spliced	DTWEXBGS + DTWEXM	96	104.7	10.2	92.4	121.4
Fed Funds rate	% p.a.	DFF	96	1.79	1.85	0.08	5.05
US 10Y yield	% p.a.	DGS10	96	3.13	1.18	1.59	4.71
BAA – 10Y spread	pp	BAA10Y	96	2.62	1.05	1.74	3.78
Composite GFC factor	z-score	PCA · this paper	96	0.00	1.55	−2.13	1.94

PC1 loadings: VIX +0.41, BAA spread +0.56, FFR −0.55, US10Y −0.41, USD index −0.21. Interpretation: positive PC1 = high VIX, wide credit spread, accommodative US rates (Fed cutting), low long yields — the canonical risk-off signature.

3. Figure 2 — Bank-level FX exposure proxy

Figure 2. Cross-sectional distribution of FX exposure across BUK, 2002–2025

FX exposure = sum of foreign-currency off-balance-sheet commitments (eleven matched account labels in komitmen-kontinjensi) divided by total assets. Panel (a) shows the cross-sectional median + p25–p75 band over time; panel (b) the five persistent-high-exposure banks (foreign branches); panel (c) the share of banks in each FX-exposure quartile; panel (d) data availability by period.

File: figure_2_fx_exposure_dist.png · 320 DPI · cross-validated bulanan-vs-triwulanan Pearson r = 0.925

Stylised fact — median FX exposure ratio rises from ~0.5 % of assets (2002–2015) to ~3 % (2025); the p90 rises from ~40 % to ~140 %. Indonesian banking has gradually FX-financialised, and the heterogeneity across banks has fanned out — a key heterogeneity dimension the paper will exploit.

4. Figure 3 — CAR & NPL gross, bridged panel

Figure 3. Capital adequacy and credit quality of Indonesian commercial banks (BUK), 2002–2025

Two regulatory-template formats bridged: kpmm_konsolidasi & kualitas_aktiva_produktif (Basel I → II.5 era, 2002–2015) and rasio_keuangan & kualitas_aset (Basel III era, 2021–2025). Panels (a) and (b) are within-new-format concordance scatters (r > 0.98 confirms extraction consistency). Panels (c) and (d) show cross-sectional bands over time. Grey 2016–2020 band = parsing gap in the source archive.

File: figure_3_car_npl_panel.png · 1,481 CAR obs + 1,677 NPL obs · 320 DPI

2016 gap (source-side) — OJK CFS public endpoint serves no 2016 data for any BUK report type (verified 2026-05-26 — server returns no_report across all 98 banks). The gap reflects the POJK 14/2017 reporting-format transition. CAR / NPL panel coverage is therefore 2002–2015 ∪ 2017–2025 for the old-format reports (kpmm_konsolidasi, kualitas_aktiva_produktif) and 2021–2025 for the new-format reports (rasio_keuangan, kpmm_car, kualitas_aset). All six risk-off episodes (GFC, Taper, CNY, EM, COVID, Fed hike) remain covered.

5. Figure 4 — Outcome panel: assets, credit, deposits

Figure 4. Aggregate and cross-sectional balance-sheet dynamics of Indonesian commercial banks (BUK), 2002–2025

Continuous annual coverage from posisi keuangan bulanan (April snapshots), with a label splice total aktiva → total aset at 2010 and kredit yang diberikan → kredit over the same window. Panel (a) aggregate banking-system assets; panel (b) the cross-sectional credit-to-asset ratio; panel (c) cross-sectional credit growth YoY; panel (d) aggregate assets, credit, and DPK indexed to 2002 = 100 on a log scale. Grey 2016 band = single missing snapshot.

