Bank Size, Growth, and Net Interest Margin

https://smkwray.github.io/banknim/

A scripted public-data pipeline studying the relationship between bank size, asset growth, and net interest margin (NIM) across 8,032 FDIC-insured commercial banks over 64 quarters (2010–2025).

---

What this project does

This project estimates how bank size is associated with net interest margins, and whether the relationship operates through being big (cross-sectional), becoming bigger (within-bank growth), franchise dilution, rate-cycle amplification, regulatory threshold crossing, or time-varying size effects. Every data source is free, the pipeline is scripted end to end, and full reruns depend on external-data availability plus the documented setup prerequisites.

Research question

What explains changes in bank net interest margin, and is the lower NIM of large banks mainly a fact about being big or becoming bigger?

---

Findings

Results are organized by hypothesis. All coefficient estimates come from panel models on 373,220 bank-quarter observations unless otherwise noted.

Headline caveat

• Raw NIM (no winsorization): size effect is insignificant (−0.01, p=0.87)
• The most reassuring evidence comes from first-difference (−0.11, p<0.001) and lagged-controls (−0.12, p<0.001) specifications
• The baseline winsorized result is informative, but it should not be read in isolation

H1: Between-bank size penalty

• Coefficient: −0.085 (p<0.001)
• Banks 10× larger run ~20 bp lower NIM in the cross-section
• OLS on 8,023 bank-level means, R² = 0.23

H2: Within-bank size penalty

• Coefficient: −0.090 (p<0.001)
• Bank fixed effects absorb all time-invariant bank characteristics
• Survives first-differencing (−0.11, p<0.001) and lagged controls (−0.12, p<0.001)

H3: Franchise dilution

• Branch-growth interaction: −0.58 (p<0.01)
• Growth without franchise deepening hurts margins
• Rapid branch expansion dilutes the deposit franchise

H4: Acquisitions

• Event study: NIM V-shape around deals
• Pre-event decline from +0.11 (t−8) to +0.02 (t−2) relative to t−1
• Post-event rebound peaking at t+2 (+0.12) before fading
• Acquirers may be buying when their own margins are under pressure

H5: Rate cycles

• Size×FedFunds: −0.012 (p<0.001)
• The size penalty on NIM worsens during rate-hiking cycles
• A steeper yield curve partially offsets (+0.015)

H6: Fee offset

• ROA rises with size: +0.205 (p<0.001)
• Noninterest margin: +0.31 (p<0.001)
• Large banks offset lower NIM with higher fee income — a different earnings model, not a worse one

NIM decomposition

• Interest income FE: +0.154 (p<0.001)
• Interest expense FE: +0.245 (p<0.001)
• The NIM penalty appears to come more from the funding-cost side than from lower asset yields

Equity, efficiency, and loan mix extensions

• Equity ratio FE: −0.0177 (p<0.001), consistent with larger banks running thinner capital buffers
• Efficiency pass-through: each 1 pp of NIM translates into about +0.514 pp of ROA after controls
• Loan mix FE: adding detailed CRE, residential, C&I, and consumer shares attenuates the size coefficient from −0.090 to −0.070, but does not eliminate it

Time-varying coefficients

• Rolling 20-quarter within-bank FE: size coefficient ranges from about −0.033 to −0.144
• Weakest around the mid-2010s, strongest in 2019–2024, and weaker again in the latest window
• The size penalty is persistent, but not constant over time

Distress extension (exploratory)

• Failure-only 4-quarter-ahead logit: higher NIM and larger size predict the event label, while higher ROA lowers it
• This is not a clean “low NIM predicts failure” result, so it should be treated as exploratory rather than headline evidence

Null results

• NIM volatility: size coefficient insignificant (p=0.19)
• Market power: local deposit HHI does not predict NIM (p=0.70)

Robustness flag

• The baseline result depends on trimming extreme NIM outliers
• First-difference (−0.11) and lagged-controls (−0.12) specifications are the most reassuring

---

Data sources

All data are from official government agencies. No proprietary, scraped, or paywalled data are used.

