2nd/10_Wiki/Topics/Other/이탈률(Churn Rate).md

---
id: wiki-2026-0508-이탈률-churn-rate
title: 이탈률(Churn Rate)
category: 10_Wiki/Topics
status: verified
canonical_id: self
aliases: [Churn, User Churn, Attrition Rate, 이탈]
duplicate_of: none
source_trust_level: A
confidence_score: 0.9
verification_status: applied
tags: [analytics, retention, kpi, business-metrics]
raw_sources: []
last_reinforced: 2026-05-10
github_commit: pending
tech_stack:
  language: python
  framework: pandas-lifelines
---

# 이탈률(Churn Rate)

## 매 한 줄
> **"매 churn 매 retention 의 의 mirror — 매 leak rate 의 quantify"**. 매 SaaS (5% monthly = catastrophe) 매 game (D30 retention 30% = norm), 매 의 의 의 의 의 metric 매 universal — 매 user lost / total user 의 unit time 의 의 의. 매 2026 modern stack 매 Cox proportional hazards + churn-prediction LightGBM 매 standard.

## 매 핵심

### 매 churn 의 정의 variants
- **logo churn (count)**: 매 #lost / #total. 매 simple.
- **revenue churn (gross)**: 매 lost MRR / total MRR. 매 enterprise SaaS 매 critical.
- **net revenue churn**: 매 (lost - expansion) / total. 매 < 0 (negative churn) 매 holy grail.
- **active churn vs passive churn**: 매 cancel button 매 / 매 expired card. 매 win-back 매 different strategy.

### 매 measurement window 매 매 매
- 매 daily / weekly / monthly / quarterly. 매 game 매 D1/D7/D30 매 standard. 매 SaaS 매 monthly.
- **cohort-based**: 매 sign-up week 의 의 group 의 churn curve 의 trace.
- **rolling**: 매 last-N-day window 매 trailing.

### 매 응용
1. 매 SaaS health: <5% monthly churn 매 healthy, >7% 매 alarm.
2. 매 mobile game: D1 ~40-50%, D7 ~20%, D30 ~5-10% retention (= 90-95% churn by D30).
3. 매 telecom: 매 churn prediction 매 ML 매 retention offer trigger.
4. 매 streaming (Netflix etc.): 매 voluntary cancel + 매 dunning (passive).
5. 매 freemium → premium conversion 매 churn 의 의 inverse calculation.

## 💻 패턴

### Basic churn calculation
```python
import pandas as pd

def monthly_churn(users: pd.DataFrame, month: str) -> float:
    """users: [user_id, signup_date, last_active_date]"""
    start = pd.to_datetime(month)
    end = start + pd.offsets.MonthEnd(1)
    active_at_start = users[users["last_active_date"] >= start - pd.Timedelta(days=30)]
    churned = active_at_start[active_at_start["last_active_date"] < end - pd.Timedelta(days=30)]
    return len(churned) / len(active_at_start) if len(active_at_start) else 0
```

### Cohort retention curve
```python
def cohort_retention(events: pd.DataFrame) -> pd.DataFrame:
    """events: [user_id, event_date]. Returns cohort x days_since_signup matrix."""
    first = events.groupby("user_id")["event_date"].min().rename("cohort")
    df = events.merge(first, on="user_id")
    df["days_since"] = (df["event_date"] - df["cohort"]).dt.days
    df["cohort_week"] = df["cohort"].dt.to_period("W")
    matrix = (
        df.groupby(["cohort_week", "days_since"])["user_id"]
        .nunique()
        .unstack(fill_value=0)
    )
    return matrix.div(matrix.iloc[:, 0], axis=0)  # normalize to D0
```

### Survival analysis (Kaplan-Meier)
```python
from lifelines import KaplanMeierFitter

def survival_curve(durations, event_observed):
    kmf = KaplanMeierFitter()
    kmf.fit(durations, event_observed)
    return kmf.survival_function_

# durations: days until churn (or censoring)
# event_observed: 1 if churned, 0 if still active (right-censored)
```

### Cox proportional hazards (predictive)
```python
from lifelines import CoxPHFitter

def churn_hazard(df: pd.DataFrame):
    """df: [duration, event, plan, monthly_usage, support_tickets, ...]"""
    cph = CoxPHFitter()
    cph.fit(df, duration_col="duration", event_col="event")
    return cph  # cph.hazard_ratios_ shows feature impact
```

### LightGBM churn prediction
```python
import lightgbm as lgb
from sklearn.model_selection import train_test_split

def train_churn_model(features: pd.DataFrame, churn_label: pd.Series):
    X_tr, X_te, y_tr, y_te = train_test_split(features, churn_label, stratify=churn_label)
    model = lgb.LGBMClassifier(
        n_estimators=500, learning_rate=0.05,
        class_weight="balanced", num_leaves=63
    )
    model.fit(X_tr, y_tr, eval_set=[(X_te, y_te)], callbacks=[lgb.early_stopping(20)])
    return model
```

### Negative churn (expansion > churn)
```python
def net_revenue_churn(start_mrr, lost_mrr, expansion_mrr) -> float:
    return (lost_mrr - expansion_mrr) / start_mrr  # negative = good
```

## 매 결정 기준
| 상황 | metric |
|---|---|
| 매 SMB SaaS | logo churn (monthly) |
| 매 Enterprise SaaS | net revenue churn |
| 매 mobile game | D1/D7/D30 retention curve |
| 매 telecom / streaming | survival + ML prediction |
| 매 marketplace | cohort retention + GMV per cohort |

**기본값**: 매 cohort retention 매 D1/D7/D30 + 매 monthly logo churn 매 dual-track tracking.

## 🔗 Graph
- 부모: [[Retention]] · [[KPI]] · [[Customer Analytics]]
- 변형: [[Logo Churn]] · [[Revenue Churn]] · [[Net Revenue Retention]]
- 응용: [[유니버스 LTV(Universe LTV)]] · [[CAC Payback]] · [[Cohort Analysis]]
- Adjacent: [[Survival Analysis]] · [[Engagement Score]]

## 🤖 LLM 활용
**언제**: 매 product / business 매 health 의 measurement, 매 retention 매 effort 의 ROI 의 prove, 매 churn-prediction model 의 build.
**언제 X**: 매 single-purchase 매 product (no recurring) — 매 LTV / repeat-rate 의 의 substitute.

## ❌ 안티패턴
- **매 averaging across cohorts**: 매 hide newer-cohort 매 improvement / regression.
- **매 treating active churn = passive churn**: 매 dunning fix 매 retention campaign 매 confused.
- **매 vanity tracking**: 매 churn 매 measure 만 매 매, 매 root-cause 의 의 의 segmentation 의 의 의 X.
- **매 D30 only**: 매 long-tail (D90/D180) 매 ignore — 매 LTV 매 underestimate.

## 🧪 검증 / 중복
- Verified (Reichheld *The Loyalty Effect*, ChartMogul SaaS benchmarks 2025-2026, lifelines library docs).
- 신뢰도 A.

## 🕓 Changelog
| 날짜 | 변경 |
|---|---|
| 2026-05-08 | Phase 1 |
| 2026-05-10 | Manual cleanup — churn variants + cohort curves + KM/Cox/LightGBM code |