File: figure_4_outcome_panel.png · 2,152 bank-years for assets; 2,150 for credit and DPK · 320 DPI

Table 4 — Outcome variables, 2002–2025 full-sample

Variable	N	Mean	SD	p10	Median	p90
Total assets (IDR trillion)	2,152	69.2	192.5	1.0	11.5	175.4
Total credit (IDR trillion)	2,150	42.6	122.1	0.6	6.9	110.2
Total DPK (IDR trillion)	2,150	52.4	148.3	0.7	8.5	134.6
Credit / assets (%)	2,150	61.3	15.7	40.8	62.7	78.4
Asset growth YoY (%)	2,050	13.4	20.5	−5.8	10.7	34.6
Credit growth YoY (%)	2,043	14.7	27.9	−10.6	11.9	40.3
DPK growth YoY (%)	2,049	13.5	21.7	−5.2	10.9	34.8

Figures rounded; exact values in table_4_outcome_summary.csv. Total panel = 24 years × ~89 active BUK banks.

6. Figure 5 — Master estimation panel overview

Figure 5. Master estimation panel — overview, descriptives, and headline correlations

The four component streams (outcome panel, GFC shock series, FX exposure proxy, CAR / NPL bridge) merge into a single bank-year master panel. Panel (a) shows the unconditional Pearson correlation matrix; panel (b) the composition of the BUK universe by ownership type (derived from 2021–2024 susunan top-owner); panels (c) and (d) collapse to annual cross-sectional medians and plot against the composite GFC factor with linear-fit overlay and per-year labels.

File: figure_5_panel_overview.png · 320 DPI · 2,152 bank-year obs · 97 banks

Headline annual correlations — GFC factor vs median credit growth: r = +0.19 (N = 22). GFC factor vs median ΔNPL: r = +0.05 (N = 17). The weak unconditional signal is the paper's hook: aggregate Indonesian banking looks unresponsive to global cycle, but the local-projection panel reveals that the average masks sharp heterogeneity across bank-type, size, and FX-exposure dimensions.

Table 5 — Master estimation panel descriptives, 2002–2025 excl. 2016

Variable	N	Mean	SD	p10	Median	p90
Total assets (IDR trillion)	2,152	69.2	192.5	1.0	11.5	175.4
Total credit (IDR trillion)	2,150	42.6	122.1	0.6	6.9	110.2
Credit / assets	2,150	0.61	0.16	0.41	0.63	0.78
Asset growth YoY (∆ln)	2,050	0.134	0.205	−0.058	0.107	0.346
Credit growth YoY (∆ln)	2,043	0.147	0.279	−0.106	0.119	0.403
CAR (%)	1,481	28.4	—	13.7	22.2	45.9
NPL gross (%)	1,677	3.7	—	0.3	2.4	6.7
FX exposure ratio	1,879	—	—	~0.005	~0.03	~1.4
GFC factor (annual)	2,152	0.00	1.55	−2.13	−0.04	1.94

N = 2,152 bank-year observations, 97 distinct BUK banks, 23 years. Bank-type composition: 4 BUMN, 20 BPD (regional state), 38 private domestic, 31 foreign-owned, 4 unknown.

7. Baseline Local Projections — first headline results

Headline LP findings — A one-standard-deviation increase in the composite GFC factor (the Q1 reading) is associated, at the 3-year horizon, with +4.6 log points more credit growth (***), +3.0 log points more asset growth (***), −25 bp lower NPL (***), and −80 bp lower CAR (**) in the average BUK bank — controlling for bank fixed effects, lagged log assets, DPK share, and credit share. The composite GFC factor loads positively on VIX and credit spreads but negatively on the Fed Funds rate, so the medium-term association captures both the risk-off shock and the subsequent global monetary accommodation cycle that fuels EM credit booms.

Table 6 — Baseline LP, bank FE, SE clustered by bank

Outcome	h = 1	h = 2	h = 3	h = 5
Δ log credit	−0.003 (0.005)	+0.016* (0.009)	+0.046*** (0.012)	+0.049*** (0.012)
Δ log assets	−0.001 (0.003)	+0.008 (0.005)	+0.030*** (0.007)	+0.030*** (0.009)
Δ NPL gross (pp)	+0.055 (0.040)	−0.034 (0.064)	−0.254** (0.126)	−0.307*** (0.117)
Δ CAR (pp)	+0.162 (0.217)	+0.120 (0.381)	−0.795** (0.333)	−0.498 (0.371)

Coefficient on the composite GFC factor (Q1 reading). Bank fixed effects, SE clustered by bank. *p<0.10, **p<0.05, ***p<0.01. Controls: lag log assets, lag DPK share, lag credit share. N ≈ 2,055 bank-years per regression (h=1 sample); fewer at longer horizons.

