Churn-to-win-back targeting

Win-back targeting starts with customers who already left, not with the live book. Pull recently closed accounts from the portfolio, restrict the window to a period your campaign can still act on, and score that slice for re-acquisition likelihood. The output is a ranked queue, not a promise that anyone will come back.

The model should query what you already hold: product held at close, tenure, balances in the months before exit, channel of last service, and any recorded closure reason classification. Bureau files from Experian or Equifax can add whether the person still looks credit-active, but they do not tell you they want your product again. A CRM such as Salesforce is a destination for the approved list, not a substitute for the score.

You are ranking people who closed so an outbound team spends time on names that look recoverable, with a reason attached. You are not converting every closer. Many closures are the right outcome for the customer and for the bank.

Consider two current-account closures in the same week. Person A moved a salary to another bank after a rate comparison and left no complaint. Person B closed after a card-fraud case, even if the case was later resolved. A ranker that treats "closed last month" as a single bucket will put both on the same dialer file. Person A may be a fair win-back candidate if timing and reason line up. Person B is usually a suppress, not a prospect. The split is the point of the example. It is not a measured campaign result.

Do not treat the entire closed population as winnable. Estate closures, deceased markers, product replacements you already fulfilled, and customers who left after a serious service failure belong off the campaign. Scoring is useful only when the input set is recently closed and still eligible to hear from you.

Define the cohort with operations and conduct: close-date window, products you may re-offer, and drop warehouse records tagged as closes that are internal product switches.

A rank without a cite is a guess with a sort order. For each person on the queue, attach why this person and why now in language a reviewer can check against source systems.

Why this person: the features that moved the score. Legitimate cites include a closure coded as price or competitor switch, a still-active relationship on another product, or a bureau signal that the customer remains in-market for credit. Why now belongs with winback timing prediction: a cooling-off window after close, a rate-review season, or a life event you are allowed to use. If you cannot name the timing logic, you are not ready to call.

Write the cite as a short, inspectable string. Closed a packaged current account twenty-three days ago; reason code competitor-switch; still holds a mortgage; no open complaint. That is usable. A two-decimal propensity printed as a percentage chance is not. A reviewer cannot confirm a naked probability, and a campaign brief that treats that number as a fact will leak into scripts. Do not tell a customer they are a percentage likely to come back.

The human who confirms the list is checking the cite against policy, not re-scoring the model. If the cite points at data the bank should not use for marketing (health, hardship notes, or a vulnerability flag), drop the person even if the rank is high.

This work sits downstream of pre-closure churn interception. Interception tries to keep an account open. Win-back starts after the close. Mixing the two files will contact people who already chose to leave and people who have not left yet with the same script.

Suppression is the quality gate. Score first if you want, but never release a name that fails a hard stop.

Open complaints: do not email, SMS, or call a closed account that still has a complaint open. The contact will look like pressure while a case is live, and it can reopen harm. Legal holds, subject-access or litigation markers, and collections-sensitive statuses belong on the same hard stop.

Fraud and scams: a customer who left after a fraud event is not a win-back lead. Contacting them to invite them back can reopen the harm of the incident and can look like the bank is using a crime as a marketing trigger. If the closure reason or case notes show fraud, an authorised push payment scam, or unauthorised access, suppress. Do not wait for the model to learn that those people do not convert. Unknown reason is not a free pass onto the dialer.

Vulnerable customers: run the same flags you use elsewhere, including vulnerable customer detection. Temporary hardship, mental-health indicators your policy treats as marketing-restricted, and known financial-abuse markers take the person off outbound win-back. Experian and Equifax attributes do not replace those flags. Salesforce campaign membership does not either. The suppress list wins.

Build the suppressions as an explicit layer: complaint open, legal, deceased, fraud-related close, vulnerability, do-not-contact, and any regulator-facing restriction your bank already maintains. Apply them after scoring so a high rank cannot override a stop. Log the suppress reason so a reviewer sees held for open complaint rather than a silent hole in the file.

Refresh suppressions on the send day, not only on the score day. A complaint filed overnight is enough to kill the row.

The ranked, cited, suppressed file is still a draft. A named human, typically a win-back or growth lead plus a conduct or complaints check, confirms before any channel fires.

Confirmation is a review of the queue, not a second model. Sample the top of the list and a slice of the middle. For each sampled row, can you see the cite, the source field, and the suppress result? Would you be willing to read the outbound copy next to that cite? If a row's story is left after fraud or complaint still open, it should not be in the sample because it should not be in the file.

Only after confirmation do you load Salesforce or the equivalent campaign object, assign owners, and allow dialer or email. The CRM is the release mechanism. It is not the decision.

Keep the approved list versioned. If a complaint lands between score and send, the name must drop without waiting for the next model run. Win-back of a closed account is still a customer contact. The fact that the product is closed does not lower the bar.

Lookalikes are a different job. Lookalike audience generation finds people who resemble valued customers and have not already left you. Do not pipe closed-account scores into a prospect lookalike and call it win-back. The legal basis, the script, and the suppressions are not the same.

The score is an ordering device. It says the person at the top of the eligible file is a better use of an agent's hour than someone far down the same file, given the features you used. It does not say that person will return, and you must not invent a re-acquisition rate for the page, the dashboard, or the call script.

Treating every closer as winnable. Volume targets push teams to fill the queue with anyone who closed. That produces contact with bereaved households, fraud victims, and people who already told you they were done. Eligibility first, then rank.

Contacting someone who left after a fraud event. If the reason code is missing, do not assume unknown is safe. Unknown reason is a research task or a suppress, not a default into outbound.

Publishing the probability as a fact. Dashboards that label a score as a come-back chance train managers to treat a model output as a measured forecast. You have a rank. Say so. If you later count reinstatements, report those as outcomes of a named campaign and period, not as the model's personality.

Quality for this use case is a queue a human can defend: recently closed, scored from portfolio data, each row cited, complaint and legal and vulnerable and fraud stops applied, confirmed before contact. Anything that skips those steps is a list of people you used to bank, not a win-back programme.