Source	Agency	Frequency	Coverage
BankFind Financials	FDIC	Quarterly	2010–2025
Summary of Deposits (SOD)	FDIC	Annual	2010–2025
Institution History	FDIC	Event-based	Merger/acquisition events
FRED Rates	Federal Reserve	Daily/Quarterly	Fed funds, prime, treasury yields

373,220 bank-quarter observations. 8,032 unique FDIC-insured commercial banks.

---

Data QC / sample construction

The review-facing sample flow is explicit rather than implicit:

• Raw FDIC financial rows loaded: 373,220
• Rows with required bank/date/assets parsed successfully: 373,220
• Rows after bank-quarter deduplication: 373,220
• Final core panel rows: 373,220 across 8,032 banks from 2010-03-31 through 2025-12-31

The current public build records zero duplicate bank-quarter keys in duplicate_keys.csv. Column-level missingness is exported separately in missingness_report.csv, and the site publishes the same staged flow as sample_selection.json.

---

Model inventory

Model	Description	Key variable	N
Within-bank FE	Bank fixed effects, size on NIM	Log assets	372,644
Between-bank means	OLS on bank-level averages	Avg log assets	8,023
Growth dynamics	Asset and deposit growth	Asset growth	364,621
Franchise dilution	Branch-growth interaction	Branch growth × size	283,636
Rate cycle	Size × fed funds, size × yield slope	Log assets × fed funds	372,644
Acquisition event study	Event-time dummies around deals	t−8 through t+7	23,837
Threshold crossing	Event study at $10B / $50B / $100B	Crossing dummies	2,217 / 619 / 426
NIM decomposition	Interest income and expense FE outcomes	Log assets	372,644
Rolling coefficients	Rolling 20-quarter within-bank FE	Log assets	45 windows

Robustness specifications: Baseline (winsorized), raw NIM (no winsorize), exclude <12 quarters, log deposits as size, first differences, lagged controls.

Extensions: NIM volatility, fee offset (ROA, interest expense, noninterest margin), market power (HHI), lagged controls, equity ratio, efficiency pass-through, loan mix, and exploratory distress prediction.

---

Repository structure

banknim/
  config/           Project settings and source inventory
  data/             Raw inputs and processed panels (gitignored)
  docs/             Static site (GitHub Pages)
  output/           Model results, figures, frontend JSON (gitignored)
  scripts/          Reproducible command-line entry points
  src/              Python package: data parsers, model estimation, utilities
  tests/            Automated test suite

Reproducing the results

# set up the environment
python3 -m venv "$HOME/venvs/banknim"
source "$HOME/venvs/banknim/bin/activate"
pip install -e .[test]
pip install -r requirements.txt
source .env

# download raw data
python3 -B scripts/00_fetch_metadata.py
python3 -B scripts/01_download_fdic_financials.py --start 2010-03-31 --end 2025-12-31
python3 -B scripts/02_download_fdic_sod.py --start-year 2010 --end-year 2025
python3 -B scripts/03_download_fdic_history.py --start-date 2010-01-01 --end-date 2025-12-31
python3 -B scripts/04_download_rates.py

# build panels
python3 -B scripts/06_build_core_panel.py
python3 -B scripts/07_build_sod_panel.py

# run models
python3 -B scripts/08_run_baseline_regressions.py
python3 -B scripts/09_event_study_acquirers.py
python3 -B scripts/10_run_franchise_dilution.py
python3 -B scripts/11_run_rate_cycle.py
python3 -B scripts/12_run_threshold_crossing.py
python3 -B scripts/13_run_robustness.py
python3 -B scripts/14_run_extensions.py
python3 -B scripts/16_run_nim_decomposition.py
python3 -B scripts/17_run_equity_as_outcome.py
python3 -B scripts/18_run_efficiency.py
python3 -B scripts/19_run_loan_composition.py
python3 -B scripts/20_run_failure_prediction.py
python3 -B scripts/21_run_rolling_coefficients.py

# export frontend data for the site
python3 -B scripts/15_export_frontend_data.py
cp output/frontend/*.json docs/data/

# run tests
python3 -B -m pytest -q

Or use the Make targets:

make setup
source .env
make all
make test

License

See LICENSE.