Figure 6. Heterogeneous GFC transmission, h = 2 LP

Coefficient (with 95 % CI bars) on the GFC × group interaction in a 2-year horizon LP with bank + year fixed effects. With year FE absorbing the main effect, each row is the additional response of the listed group relative to the omitted base (bank_type = foreign; size_quintile = 1; fx_quartile = 1). Panel (a) shows the credit-growth response, panel (b) the NPL-gross response.

File: figure_6_forest_plot.png · 320 DPI · ≈ 1,414 bank-year obs per regression

Heterogeneity highlights —

State (BUMN) banks see NPL rise more during stress (+0.78 pp**) but CAR also rises (+1.65 pp**) — consistent with government recapitalisation support.
High-FX-exposure banks (Q3, Q4) deleverage assets more strongly during global stress (Δ log assets −0.038**, −0.046**).
Size quintiles are not statistically distinguishable — the cross-section is dominated by ownership type, not size.

Table 8 — Episode event-study, h = 2 LP

Episode	Δ log credit	Δ log assets	Δ NPL (pp)	Δ CAR (pp)
GFC 2008–09	−0.115** (0.056)	−0.052 (0.032)	+1.217** (0.514)	+2.123 (1.774)
Taper tantrum 2013	+0.111*** (0.034)	+0.106*** (0.029)	—	—
CNY / commodity 2015–16	+0.065 (0.067)	+0.037 (0.047)	—	—
EM contagion 2018	−0.033 (0.032)	−0.040* (0.024)	—	—
COVID 2020	−0.156*** (0.052)	+0.027 (0.049)	—	—
Fed hike 2022–23	+0.090* (0.047)	+0.004 (0.028)	−1.290** (0.654)	−3.471 (2.282)

Episode dummy = 1 if year overlaps the pre-specified window. Bank fixed effects, SE clustered by bank. Several CAR / NPL cells absorbed by FE due to the 2016 + 2017–2020 NPL sample structure (signalled by —). Cells in italics are stylised facts confirmed by both descriptive Figure 4 and the regression. *p<0.10, **p<0.05, ***p<0.01.

What the episode study reveals — Two episodes show the canonical risk-off signature: GFC 2008–09 (credit −11.5 log points**, NPL +1.22 pp**) and COVID 2020 (credit −15.6 log points***). The post-Fed-hike 2022–23 episode shows credit growth holding up (+9.0**) and NPL falling (−1.29 pp**), suggesting that Indonesian banks navigated the most recent rate-cycle materially differently from earlier shocks — possibly reflecting the post-2016 macroprudential framework. Taper 2013 + CNY 2015–16 + EM 2018 are quieter at h = 2, consistent with the descriptive credit-growth panel in Figure 4(c).

8. Robustness — does the baseline survive?

The baseline GFC coefficient on Δ log credit at h = 3 (β = +0.046***) is stress-tested with six robustness checks. The headline finding survives all six, with one substantive revision to the narrative on the post-2016 period.

Figure 7. Robustness of the baseline GFC transmission coefficient

Each row plots the coefficient (with 95 % CI bars) on the composite GFC factor in the focal regression of Δ log credit at h = 3. Brick dotted line marks the baseline point estimate.

File: figure_7_forest_robustness.png · 320 DPI

Table 9 — Robustness of baseline coefficient (Δ log credit, h = 3)

Specification	Coefficient (SE)	N obs
(0) Baseline	+0.046* (0.012)**	1,757
R1. + Domestic macro controls (CPI, IDR/USD, BI rate)	+0.047*** (0.015)	1,757
R2a. GFC pre-2016 (main)	+0.038*** (0.012)	1,757
R2b. GFC × post-2016	+0.081 (0.033)**	1,757
R3. Two-way clustered SE (bank + year)	+0.046*** (0.016)	1,757
R4a. GFC factor — annual mean	+0.058*** (0.014)	1,757
R4b. VIX (Q1) only	+0.002 (0.002)	1,757
R4c. BAA spread (Q1) only	+0.047** (0.021)	1,757
R5a. Subsample 2002–2014	+0.033*** (0.012)	1,145
R5b. Subsample 2017–2025	+0.130* (0.027)**	521
R6. Drop \|Δ credit\| > p99	+0.040*** (0.008)	1,740

All include bank fixed effects + baseline lag controls. *p<0.10, **p<0.05, ***p<0.01.

