# Turning client data into flows and retention
A pension fund with a large mandate goes quiet. No calls, no new allocations, just a slow drift in tone during quarterly reviews. Six weeks later, a redemption notice arrives. The relationship team is blindsided.
The uncomfortable truth: the warning signs were sitting in the data the whole time. Reduced portal logins, unopened performance reports, a shrinking share of the client's total assets with your firm. Nobody connected them.
This lesson is about connecting them. We will use three data sources most asset managers already have (CRMCRMCustomer Relationship Management: software and strategy to manage and analyse customer interactions throughout their lifecycle.Voir la définition complète →, subscription and redemption records, and channel activity) to segment clients, flag redemption risk early, and focus distribution effort where it actually protects AUM (assets under management, the total market value of the money you manage).
Before any modeling, know what you are working with.
CRM data. Your system: contacts, meeting notes, relationship tenure, mandate type, and who owns the relationship. Often messy and incomplete, but it holds the human context.
Subscription and redemption data. The actual money movements. Subscriptions are inflows (clients adding money); redemptions are outflows (clients pulling money out). Net flow is the difference. This is the ground truth of retention.
Channel data. How clients engage. For institutional clients this means portal logins, report downloads, and consultant interactions. For wholesale clients (intermediaries like private banks, platforms, and independent financial advisers who distribute your funds to their own clients) it means platform activity, sales calls, and event attendance.
Each source alone is thin. Combined, they tell you who is happy, who is drifting, and where to spend the next hour of a salesperson's time.
SegmentationSegmentationDividing a market into distinct groups of customers who share similar needs, characteristics or behaviours, so each group can be served with a tailored approach.Voir la définition complète → just means grouping clients so you can treat different groups differently. Do not overcomplicate it.
Think pension funds, insurers, endowments, and sovereign wealth funds. Few relationships, very large tickets. Losing one can move your firm's total AUM.
Useful segmentationsegmentationDividing a market into distinct groups of customers who share similar needs, characteristics or behaviours, so each group can be served with a tailored approach.Voir la définition complète → axes:
Think fund platforms, private banks, and adviser networks. Many relationships, smaller individual tickets, but flows can be fast and herd-like.
Useful axes:
A simple, defensible approach is to score each client on two dimensions: value (revenue or AUM contribution) and risk (likelihood of leaving). Plot them. High value plus high risk is your priority quadrant. That is where retention effort pays off most.
You do not need a deep learning model to start. You need a handful of behavioral signals that reliably precede outflows.
Common leading indicators (patterns widely observed across the industry, not guarantees):
Here is a lightweight risk score you can build in a spreadsheet or a few lines of code. It is a starting point, not a black box.
import pandas as pd
# One row per client with pre-computed features
df = pd.DataFrame({
"client_id": [101, 102, 103],
"excess_return_12m": [-0.03, 0.01, -0.06], # vs benchmark
"login_change_pct": [-0.40, 0.05, -0.10], # QoQ engagement
"wallet_share_change": [-0.08, 0.00, -0.02], # falling = risk
"had_partial_redemption": [1, 0, 1],
})
# Simple weighted score; weights set with your distribution team
df["risk_score"] = (
(df["excess_return_12m"] < -0.02).astype(int) * 0.30 +
(df["login_change_pct"] < -0.20).astype(int) * 0.25 +
(df["wallet_share_change"] < 0).astype(int) * 0.25 +
df["had_partial_redemption"] * 0.20
)
print(df.sort_values("risk_score", ascending=False)[["client_id", "risk_score"]])The point is not the exact weights. The point is that you make the logic explicit, review it with the people who know the clients, and rank the book. Later you can replace the hand-set weights with a trained model once you have enough labeled redemption history.
A word of caution on models: correlation drifts. A signal that predicted redemptions in a rising-rate environment may behave differently in a falling one. Retrain and re-validate. And never let a score override an experienced relationship manager's judgment; let it prompt the conversation.
For a solid, free grounding in evaluating this kind of prediction responsibly, Google's Machine Learning Crash Course covers the fundamentals of features, evaluation, and fairness in plain terms.
A ranked list is useless without a playbook. Match the action to the segment and the risk driver.
High value, high risk, institutional. Senior-led intervention. If the driver is performance, get the portfolio manager in front of the client with a clear, honest attributionattributionA framework for assigning credit to the touchpoints that contributed to a conversion, so you can measure which channels and interactions actually drive results.Voir la définition complète → story (why performance happened). If the driver is a consultant review cycle, prepare early and get ahead of it.
High value, high risk, wholesale. Reinforce with the intermediary's own gatekeepers. Provide sales enablement, model portfolio positioningpositioningThe mental space you want your brand to occupy in your target customer's mind relative to alternatives.Voir la définition complète →, and timely commentary. Broaden the relationship beyond a single fund so it is harder to unwind.
High value, low risk. Protect and grow. These fund your firm. Look for cross-sell into adjacent strategies, but do not smother them.
Low value, high risk. Be honest about economics. Some relationships cost more to service than they return. Triage accordingly.
The discipline that matters: assign each flagged client an owner, a next action, and a date. Track whether the action changed the outcome. That feedback loop is how the score gets smarter and the team starts trusting it.
🎬 [VIDEO: "Customer ChurnCustomer ChurnChurn rate is the percentage of customers or revenue lost over a period. It measures how fast a business loses its existing customer base.Voir la définition complète → Prediction Explained" — youtube.com — a clear, non-technical walkthrough of how behavioral signals are turned into a churn (attrition) risk score, directly transferable to redemption risk]
Vérification des acquis
1. In the opening scenario, the pension fund's redemption caught the relationship team off guard. What is the core lesson the scenario is meant to illustrate?
2. Why does the lesson argue that combining CRM, flow, and channel data is more valuable than using any one source alone?
3. Subscription and redemption data is described as 'the ground truth of retention.' What does this framing imply relative to CRM and channel data?
4. Select ALL correct answers. Which of the following are legitimate early warning signals of redemption risk according to the lesson's logic?
Sélectionnez toutes les réponses correctes.
5. Select ALL correct answers. Which statements correctly distinguish institutional from wholesale clients in the context of channel data?
Sélectionnez toutes les réponses correctes.
Treating flows as the only signal. By the time redemptions show up, defense is expensive. Engagement and share-of-wallet data give you weeks or months of lead time.
Garbage CRM data. If meeting notes are blank and mandate fields are stale, no model saves you. Data hygiene is unglamorous but decisive. Assign ownership for keeping key fields current.
Confusing activity with retention. A client can log in daily while quietly preparing to leave (checking performance nervously). Read signals in context, not isolation.
Privacy and consent. Client engagement data is subject to regulation such as the EU's GDPR (General Data Protection Regulation, which governs how personal data is collected and used). For institutional entities the rules differ from individual data, but wholesale and private-client contexts can involve personal data. Involve compliance before you build. This is guidance, not legal advice.
Over-automating the human part. The score prioritizes. People retain. The best outcomes come from a good analyst handing a good salesperson a short, ranked, well-explained list.
Start small. Pick one channel or one strategy. Build the risk score with three or four transparent signals. Sit with the distribution team and pressure-test the top twenty names: do the flags match their gut? Adjust. Then run it live for a quarter and measure whether flagged clients who got intervention retained better than those who did not.
That single controlled comparison, even rough, is what turns a data project into a budget line.