NCUA credit union data

ncua-data-analysis

Clean, documented, quarterly credit union data from NCUA call reports, 2018 to today: lending, deposit mix, earnings, staffing and efficiency.

The MCP server below also covers FDIC-insured banks: quarterly income, efficiency, ROA and peer benchmarking across both sectors.

NCUA publishes every federally insured credit union's quarterly 5300 call report as free bulk files. They are hard to use: 3,300+ account columns split across 17 wide tables, form changes that move accounts around, year-to-date income, and a data dictionary written as form instructions. This project turns that into four tidy tables you can query with DuckDB, pandas or anything that reads Parquet.

Status: v0. 115 curated fields and 52 computed metrics, 34 quarters (2018 Q1 to 2026 Q2), checked against NCUA's own published totals. See what is not done yet.

Tables

Join on quarter + cu_number. Dollar fields are dollars. Fields ending in _ytd are year to date and reset each January (the Q4 value is the full year); metrics annualizes them.

-- Which large credit unions grew auto lending fastest last year?
SELECT d.name, d.state, m.auto_loan_growth_yoy, f.loans_new_vehicle + f.loans_used_vehicle AS auto_loans
FROM fact_call_report_curated f
JOIN dim_credit_union d USING (quarter, cu_number)
JOIN metrics m USING (quarter, cu_number)
WHERE f.quarter = '2026-06' AND d.is_federally_insured AND d.peer_group = 6
ORDER BY m.auto_loan_growth_yoy DESC LIMIT 10;

Run it

pip install -e .
ncua-data all          # download 34 quarters (~270 MB), build tables, reconcile
python -m unittest discover -s tests

Output lands in data/out/ as Parquet. Raw ZIPs are never committed. Links are scraped from NCUA's quarterly data page.

Numbers you can trust

For June 2026 and December 2025 the totals reconcile to the figures in NCUA's Quarterly Credit Union Data Summary: credit union count, members, loans by type, shares, net worth ratio, delinquency, income and expense. 60 of 67 checks match across six year-ends (2021 to June 2026). The 7 that do not are historical deposit lines that differ by under $0.5B (under 0.1%), for example Dec 2025 money market $367.5B versus $368.0B published. All June 2026 lines match. The likely cause is restated history in NCUA's table, which has not been confirmed. Full table: docs/RECONCILIATION.md.

Earnings ratios were audited against all 32 quarterly summaries from 2018-Q3 to 2026-Q2. Pooled ROA, net interest margin and net charge-off ratio, and the median yield on loans, cost of funds, margin and ROA, match NCUA's published figures once the denominator is the average of prior-December and current balances (the *_ncua_ytd fields). Year-to-date annualized ratios step each January, in NCUA's own numbers too; the *_quarterly fields are single-quarter versions derived here, not published by NCUA. Details and the few mismatches: docs/RATIO_AUDIT.md.

Things to know before using it:

• Filter to is_federally_insured. NCUA's raw files include about 85 state-chartered credit unions it does not insure. Its published totals leave them out.
• The form changed in 2022 and 2023. NCUA redesigned the call report and adopted CECL, so some accounts disappear and new ones appear. Fields that span the change map both account codes (for example the allowance is Acct_719 before CECL and Acct_AS0048 after). Fields that only exist after the change say so in the dictionary.
• Net worth ratio. NCUA's published ratio excludes the CECL transition provision from 2023 on. This dataset carries that provision (cecl_transition_provision) and metrics.net_worth_ratio_ex_cecl matches NCUA.
• Employees are estimated. employees_fte_estimate is full-time plus half of part-time. It is not an NCUA definition and is not reconciled to a published figure.
• Net charge-off ratio, ROA and NIM use NCUA's own average balances, which are not public. Metrics here use a four-quarter average and land within a few basis points of NCUA's published values, not on them.
• Mergers. Credit union counts fell from 5,375 to 4,214 since 2018. A merged credit union's history stays under its old charter number, so per-institution growth across a merger is not meaningful.
• Reports are self-reported and occasionally reposted as "Revised". The ZIPs are used as NCUA publishes them today.

Not done yet

• Commercial-loan delinquency by type (NCUA moved these to new account codes in 2022).
• More fields. The target is 150 to 200; 115 are in and verified.
• Years before 2018 (the download links use two other naming patterns).
• A static analytics site and natural-language querying. The dictionary is built to be the semantic layer for that.

MCP server (v1)

A local MCP server lets an AI assistant query this dataset in plain language. It runs over stdio, reads the release Parquet files with DuckDB (downloaded once to ~/.cache/ncua-data-analysis), and builds its tool descriptions from the dictionary table.

{
  "mcpServers": {
    "ncua-data": {
      "command": "uvx",
      "args": ["--from", "ncua-data-analysis", "ncua-data-mcp"]
    }
  }
}

The package is on PyPI: https://pypi.org/project/ncua-data-analysis/. To run the latest development version instead, use "args": ["--from", "git+https://github.com/bnovarini/ncua-data-analysis", "ncua-data-mcp"].

