Signup buckets — who they are, when to charge them, and how to chase them

The twelve buckets

Measured on FIT-CARD: 21,089 agreements that have a known card class and at least 90 days of history behind them. Projected 12-month cash is not measured on these agreements — they are not old enough. It is the bucket’s own first-charge rate applied to the two values measured on FIT-LONG, where an agreement whose first charge lands banks $1,367 over twelve months and one whose first charge fails banks $418. That single gap of $949 an agreement is what all of this is actually about.

Download the bucket table — buckets.csv · filter it on the bucket column

The cross is the point — neither dimension says this on its own

A bigger deposit lifts every class. But an early-access customer who put down 30%+ still fails more often than a traditional-debit customer who put down under 20%. The card is the floor and the deposit is the lever, and you need both to place someone. On its own, issuer class spreads first-charge success by 57.1pp and deposit band by 28.0pp; crossed, the best cell and the worst are 70.6pp apart.

What to do with each one

Every bucket gets the same ladder shape, because the branch that matters is the decline reason, not the customer — see the ladder tab. What changes per bucket is the charge day, how hard it is worth working, and how much of the failure pile a collector should ever be asked to touch. The chased by a collector column counts what the collections team did — nothing in this business charges a failed card again on its own.

Bucket 12 is a join gap, not a kind of customer

9,110 fitted agreements and 149 of the last 500 sit here because the CRM booking could not be matched to a unique schedule — not because anything is unknown about the person. Their card class is known, so they can still be routed on it today, and they move into the grid the moment the deposit is joined:

Separately, 11,162 agreements are outside the bucket model altogether — they opened before the BIN feed starts (2024-09), so no card class exists for them and none ever will. They are in the master CSV marked other, and they are excluded from every rate on this page.

Buckets that are too small to act on

Benefits card (Direct Express) (n=86). Treat these as a routing rule with a reason behind it, not as a measured rate. The benefits-card bucket in particular is worth keeping separate even at n=86: Direct Express pays on the 1st and 3rd and on a Wednesday set by birth date, so a Friday payroll rule is actively wrong for it.

Every issuer on the book

Download all 971 issuers — issuers.csv · filter it on the issuer_class column

Download every one of the 32,607 agreements with every column on this page — agreements.csv · filter it on the set column

The weights are fitted, not assumed

Neil asked to weight the metrics so we can decide where people go. Six predictors each have a real univariate spread, but they overlap — stacking them as additive weights would count the same signal several times. So each one is put in a logistic regression against all the others, and the question is only which ones still move the needle once the rest are in. The outcome modelled is whether the first instalment lands.

A — what the CRM booking alone knows n=15,955 · McFadden R² 0.051

B — plus the card’s issuer class n=12,052 · McFadden R² 0.143

C — plus how the deposit was paid n=7,186 · McFadden R² 0.063

D — everything at once (this is the one that decides) n=7,186 · McFadden R² 0.155

D + $500 — the same model on the $500+ product only n=6,360 · McFadden R² 0.151

E — with the sub-$500 flag put in explicitly n=12,052 · McFadden R² 0.142

That does not make the charge day useless — it makes it a different kind of lever. Day of week does not tell you which customer will pay; it tells you which day a charge lands. It is a scheduling decision for people you have already bucketed, not a bucketing dimension. That is why it has its own tab and no place in the grid.

Adding the issuer class takes the model from R² 0.051 to 0.143 — roughly 2.8× the explanatory power of everything the CRM booking knows on its own. If only one field could be captured for every new customer, it is the card’s issuing bank.

What separating the sub-$500 sales actually moved

Model D is fitted across both products; D + $500 is the identical model on the $500+ product alone. If a weight is stable between them, the blend was not hiding anything. If it changes status, the blended figure was an average of two different things and should not be quoted.

How much the predictors overlap

So the weights are not taken on trust. Every pair is weakly correlated at worst — the strongest is deposit % against agreement length at 0.28 — which is why the issuer × deposit cross tells you something that neither column does alone, and why cash-in-the-deposit still earns a place even though a cash deposit is also a slightly smaller one.

Download every model, every coefficient — weights.csv · filter it on the model column

How the deposit was paid

Cash in the deposit survives the full model at ×? odds. It weakens once deposit size is in the model — a cash deposit is a smaller deposit — but it does not vanish, so it is a real second signal and not a restatement of the first.