Key robustness takeaways

R1 (domestic controls): β virtually unchanged (+0.046 → +0.047***), so GFC is not proxying for BI rate, IDR/USD, or CPI inflation.
R2 + R5 (the surprise): The post-2016 GFC sensitivity is materially larger, not smaller — interaction +0.081**, subsample β almost 4× the pre-2014 magnitude (+0.130 vs +0.033). Indonesian banks became more coupled to the global financial cycle after the macroprudential framework activation, not less.
R4b vs R4c: The composite factor's transmission is driven by credit-spread movements (BAA spread), not volatility (VIX). VIX alone has zero predictive power for Indonesian bank credit. This refines the channel interpretation.
R6 (outliers): Trimming the top 1 % of |Δ log credit| (M&A spikes) leaves β nearly identical (+0.040***). Result is not driven by a handful of merger-fueled jumps.

Narrative revision — asymmetric macroprudential effectiveness

The post-2016 macroprudential framework appears to work asymmetrically:

Stress side: the Fed-hike 2022–23 episode (Table 8) shows declining NPL and stable credit — macroprudential buffers do contain stress propagation.
Boom side: the same banks transmit global accommodation phases more strongly than pre-2016 — macroprudential does not contain credit booms during easy-money periods (R2b + R5b).

This is the classic asymmetric macroprudential result documented in cross-country panels (Cerutti et al. 2017, Akinci & Olmstead-Rumsey 2018) — buffers limit downside but not upside transmission. Indonesia fits the pattern.

9. Caveats — what reviewers will ask

Frequency is annual, not quarterly

OJK CFS public release is one snapshot per year (April / Q1). The paper is therefore a bank-year panel. Most GFC episodes (GFC 2008, COVID, Fed hike) span multiple quarters and remain visible in annual data, but Taper tantrum (May–Sep 2013) shows up only as a single point in 2013-04 / 2014-04 averages — power against this episode will be modest.

2016 gap is structural (OJK source)

OJK CFS public endpoint returns no_report for fiscal year 2016 across all BUK banks and all report types — verified through two scrape attempts (May 2026, May 2026 re-test). The gap reflects the POJK 14/2017 reporting-format transition rather than any pipeline issue. The paper treats 2016 as a one-year structural gap absorbed into year fixed effects; the 2002–2015 ∪ 2017–2025 windows together cover 23 years of bank-level data including all six risk-off episodes.

FX exposure is gross, not net

The proxy sums off-balance-sheet FX-tagged commitments and normalises by total assets. It measures exposure intensity, not unhedged open FX position. Foreign-branch ratios above 500 % were dropped as suspect (likely double-counted notionals).

Identification — single small open economy

GFC shocks are exogenous to any individual Indonesian bank but not to Indonesia's macroeconomy. The paper controls for domestic macro (BI rate, IDR/USD, CPI) and identifies via bank-level heterogeneity in the GFC response, not via the existence of the GFC itself.

10. Next steps

✅ Master estimation panel — 2,152 bank-year obs, 97 banks (Stream E). Done.
✅ Baseline LP + heterogeneity + episode study — Tables 6, 7, 8 + Figure 6 (Stream F). Done — see results above.
Write the paper — manuscript HTML in the same style as the MBG paper preview, with sections (i) introduction, (ii) data, (iii) methodology, (iv) main results, (v) heterogeneity, (vi) episode study, (vii) discussion, (viii) policy implications.
Domestic macro controls (robustness) — add BI rate, IDR/USD, CPI as time-varying covariates to check that the GFC factor isn't proxying for domestic policy.
Macroprudential resilience formal test — interact GFC × post-2016 dummy explicitly to confirm the regime shift hinted at by the Fed-hike-2022 episode coefficient.
Two-way clustered SE robustness — currently bank-only; add year clustering for the bigger samples.