U.S. Bank Deposit Stickiness & ALM Mismatch Pipeline

smkwray.github.io/bankalm

Site

The bankALM static site provides an interactive atlas of bank fragility scores, league tables, and bank detail pages — all generated from the pipeline's enriched panel.

Live at smkwray.github.io/bankalm | served from site/ via GitHub Pages.

• Homepage — hero metrics, risk index cards, backtest comparison
• League Table — sortable, filterable, paginated rankings for each index (Run Risk, ALM Mismatch, Composite Fragility)
• Bank Detail — per-bank score breakdown, profile metadata, methodology notes

To run locally:

cd site && python3 -m http.server 8000

About

bankALM is an independent research project that screens U.S. bank fragility using only free public regulatory data. The live atlas is at smkwray.github.io/bankalm.

It is not affiliated with or endorsed by the FDIC, FFIEC, Federal Reserve, or any financial institution. Scores are computed from public filings and should not be interpreted as supervisory ratings or investment advice.

---

A reproducible pipeline using only free public data to estimate, at the U.S. bank-quarter level:

1. Deposit stickiness / run risk — transparent score + experimental supervised overlay
2. Effective maturity scenarios — baseline, adverse, severe deposit-life scenario proxies
3. Asset-liability mismatch — 13 structural ratios (pre-hedge proxy)
4. FFIEC repricing buckets — loan/deposit/borrowing maturity + repricing gaps + duration-gap-lite
5. Treasury buffer extensions — shock scenarios, encumbrance-adjusted coverage
6. Interpretable indices — peer-group normalized, per-row versioned

Quickstart

python3 -m venv ~/venvs/bankalm
source ~/venvs/bankalm/bin/activate
pip install -e .
pip install pytest

source ./.env
make universe    # download all data + build pipeline + generate reports
make test        # run the test suite
make status      # check data freshness, coverage, integrity

Pipeline Scale

Counts depend on the current processed build. The default universe spans quarterly FDIC bank filings from 2007 forward and persists intermediates in Parquet.

Failure Cohort Backtest

The validation layer now uses quarter-aligned bank-quarter cohorts rather than one-row-per-bank comparisons. For each report quarter, the pipeline labels whether a bank fails within 1Q, 2Q, or 4Q, tracks whether that horizon is observable, and produces reproducible metrics tables by horizon and episode slice.

Data Sources

Source	What it provides
FDIC BankFind financials	Deposits, assets, capital, securities, borrowings
FDIC BankFind institutions	Identifiers, charter, location (current snapshot metadata)
FDIC Summary of Deposits	Branch geography, deposit concentration
FDIC failures list	Historical bank failure dates and institutions
FFIEC CDR Call Reports	Loan/deposit/borrowing maturity buckets, repricing data
FFIEC NIC	Holding-company structure and parent chains
SEC EDGAR	Company tickers, 10-Q/10-K filing parser for uninsured deposits

Validation

The repo ships:

• quarter-aligned failure cohort labels and metrics
• consistency checks on the bank-quarter panel
• historical stress validation scripts and reporting outputs

These artifacts support exploratory public-data screening, not a claim that bankALM is a decision-grade failure predictor.

Package Structure

src/bankfragility/
  downloads/         FDIC API, Treasury yields
  staging/           Bank panel (5-way join), FFIEC repricing extraction
  features/          Deposit stickiness, ALM mismatch, Treasury extensions
  models/            Index construction, supervised stickiness overlay
  validation/        Failure backtest, 11 consistency checks
  entity/            SEC CIK mapping, 10-Q/10-K parser, NIC structure
  reporting/         League tables, drill-downs, peer-group summaries

Make Targets

Target	Description
make universe	Full download + build + reports
make smoke	Quick rebuild from existing data (no downloads)
make test	Run the test suite
make status	Check data freshness, coverage, integrity
make validate	Run 2023 SVB stress validation
make download-all	Download default raw data
make download-treasury	Download Treasury yield history
make build-all	Build panel through indices + supervised
make backtest-failures	Run quarter-aligned failure cohort validation
make reports	Generate league tables and reports
make clean	Remove processed outputs (keeps raw)

Outputs

The pipeline produces per bank-quarter:

• Raw features: deposit composition, ALM ratios, Treasury coverage, FFIEC repricing buckets
• Scenario estimates: deposit WAL under baseline/adverse/severe, stable-equivalent amounts
• Indices: run_risk_index, deposit_stickiness_index, alm_mismatch_index, treasury_buffer_index, funding_fragility_index
• Supervised overlay: experimental SUPERVISED_OUTFLOW_SCORE plus next-quarter observability and label fields
• Failure backtest: FAIL_WITHIN_1Q, FAIL_WITHIN_2Q, FAIL_WITHIN_4Q labels plus cohort metrics tables
• Split publishable panels: data/processed/universe_core_panel.parquet for the historically consistent full-history core output data/processed/universe_enriched_panel.parquet for the recent-history advanced-module output
• Mart + site exports: integrated data/processed/universe_publishable_mart.parquet plus generated site/data/manifest.json and site/data/league.json
• Reports: bank drill-downs, quarter league tables, peer-group summaries, scenario comparisons
• Macro context report: data/reports/universe/treasury_regime_summary.parquet with one row per quarter of the observed Treasury backdrop

Configuration

File	Purpose
config/peer_groups.yaml	Asset-size peer group thresholds (in FDIC thousands)
config/index_weights.yaml	Component weights for composite indices
config/stickiness_scenarios.yaml	Deposit life parameters by scenario
config/ffiec_repricing_map.yaml	MDRM code → standard horizon mapping
config/sec_overrides.yaml	Manual CERT → SEC CIK/ticker overrides
config/field_definitions.yaml	Field definitions and known changes

Limitations

• No internal deposit behavior or management assumptions
• Annual SOD geography, not daily funding flows
• ALM model is pre-hedge (derivative overlay available via FFIEC RC-L)
• Deposit-life outputs are scenario assumptions, not empirical decay estimates
• ALM outputs are structural public-data proxies, not a bank's internal hedge-adjusted model
• Holding-company disclosures and institutions metadata are not historical as-of mappings
• Supervised overlay is backward-looking and experimental
• Treasury yield history is integrated into the recent-history enriched panel, not the full-history core panel

The transparent score is the primary product. The supervised overlay is secondary. bankALM is a transparent public-data fragility screen and scenario engine, not a decision-grade probability model.

Published Artifacts

Use the split outputs intentionally:

• universe_core_panel.parquet: full-history panel for historically consistent use. Excludes reliance on the recent-only advanced-module surface.
• universe_enriched_panel.parquet: recent-history panel beginning when FFIEC/derivative enrichment is available.
• universe_publishable_mart.parquet: integrated internal mart used to derive the split panels and site exports.
• universe_failure_backtest_metrics.csv: quarter-aligned validation metrics by score, horizon, and episode slice.

The static site is generated from the recent-history enriched panel and includes manifest metadata describing both published panel variants.