Annualization

Income-statement fields are year to date and reset each January, so a March filing holds one quarter of expense and December holds four. Every flow is multiplied by a factor set by the cycle month before it is divided by a balance: March x4, June x2, September x4/3, December x1. Every annualized figure is labelled as such, and first-quarter figures are the noisiest, because one lumpy month is amplified fourfold.

Percentiles

A chip reading p94 means 94th percentile among credit unions in the same asset band in the same quarter. It is computed across the whole panel before the latest row per institution is selected, so the comparison is never against a different period. Bands are under $50M, $50-100M, $100-200M, $200-500M, $500M-1B, $1-10B and over $10B. A fixed threshold would flag every small institution on expense and none on scale, which is why nearly every judgement here is band-relative rather than absolute.

CECL coalesce

A credit union reports its allowance and provision under the legacy codes (719, 300) or the CECL codes (AS0048, IS0017), never both, depending on whether it has adopted the newer accounting standard. The pipeline sums each pair so every filer has one usable value without a branch in the code.

Table headers are mixed case. ACCT_010 in FS220 but Acct_997 in FS220A. Match case-insensitively or you will silently read zero for net worth, which is exactly what happened the first time the verification script was written.

Files are named .txt but are comma-separated. They open in Excel.

Ten years of quarters, one form redesign in the middle. The panel runs from 2016-03. NCUA redesigned the 5300 in March 2022 and renamed several lines; before that date five fields are read from their legacy codes: repossessed vehicles from 798A2, cash from 730A + 730B + 730C (the 730 subtotal is blank in the bulk files), first mortgages from 703, other real estate from 386, investments from 799I. One series does not splice cleanly: 799I was Total Investments including certificates at other institutions, while AS0013 after 2022 is investment securities only, so the investments line and the cash percentage step at 2022-03 for many small credit unions. Junior liens (RL0030), vehicle participations by type (SL0036), the CECL pair and the commercial split (400P, 718A5) do not exist before their introduction and read zero.

Codes retire. ACCT_618B and ACCT_618C went inactive. Notebook 02 checks the STATUS column on every code every cycle and reports anything retired, which is how a dead field gets caught instead of quietly becoming a column of zeros.

Most of the form is blank for most institutions. The derivatives table is 145 columns and not one credit union in the country reports a value in the interest-rate-swap column. That is correct data, not missing data.

The indirect field is self-reported. One $1.7B institution's indirect balance tripled in a year while its total loans shrank, which is a reclassification rather than lending. Four data-quality checks flag that pattern rather than silently dropping the row.

The delinquency maturity buckets do not sum to the total. Adding ACCT_021B, 022B and 023B gives $10.4B against the filed total ACCT_041B of $14.7B, an understatement of 29% system-wide, and the two disagree at 91% of the credit unions that report both. The maturity series covers only part of the book; the category detail lives across seven other tables. Use the filed aggregate, which is what this pipeline does and has always done.

The CECL fields are exclusive for the allowance and overlapping for the provision. No credit union reports both ACCT_719 and ACCT_AS0048. But 587 filings across eleven cycles report both ACCT_300 and ACCT_IS0017, because 300 is the loan-loss provision while IS0017 is the wider CECL total credit loss expense. This pipeline summed them until 3 September 2026, which double-counted the provision at 37 credit unions in the latest cycle, several by exactly 100%. It now coalesces, preferring the CECL figure, and asserts on every build that it is choosing rather than adding.

The funding percentages divide by shares and deposits, not by shares. The ten share components sum to ACCT_013 exactly, to the dollar in aggregate and at 94.6% of filers individually. They do not sum to ACCT_018, which is $24.5B larger because it also holds non-share deposits at 1,138 institutions. The funding ratios use 018 on purpose, since the question is what share of all member money is sticky, and they are labelled "of member funds" rather than "of shares" so the denominator is not misread. For the median institution the two are identical; at the most affected the gap reaches 24 points.

Vehicle charge-offs are gross. The 5300 carries no vehicle-specific recovery field, so veh_nco_pct is gross where total_nco_pct is net. The two are not directly comparable.