Measured on FIT-LONG so the 12-month column is real. The exports that carry this column stop at 2026-04, so it can be measured on history but cannot be read for anyone signing up today — which is why it is a flag on the bucket rather than an axis of the grid. If Neil wants it live, the Daily Sales export needs to keep coming.

How long the agreement is

Shorter agreements do worse, not better. Rows under 25 agreements are hidden. It survives the full model, so it is not simply a restatement of instalment size — and note that instalment size itself does not survive.

Download every one of the 32,607 agreements with every column on this page — agreements.csv · filter it on the set column

Sub-$500 sales are a different product

Across the whole CRM ledger (60,784 agreements, no Square join needed)

7,655 of 60,745 agreements (12.6%) are sub-$500 on one side or the other. A small slice of the money and a real slice of the count — exactly the shape that distorts a blended rate.

What the segment does to the outcome (FIT-LONG, a full 12 months each)

And on the card-era set the buckets are fitted on

What the small sales need is not a different ladder — a soft decline is a soft decline at any size — but a different economic threshold. Rung 4 puts a collector on the phone, and that call costs the same whether the balance is $260 or $2,400. On a sub-$500 financed balance it can cost more than it recovers. Run rungs 0–3, which are all automated, on every segment; reserve the human rung for the $500+ product.

Kids vs adult campaigns

A kids campaign here is a marketing code containing KIDS — both the Blue Rooms {CITY}KIDSX family and the bare {CITY}KIDS codes Neil confirmed on 29 Jun 2026 carry real volume. The supplier is a different question and is answered by suppliers.classify(), the canonical implementation, not by a regex written for this page. Note that KIDS is a campaign type and says nothing about whether a customer has children.

By lead supplier ($500+ only, so sale size is not doing the talking)

Afterpay and the deposit — and what we cannot see

The shape matters more than the count. The Deposits locations are the $50 booking fee, and that is where almost all the BNPL sits — someone spreading a booking fee, not a deposit. The real money taken on the day (the Less Card component, hundreds of dollars) goes through Studio Sales, and only 62 BNPL attempts ever landed there.

Compared on the card-era set at 90 days — Square only started reporting the tender in 2024-09, which is after the 12-month window closes.

Download sale size, campaign type and supplier, every combination — segments.csv · filter it on the size_segment column

Download every one of the 32,607 agreements with every column on this page — agreements.csv · filter it on the set column

The day is a scheduling lever, not a bucket

The regression says the day of the week and the day of the month drop out once you know the card and the deal — they do not tell you which customer will pay. What they do tell you is which day converts, for a customer you have already got. That is worth real money at 196,606 charges a year, and it costs nothing to change.

Day of the week

Where in the pay cycle

The start of the month is the best window for every class, and the end of it is the worst for every class. That is not a subtle finding and it is being ignored: the busiest charge days on the book are the 15th, 20th, 28th and 31st, and three of those four sit in the worst windows in this table.

Every day of the month

Download day of month, split by card class — day_of_month.csv · filter it on the issuer_class column

Download day of week, split by card class — day_of_week.csv · filter it on the issuer_class column

The time of day — we cannot answer this, and here is why

91.6% of all recurring charges fire inside two hours (14:00 UTC and 15:00 UTC), which is the automated batch. They convert 44.3%. Everything charged outside that window converts 92.6%.

One thing to fix regardless. The batch is pinned to a fixed UTC time per city, not to a local one. When US clocks go back on 1 November it will silently start firing an hour earlier in local time — 09:00 instead of 10:00 — without anyone changing anything. Whatever the right time turns out to be, it should be pinned to the city’s local time, not to UTC.

Download every one of the 32,607 agreements with every column on this page — agreements.csv · filter it on the set column

The chase is the mechanism, so here it is measured

First, a number I nearly put on this page and should not have

The unselected number: who never gets chased at all

What the collectors did, by card class

Failed opening instalments from 2025-09 onward with a full 90 days behind them: 2,408 agreements.

The same table across all twelve buckets

Thin once the failures are split twelve ways — rows under 100 are marked. Use the card-class table above for anything you intend to act on.

Download the chase against every failed opening instalment — chase.csv · filter it on the bucket column

Does more chasing recover more? No — and the reason matters

The one comparison that is worth something

Within the chased population only — everyone here got at least one text, so the “did we bother” selection is held constant. What differs is whether they wrote back.

