Decline trends

Every recurring instalment charged since Jan 2026, sorted into buckets, month by month. Built 2026-09-19 07:36 UTC.

Recurring only. Studio-counter, deposit and office declines are excluded from every number on this page — a card failing at the till is a sale that did not go through, not arrears. Read the percentage, not the count. Two kinds of table here, and they answer different questions. Where the bucket is a property of the charge (plan position, payment history, card, issuer) the figure is the decline rate — declines divided by charges inside that same bucket and month, so a bucket getting bigger cannot read as a bucket getting worse. Where the bucket only exists on a decline (the bank's reason code) there is no such rate, so the figure is that reason's share of the month's declines and the label says so. Colour is the rate: under 45%, 45–55%, over 55%. The number is the count and the white figure beside it is the rate.

By what the bank actually said — share of the declines that month

The issuer's decline code folded into the action it implies — the same five buckets the GHL sequence branches on. MAXED is TRANSACTION_LIMIT, which is not a per-charge cap: in practice the card is at its limit, and we see it on $25 charges. MAXED Card is at its limit — needs a different card or a re-present after payday. RETRY No money in the account right now — worth re-presenting. NEW_CARD The card is dead — expired, closed or the number is wrong. CALL_BANK The issuer has blocked this kind of charge — the customer must call them. RE_KEY Card details entered wrong — re-key them. CHASE No reason given by the bank — treat as a normal chase.

By the issuer's raw decline code — share of the declines that month

The code underneath the action above, unfolded. Watch the split between the codes that mean no money (TRANSACTION_LIMIT, INSUFFICIENT_FUNDS) and the ones that mean the card is dead or blocked (INVALID_ACCOUNT, CARD_EXPIRED, PAN_FAILURE, CARDHOLDER_INSUFFICIENT_PERMISSIONS): the first is the economy, the second is something we can fix by getting a new card on file.

By where they are in the plan

An agreement is a run of recurring charges; a new one starts after 12 months of silence from that customer — the same test /first-payments and /instalment-recovery use. History is read from June 2023, so a charge in Jan 2026 has two and a half years of visibility behind it and cannot be mistaken for an opening.

The rise down this table is NOT proof that people get worse over time. It is heavily survivorship-biased: someone who pays cleanly finishes their plan and leaves the book, so the further into a plan you look, the more the remaining population is the people who were already struggling. The honest reading is "a charge deep in a plan is far more likely to fail than an opening charge" — useful for deciding where to put effort — not "plans decay". Separating those two needs a fixed cohort followed forward, which is what /instalment-recovery does.

By how well they had paid up to that point

The share of their own agreement they had already paid at the moment of that charge — never with hindsight, or every decline would look like it came from a bad payer. First charge is the opening instalment, which has no history behind it yet.

How much of this is BIN-backed

The card and issuer tables below can only see charges that carry a BIN, which comes from a separate pull with its own window. A month with low coverage here is not a month with no prepaid cards — it is a month the feed has not reached. Charges with no BIN are excluded from those two tables rather than bucketed as unknown.

By what kind of card it is

Prepaid wins over debit/credit where a card is both — a prepaid debit card behaves like prepaid, not like debit. Card type comes from Square per card where it has one, and falls back to the BIN reference.

By what kind of bank issued it

Read this one carefully. A BIN identifies the ISSUER, not the cardholder. "Early-access neobank" is the sponsor-bank cluster behind Cash App, Chime, Current, Varo, Dave, SoFi and MoneyLion — the population that can be paid wages early, not a proven list of people who are. "Benefits card" is Direct Express, the federal benefits debit card Comerica issues: that population is paid on the 1st and 3rd and on a Wednesday set by birth date, not on a payroll Friday, so it was split out of "traditional bank" on 4 Sep 2026 rather than being averaged in with Chase and Wells Fargo. Two banks changed side that day too: SoFi and Choice Financial Group (Current's sponsor) were being counted as traditional and are early-access.

Which bank the card is from — every issuer we charge 300+ times, worst first

The table above sorts banks into two boxes. This one names them. Same recurring charges, cut by the name printed on the card, 300+ charges since Jan 2026 to qualify — 36 banks, 59,815 charges — 71.4% of the 83,729 on the book, out of the 78,016 we can name an issuer for at all (the rest are spread thinly across 972 smaller banks and credit unions, all of them in the CSV). The figure is the share that went through, so it reads the same way round as the money beside it. “Charges” is attempts, not people — the People column is how many customers those charges belong to and the multiplier is how many times we hit each one. Every bank, including the ones under 300, is in the CSV.

