Churn Rescue

Churn rescue — find users slipping away

"Which whales are leaking?" / "流失预警 / 找出停止交易的用户"

Identifies users whose activity has degraded materially, segments them by stage of decay, and prioritizes by lifetime value at stake.

When to use this skill

• User asks about churn, slipping, sleeping, rescue, 流失, 沉睡
• After a daily briefing flags a downward trend in active trader count
• Quarterly retention review
• Before a re-engagement campaign

What the agent does

There is no lastTradeTime field exposed by the MCP, so the agent has to infer "last active" by comparing lifetime totals to recent-window activity.

Step 1 — Pull lifetime top contributors

{
  "name": "okx-affiliate-invitee-list",
  "arguments": {
    "page": "1", "limit": "50",
    "orderBy": "rebate", "orderDir": "desc",
    "periodType": "total"
  }
}

This is the "all-time whales" baseline.

Step 2 — Pull last-30-day top contributors

{
  "name": "okx-affiliate-invitee-list",
  "arguments": {
    "page": "1", "limit": "50",
    "orderBy": "rebate", "orderDir": "desc",
    "periodType": "last_30d"
  }
}

This is the "currently active whales" set.

Step 3 — Diff and classify

For each UID in the lifetime top set, compute:

• decay_ratio = 1 - (last30_rebate / (lifetime_rebate / months_since_join))

Bucket users:

Step 4 — For each decayed/churned user, get withdrawal ratio

For the top 5 candidates, call:

{ "name": "okx-affiliate-invitee-detail", "arguments": { "uid": "<UID>" } }

Compute wdAmt / depAmt. This separates two churn modes:

• wdAmt / depAmt > 0.7 — Type A: cashed out, hard to win back
• wdAmt / depAmt < 0.1 and volMonth ≈ 0 — Type B: parked & idle, easiest to win back (capital still on platform)
• Otherwise — Type C: gradual disengagement

Step 5 — Optional: probe when they last traded

The new MCP does not expose lastTradeTime. To approximate, run invitee-list with a 1-month custom date range moving backward:

{
  "name": "okx-affiliate-invitee-list",
  "arguments": {
    "page": "1", "limit": "50",
    "orderBy": "rebate", "orderDir": "desc",
    "begin": "<one month ago in ms>",
    "end": "<two months ago in ms>"
  }
}

Check whether the UID appears with nonzero totalVol. Walk backward in 30-day chunks until you find activity. This is expensive — only do it for the top 5 users you actually plan to DM.

Insights to extract

• Total LTV at risk — sum lifetime rebate across decaying + churned users
• Largest single whale at risk — flag if any one user's lifetime rebate exceeds 5% of the node's total
• Type A vs Type B ratio — if mostly Type B (capital still on platform), reactivation is feasible; if mostly Type A, the channel may have a quality problem
• Decay timing pattern — if multiple users decayed in the same calendar month, look for an event (market crash, OKX policy change, competitor campaign)

Sample output format

⚠️ Churn watchlist — <YYYY-MM-DD>

📊 At-a-glance
• Lifetime top 50 users: 50
• Of those, currently active (last 30d top 50): X
• Effectively churned: Y
• Combined LTV at risk: $X,XXX,XXX

🔴 Type A — cashed out (urgent, but hard to win back)
| UID | Lifetime rebate | Withdrawal ratio | Last seen |
| ...XXXXXX | $XXX,XXX | 95% | YYYY-MM |

🟡 Type B — parked & idle (easiest win)
| UID | Lifetime rebate | Capital parked | Last seen |
| ...XXXXXX | $XX,XXX | $X,XXX,XXX | YYYY-MM |

🟠 Type C — gradual disengagement (intervene now)
| UID | Lifetime rebate | Recent trend |
| ...XXXXXX | $XX,XXX | declining MoM for N months |

🎯 Action plan
P0: <UIDs> — Type B users with parked capital, DM with VIP perks or Flash Earn pitch
P1: <UIDs> — Type C users still active, send market analysis to nudge them
P2: <UIDs> — Type A users, send exit survey to learn why they left

Recommended follow-ups

• "Generate outreach scripts for each bucket"
• "What did we do differently in ?" — switch to acquisition-trends and compare peak month behavior
• "Pull full history on the largest at-risk user" → whale-deep-dive

Gotchas

• No lastTradeTime field exists. Inferring it from period queries is slow and approximate. Don't burn API budget probing > 5 users.
• Withdrawal ratio is lifetime, not period-scoped. A user with wdAmt / depAmt = 0.7 may have withdrawn the bulk months ago and recently re-deposited. Cross-check with volMonth for current activity.
• Be conservative on "churned" labels. Volume can be lumpy — a single quiet month does not equal churn. Require ≥ 2 quiet months in a row, or a clear withdrawal pattern.
• subAffiliateUid filter is broken in the current MCP — to attribute churn to a sub-affiliate downline, filter client-side after pulling the list.