A reply is worth roughly 20 to 25 points of recovery in every class. That is still selected — people who reply are people willing to engage — but it is the most informative cut available, and it points somewhere useful: the thing to optimise is the reply rate, not the text count. Median time to a first text is already 3.0 days, so there is little left to win on speed. Getting more of the silent half to answer is where the room is. And per project_debt_reply_cohort_mix, never quote a blended reply rate: first-month conversations answer at 51-57% and the re-chased pile at 12-13%, so a single number hides which pile you are looking at.

Calls: we cannot attribute a single one, and here is exactly why

Download every one of the 32,607 agreements with every column on this page — agreements.csv · filter it on the set column

First, the thing that changes what every number below means

So every “next attempt” in the data is a human action: a collector taking a card over the phone, or a customer paying a reminder link we sent them. That single fact decides how this whole tab has to be read, and it is why the ladder Neil is designing is a new capability being specified from zero, not a tuning exercise on an existing one.

The one number that survives — and it is the important one

Step one: read the reason. Most failures must never enter a ladder at all

The families come from the REASONS table report_square_declines.js already exports, so this page and /square-declines cannot drift apart. Structural means the card details we hold cannot work — expired, invalid account, bad number. Hard means the issuer has told us to stop asking. Neither is fixed by a smaller number, and putting the same dead card through again buys nothing but decline fees. The high “the collector got paid” figure on the structural row is not the old card working — it is the customer having put a new card on file, which is exactly the play those two families need.

How long after the failure the collector reached them, and how often it worked

The day 15-30 row is mostly not a collector at all — it is next month’s scheduled instalment arriving. And the day 0-1 row (n=9,642) is where the same-day keyboard work sits: 4,819 of all second attempts land within five minutes of the failure.

The amount — the same trap, which we already knew about

The ladder, specified from zero

Sub-$500 sales get rungs 0–3 and stop. Rung 4 costs a collector’s time, which is the same whether the balance is $260 or $2,400 — on a sub-$500 financed balance that call can cost more than it recovers. See the sale-size tab.

The trial that would actually tell us something

Because nothing automatic has ever charged a failed card again, the only way to learn what it would convert at is to run it. That is the main argument for the trial — not refining a number we already have, because we do not have one.

Randomise rung 2 on TRANSACTION_LIMIT and INSUFFICIENT_FUNDS only — the two codes where a smaller charge has a plausible mechanical reason to work and needs no phone call. Three arms: collectors only, exactly as today (the control), same amount charged again on day 2, and 35% of the amount on day 2. That design answers both open questions at once — does an automatic attempt add anything on top of the collectors, and does the amount matter once the human is out of it — and neither can be answered from the history, because in the history the collector is always there.

The control arm is the important half. Without it, any number the trial produces gets compared against the 90.3% in the table above, which is a rate from conversations and would make an automated rung look like a failure no matter how well it did.

Download every collector-contact band, by decline family — ladder.csv · filter it on the decline_family column

Download every one of the 32,607 agreements with every column on this page — agreements.csv · filter it on the set column

The last 500, scored

Download all 500 scored signups, every column — last500.csv · filter it on the bucket column

What we can and cannot see about them

The deposit gap is the one worth closing, and it is closable: it is a reporting join, not missing data. If the CRM booking reference were written onto the Square customer at signup, every one of these would land in a real bucket on day one instead of 55.8% of them.

You can bucket them on the day of the shoot — you already have the card

16,450 agreements had a studio card charge before their first instalment — the deposit taken at the till. That card’s issuing bank predicts the first instalment on its own, before a single recurring charge has run: an early-access neobank card at the studio goes on to land its first instalment 28.8% of the time against 70.6% for a credit card. Of the last 500, 344 have one. That means the bucket does not have to wait for the first instalment to fail — it can be set the moment they pay their deposit.

What has already happened to them

220 of the 500 have already had their opening instalment fail. 200 of those (90.9%) are soft declines — the ladder applies to every one of them. Only 59 (26.8%) were chased by a collector within ten days — and every one of those was a person, because nothing charges a failed card again on its own. That is the argument on one line: the ladder does not need to be cleverer than the one Neil sketched, it needs to exist.

Download every one of the 32,607 agreements with every column on this page — agreements.csv · filter it on the set column