The 8 worst banks on the whole book are all early-access, and the break is clean. The best early-access bank we charge at volume is SOFI BANK at 23.9%; the worst bank of any other kind is M&T BANK at 24.8% — 0.8 points clear, with nothing in between. That is not a gradient with neobanks at one end; it is two populations that barely overlap.

It is not the prepaid card — it is the bank

The obvious explanation for the table above is the plastic: prepaid cards have no overdraft, so of course they bounce. The data does not support it. Hold the bank constant and vary the card format, and the format barely moves the number; hold the format constant and vary the bank, and the number halves.

A normal, non-prepaid card from an early-access bank went through 19.3% of the time (8,221 charges). A prepaid card from a high-street bank or credit union went through 36.0% of the time (955 charges). The prepaid card at the ordinary bank beat the ordinary card at the neobank. And the single worst performer on the book, GREEN DOT BANK DBA BONNEVILLE BANK at 10.0%, is 0.0% prepaid, while SUTTON BANK is 97.9% prepaid and does better at 13.0%. The card format is not what predicts failure. Which bank the customer chose is.

What that is worth in money — and how many people it is

“6,000 charges” reads like 6,000 people. It is not: we charge the same person every month, and a failed charge gets re-presented, so one customer can be a dozen rows. Here are the same banks with the customer count beside the attempt count, and what we asked those customers for against what actually arrived.

SUTTON BANK and JPMORGAN CHASE BANK N.A. - DEBIT are almost the same size of group — 1,254 people against 1,500 — and we asked them for almost the same money, $1,147,597 against $1,396,440. We collected $140,813 from the first and $713,553 from the second. GREEN DOT BANK DBA BONNEVILLE BANK is smaller — 142 people, 994 attempts — but we asked for $198,736 and collected $12,106, 6.1% of it.

Does the day of the week explain it? Not the way you would expect

If the early-access story were really “these people are paid two days early”, their charges would do best in the middle of the week — Wednesday or Thursday — when the early money has landed and the Friday payroll has not. Every charge below is one that carries an identified card (78,016 of 83,729), grouped by the weekday it was actually charged in the studio's own timezone. The Benefits column is one issuer split seven ways — Comerica, a few hundred charges in total — so every cell under 200 charges is greyed and carries no swing figure. Read the Early-access and Traditional columns; the Benefits one is there for completeness.

Both groups peak on Friday and both trough on Sunday — the same shape, so the weekday effect is not something the neobanks have and the high-street banks do not. What differs is the size of it: early-access swings 1.8× from its best day to its worst, traditional only 1.1×. These customers are far more exposed to when in the week we ask.

This is not evidence of the “paid two days early” effect. If it were, the early-access line would peak mid-week. It peaks on Friday, the same day as everyone else — which says the money lands on the ordinary payroll cycle and these accounts simply run empty faster afterwards.

The whole book by calendar day

No BIN needed for these two, so they cover every one of the 83,729 recurring charges since Jan 2026 — the widest basis on this page, including the days and months the card feed has not reached. The question is whether the days we have chosen to charge on are the days that pay.

Best day of the month is the 1st at 49.5%; worst is the 14th at 38.8%. Our four busiest days — the 15th (20th worst of 30), the 31st (2nd worst of 30), the 28th (4th worst of 30), the 20th (7th worst of 30) — carry 26.4% of every charge we make. 3 of those four sit in the worse half of the month, and between them the four go through 41.2% of the time against 43.8% on every other day — 2.6 points worse, on 22,064 charges. We have concentrated the book onto days that pay below average.

The 29th is marked too few charges to read: 325 charges on the 29th, against 2,791 on a normal day. The 29th shows 70.8%, which is an artefact of how few agreements are set to that day and which months even have one — not a good day to charge on. Do not move anything onto it on this evidence.

Friday is the best day to ask at 46.8% and Sunday the worst at 40.0% — 6.8 points between them on the whole book. Saturday and Sunday together take 25.8% of every charge we make, and both sit below the weekday average. Nothing about a weekend charge is cheaper for us to attempt.

The 12 banks we are declined by most

Ranked by total declines over the whole period, so the columns stay in one order down the months. Issuers outside the top 12 are in the CSV, not dropped.

Downloads

Every bucket this page groups by is reachable from one of these three.

• buckets.csv — every month × bucket cell for all seven views (action, reason, instalment, history, card, issuer class, issuer), including the issuers below the top 12.
• issuers.csv — all 1,008 issuing banks, not just the 36 over 300 charges: class, charges, people, charges per person, success rate, prepaid share, asked and collected.
• calendar.csv — day of month, day of week, day of week × issuer class, and issuer class × card format.