Seeing assets leave before they leave: attrition modeling in wealth management
Outflows are almost never a surprise in hindsight. A model that reads the early signals — and hands advisors a specific, timely reason to make the call.
The problem
A wealth manager with tens of billions under advisement was losing assets it never saw leaving. Attrition showed up in the quarterly report as a number, long after anything could be done about it. When the firm went back and reconstructed the departures, the pattern was uncomfortable: nearly all of them had been visible months in advance, in signals nobody had joined together.
A client stops opening statements. Beneficiary details get updated. A partial transfer goes out to another custodian. Meeting cadence quietly slips from quarterly to nothing. Each of those sat in a different system, and no advisor could watch four hundred relationships closely enough to notice.
The hypothesis
Attrition is a predictable process with observable precursors, not an event. If those signals were unified per household and modeled against historical departures, advisors could be given a ranked list with enough lead time to act — and, critically, the specific reason to act on rather than a generic score.
The build
- Household-level data foundation — accounts, holdings, transfers, service interactions, meeting history, portal engagement, and life events joined to a single household identity with lineage. This was the majority of the effort.
- Time-to-event modeling — survival analysis rather than a binary churn flag, predicting probability of significant outflow inside a rolling horizon, so the output is a timeline rather than a verdict.
- Reason codes and evidence — each risk score carries its top contributing factors, expressed in advisor language: engagement dropped, partial transfer out, a life event with no follow-up, performance dispersion versus stated goals.
- Next-best-action, drafted not decided — an agent proposes a specific outreach with the context assembled, and the advisor edits and owns it. The system never contacts a client.
Design choice that mattered: we suppressed alerts we could not explain. An unexplained high score gets ignored by advisors after the second false alarm; a score attached to "two partial transfers out and no contact in five months" gets a phone call the same day. Explainability here was an adoption requirement, not a compliance checkbox.
Rollout
We backtested against three years of known departures, validating out-of-time rather than on random splits, and then ran six months in shadow mode where scores were recorded but not shown. That gave the firm an honest lift number before a single advisor workflow changed. Deployment started with two regions whose leadership was willing to hold advisors accountable for acting on alerts — without that, the model is just a report.
Results
The firm gained months of warning on at-risk households, and roughly a third of flagged outflows were retained after intervention — a figure the business measured against a matched control group rather than against hope. The unexpected benefit was diagnostic: the reason codes revealed two service failures nobody had escalated, and fixing those did more than any individual save.
What we'd tell you
- Model time-to-event, not a binary flag. Advisors need to know how long they have.
- Reason codes are the product. A bare score is ignored within a month.
- Hold out a control group, or you will never know whether the saves were yours.
- Never let the system contact a client. The relationship is the advisor's, and so is the message.
More from the field.
Consumer lending, decided in hours
Early-warning credit risk
The 6 a.m. advisor brief
Bring us a hypothesis. Leave with a system.
Tell us what's eating your team's time. We'll give you an honest read on whether AI is the right tool — and if it is, a scoped v1 with a timeline and cost.