algo-hr-turnover
建立員工離職預測模型，找出風險與留任因素
中文翻譯
---
name: "algo-hr-turnover"
description: "建立員工離職預測模型，找出離職風險與留任的驅動因素。當使用者需要預測哪些員工可能離開、找出留任風險因子，或排定 HR 介入的優先順序時使用，即使他們說的是『attrition prediction』『who is going to quit』或『employee retention model』。"
metadata:
category: "WP-42 HR 演算法"
tags: ["hr", "turnover-prediction", "attrition", "retention"]
---
# 員工離職預測
## 概述
離職預測用分類模型（邏輯迴歸、隨機森林、XGBoost）估計員工在特定期間內（通常 6-12 個月）離開的機率。特徵包含年資、薪酬、績效、晉升歷史和敬業度訊號。
## 何時使用
**觸發條件：**
- 找出高自願離職風險的員工
- 量化是哪些因素在推動離職，以便有針對性地介入
- 把留任預算優先投向影響最大的員工
**不該使用的情況：**
- 規劃非自願資遣時（流程和倫理都不同）
- 員工數 < 200 時（資料不足，無法可靠建模）
## 演算法
```
IRON LAW: Turnover Models Predict RISK, Not Certainty
A predicted 80% turnover probability means "employees with similar
profiles historically left 80% of the time." It does NOT mean this
specific employee WILL leave. Never use model outputs as sole basis
for employment decisions — that creates legal and ethical liability.
```
### 階段 1：輸入驗證
收集：員工人口統計、年資、薪酬（相對市場）、最近一次晉升日期、績效評等、主管異動歷史、敬業度調查分數、通勤距離。結果變數：N 個月內自願離職。
**關卡：** 至少 200 筆離職事件，且特徵在離職日之前就取得。
### 階段 2：核心演算法
1. 特徵工程：年資分組、薪資比（薪水/市場中位數）、距離上次晉升的時間、主管年資、敬業度趨勢
2. 處理類別不平衡：離職率通常 10-20%。用 SMOTE 或 class weight。
3. 訓練：邏輯迴歸（可解釋，HR 偏好）或 GBDT（準確度較高）
4. 輸出：每位員工的離職機率 + 主要風險因子
### 階段 3：驗證
評估：AUC、precision-recall（在可行動的門檻下）。回測：模型有沒有正確標出過去 6 個月離職的員工？
**關卡：** AUC > 0.70，最高十分位的 precision > 50%。
### 階段 4：輸出
回傳風險分數，附上驅動因子分析。
## 輸出格式
```json
{
"risk_scores": [{"employee_id": "E123", "turnover_prob": 0.72, "risk_tier": "high", "top_drivers": ["low_comp_ratio", "no_promotion_3yr"]}],
"metadata": {"model": "xgboost", "auc": 0.78, "prediction_window_months": 12}
}
```
## 範例
### 輸入輸出範例
**輸入：** 員工：年資 4 年、薪資比 0.85、3 年沒升遷、敬業度分數下降中
**預期：** 高風險（>0.6）。主要驅動因子：薪酬低於市場、職涯停滯。
### 邊界情況
| 輸入 | 預期 | 原因 |
|-------|----------|-----|
| 新進員工（< 6 個月） | 預測不可靠 | 行為資料不足 |
| 績效頂尖、薪酬高 | 仍可能離開 | 非財務因素（主管、文化）也很重要 |
| 組織重整後的時期 | 模型很可能漂移 | 異常狀況會扭曲模式 |
## 注意事項
- **倖存者偏誤**：訓練資料只包含被雇用、而且待得夠久才能被觀察到的人。很早期就離開的人可能代表性不足。
- **特徵洩漏**：「開始找工作」或「更新 LinkedIn」是很強的預測因子，但拿來用在倫理和法律上都有問題。只用公司內部的 HR 資料。
- **自我實現預言**：如果主管對「高風險」員工另眼看待（少投資、少給專案），模型預測就會變成自我實現。
- **法律限制**：直接或透過代理變數使用受保護的屬性（年齡、性別、族群），可能違反勞動法規。要稽核是否有差別性影響。
- **留任介入的時機**：找出風險，要 HR 真的採取行動才有用。要把模型放進有具體介入觸發條件的留任流程裡。
## 參考資料
- 從 HR 資料做特徵工程，見 `references/hr-features.md`
- HR 應用中的倫理 AI，見 `references/ethical-hr-ai.md`英文原文
---
name: "algo-hr-turnover"
description: "Build employee turnover prediction models to identify flight risk and retention drivers. Use this skill when the user needs to predict which employees are likely to leave, identify retention risk factors, or prioritize HR interventions — even if they say 'attrition prediction', 'who is going to quit', or 'employee retention model'."