Hosted, nothing to install: https://ncua-data-analysis.fly.dev/mcp (streamable HTTP, read-only, rate limited to 60 requests a minute per client). Add it as a remote MCP server in any client that supports one, for example {"mcpServers": {"ncua-data": {"url": "https://ncua-data-analysis.fly.dev/mcp"}}}.

Claude Code: claude mcp add --transport http ncua-data https://ncua-data-analysis.fly.dev/mcp

Hosted usage logging

The hosted server records request timestamp, event type, tool name, client name/version (when supplied), request duration and outcome status in a persistent SQLite log. It does not store query arguments, results, IP addresses or User-Agent headers. Unknown tool names are recorded as (unknown) and return JSON-RPC -32601 (method not found), separate from errors in supported tools.

Deploy from this repository

From the repository root, run fly deploy --config deploy/fly.toml. The Docker build uses the repository root as its context and bakes the private quarterly and bank panels (cu-quarterly/, bank-quarterly/) into the image, so a local build without those directories fails by design. The usagelog volume mounts at /data and holds the usage log and cached dataset; preserve it during updates. deploy/usage-start.py starts the HTTP server with logging enabled. No credentials, usage database or dataset files belong in this repository.

The server covers both sectors, 2018-03 to 2026-06. Credit union tools: list_fields, find_credit_union, credit_union_profile, metric_series (one credit union or an aggregate across all), peer_compare (by asset group, state or charter), query_metrics (filters, ordering and limits; no raw SQL), and cu_quarter_income (single-quarter income with certified merger-event labels). Bank tools: find_bank (lookup by name), bank_quarter_income (official FDIC quarterly income with naive-difference flags and Quarterly Banking Profile reconciliation labels), bank_metric_series (efficiency, ROA, margin and income mix for one bank or aggregates, by state or asset band), and bank_peer_compare (a bank against its asset band, state or both). Bank ratios use the same formulas as the credit union metrics, so cross-sector comparisons are valid.

The quarterly-income and bank tools need panel files that are baked into the hosted image and are not in this repository; on a local install they answer with a pointer to the hosted endpoint. Set NCUA_DATA_DIR to use a folder of already-downloaded files. The hosted endpoint supports streamable HTTP; local clients can use stdio. The server is listed in the official MCP registry as io.github.bnovarini/ncua-data-analysis.

Download

Ready-made Parquet files are attached to the v0.1.1 release. GitHub caps release files at 25 MB, so fact_call_report_curated and metrics come in three parts by year (2018-2020, 2021-2023, 2024-2026) with identical columns:

import duckdb
base = "https://github.com/bnovarini/ncua-data-analysis/releases/download/v0.1.1/"
parts = [base + f"metrics_{y}.parquet" for y in ("2018_2020", "2021_2023", "2024_2026")]
duckdb.sql(f"SELECT quarter, count(*) FROM read_parquet({parts}) GROUP BY 1 ORDER BY 1").show()

Data source and license

Data: National Credit Union Administration, 5300 Call Report Quarterly Data. Bank data: FFIEC Central Data Repository call reports via the FDIC's BankFind API, reconciled against the FDIC Quarterly Banking Profile. NCUA does not state a license on the download page. As a US federal agency's work it is assumed to be public domain, but that assumption has not been confirmed. Credit NCUA when you use it.

Code: MIT, see LICENSE.

Bank employees and revenue per FTE

The hosted MCP includes bank_metric_series(field="employees_fte_reported"), list_bank_fields, and revenue_per_employee. The productivity tool compares banks, credit unions, or both in a headquarters state, with inputs, ranks, peer median, pagination and explicit exclusions.

Example: revenue_per_employee(state="NC", cu_number=18297) compares Coastal with the eligible NC-headquartered institutions in the same quarter. Use sector="credit_union" or sector="bank" for a single sector.

Net revenue means interest income minus interest expense plus noninterest income. Gross revenue means interest income plus noninterest income. The ratio annualizes year-to-date revenue and divides by quarter-end FTE. It is a run rate, not actual full-year revenue, net profit, or revenue divided by average workforce.

Credit-union FTE is our estimate: NCUA full-time employees (account 564A) plus half of part-time employees (564B). Bank FTE is reported by the bank through FDIC NUMEMP. These definitions differ, so combined-sector comparisons are indicative. Headquarters state does not mean all business occurs in that state. Outsourcing, business mix and mergers affect comparisons. Null/zero employees and known reporting-perimeter breaks are excluded and counted.

The hosted bank employee panel covers 2018Q1 through 2026Q2. Employee counts are not dollar amounts and are not scaled by 1,000. The local package does not ship the bank employee data; use the hosted endpoint or supply the separate panel when self-hosting.