metadata:
category: "WP-42 HR 演算法"
tags: ["hr", "turnover-prediction", "attrition", "retention"]
---
# Employee Turnover Prediction
## Overview
Turnover prediction uses classification models (logistic regression, random forest, XGBoost) to estimate the probability an employee will leave within a defined period (typically 6-12 months). Features include tenure, compensation, performance, promotion history, and engagement signals.
## When to Use
**Trigger conditions:**
- Identifying employees at high risk of voluntary departure
- Quantifying which factors drive turnover for targeted interventions
- Prioritizing retention budgets toward highest-impact employees
**When NOT to use:**
- For involuntary termination planning (different process and ethics)
- When headcount is < 200 (insufficient data for reliable modeling)
## Algorithm
```
IRON LAW: Turnover Models Predict RISK, Not Certainty
A predicted 80% turnover probability means "employees with similar
profiles historically left 80% of the time." It does NOT mean this
specific employee WILL leave. Never use model outputs as sole basis
for employment decisions — that creates legal and ethical liability.
```
### Phase 1: Input Validation
Collect: employee demographics, tenure, compensation (relative to market), last promotion date, performance ratings, manager change history, engagement survey scores, commute distance. Outcome: voluntary departure within N months.
**Gate:** Minimum 200 turnover events, features available before departure date.
### Phase 2: Core Algorithm
1. Feature engineering: tenure buckets, comp ratio (salary/market median), time since last promotion, manager tenure, engagement trend
2. Handle class imbalance: turnover rate typically 10-20%. Use SMOTE or class weights.
3. Train: logistic regression (interpretable, HR-preferred) or GBDT (higher accuracy)
4. Output: probability of departure + top risk factors per employee
### Phase 3: Verification
Evaluate: AUC, precision-recall (at actionable thresholds). Backtest: did the model correctly flag employees who left in the past 6 months?
**Gate:** AUC > 0.70, precision > 50% at top decile.
### Phase 4: Output
Return risk scores with driver analysis.
## Output Format
```json
{
"risk_scores": [{"employee_id": "E123", "turnover_prob": 0.72, "risk_tier": "high", "top_drivers": ["low_comp_ratio", "no_promotion_3yr"]}],
"metadata": {"model": "xgboost", "auc": 0.78, "prediction_window_months": 12}
}
```
## Examples
### Sample I/O
**Input:** Employee: 4yr tenure, comp ratio 0.85, no promotion in 3yr, engagement score declining
**Expected:** High risk (>0.6). Top drivers: below-market compensation, stalled career progression.
### Edge Cases
| Input | Expected | Why |
|-------|----------|-----|
| New hire (< 6 months) | Unreliable prediction | Insufficient behavioral data |
| Top performer, high comp | Still could leave | Non-financial factors (manager, culture) matter |
| Post-reorg period | Model drift likely | Unusual conditions distort patterns |
## Gotchas
- **Survivorship bias**: Training data only includes people who were hired and stayed long enough to observe. Early-stage leavers may be underrepresented.
- **Feature leakage**: "Started job searching" or "updated LinkedIn" are strong predictors but ethically and legally problematic to use. Stick to internal HR data.
- **Self-fulfilling prophecy**: If managers treat "high risk" employees differently (less investment, fewer projects), the model prediction becomes self-fulfilling.
- **Legal constraints**: Using protected attributes (age, gender, ethnicity) directly or via proxies may violate employment law. Audit for disparate impact.
- **Retention intervention timing**: Identifying risk is only useful if HR acts. Build the model into a retention workflow with specific intervention triggers.
## References
- For feature engineering from HR data, see `references/hr-features.md`
- For ethical AI in HR applications, see `references/ethical-hr-ai.md`