Employee Attrition Analysis & Prediction (Salifort Motors)

1. Business Context & Problem

Salifort Motors is experiencing increasing employee attrition, impacting productivity, team stability, and overall business performance.

The HR team lacked clarity on:

• What factors are driving employees to leave
• Whether attrition can be predicted before it happens
• What actionable steps can reduce employee turnover

Replacing an employee can cost up to 50–200% of their annual salary, making attrition not just an HR issue, but a significant business risk.

---

2. Exploratory Data Analysis (EDA) & Key Insights

Data Preparation & Cleaning:

• Removed 3008 duplicate records from 14,999 entries to ensure data integrity
• Checked for missing/null values (none significant)
• Standardized and cleaned column names for consistency
• Analyzed feature distributions to understand variable behavior
• Identified and reviewed outliers (especially in workload and tenure)

Key Insights:

• Attrition is not random — it follows clear, measurable patterns

• Low Satisfaction = High Risk

  • Employees with satisfaction levels below ~0.5 show significantly higher attrition

• Workload is a critical driver

  • Employees working 270+ hours/month or handling 6+ projects show extremely high attrition rates
  • Overworked employees had up to 87% likelihood of leaving

• Tenure effect (3–5 years)

  • Mid-tenure employees are the most vulnerable group, indicating stagnation or unmet expectations

• Performance paradox

  • High-performing employees are leaving at similar rates as low performers → burnout, not incompetence

• Low signal features

  • Department, promotions, and work accidents showed minimal impact on attrition

---

3. Machine Learning Approach

Feature Engineering:

• Created overwork_flag to capture extreme workload conditions (hours + projects)
• Encoded categorical variables (salary levels)
• Removed low-impact features (department, work_accident, promotion_last_5years)

Data Splitting Strategy:

• Train / Validation / Test split to ensure unbiased evaluation
• Validation set used for tuning
• Test set kept untouched for final performance

Model Optimization:

• Applied GridSearchCV for hyperparameter tuning
• Focused on balancing precision vs recall for business usability

---

4. Model Selection & Justification

Three models were evaluated:

Decision Tree

• Precision: 0.33
• Recall: 0.94
• Accuracy: 0.67
• F1: 0.41
• Insight: High recall but poor precision → too many false alarms

XGBoost

• Precision: 0.70
• Recall: 0.94
• Accuracy: 0.92
• F1: 0.80
• Insight: Strong performance but slightly less balanced than Random Forest

Random Forest (Selected Model)

• Precision: 0.93
• Recall: 0.92
• Accuracy: 0.97
• F1: 0.92

👉 Why Random Forest? It achieved the best trade-off between precision and recall, minimizing false alarms while still identifying most employees who leave — making it the most practical model for HR deployment.

---

5. Final Model Performance

• Accuracy: 98%
• Recall: 92% → Detects 9 out of 10 employees who leave
• Precision: 93% → Most flagged employees are actually at risk

Confusion Matrix Insight:

• Very low false positives → HR efforts remain focused
• Very low missed cases → high-risk employees are rarely overlooked

---

6. Business Impact

The model enables:

• Proactive intervention → Identify at-risk employees before resignation
• Targeted HR action → Focus on high-risk individuals only
• Data-driven strategy → Shift from reactive to predictive HR

👉 Key Insight: Attrition is primarily a workload and engagement problem, not a compensation problem.

---

7. Recommendation Snapshot

1. Enforce Workload Limits

• Cap project assignments at 5
• Flag employees exceeding 270 hours/month

2. Focus on Mid-Tenure Employees (3–5 years)

• Introduce career development plans
• Improve internal mobility

3. Replace Annual Surveys with Quarterly Pulse Checks

• Track satisfaction continuously
• Act before disengagement escalates

4. Protect High Performers from Burnout

• Avoid rewarding performance with more workload
• Introduce flexibility and non-monetary incentives

5. Deploy the Model as a Continuous HR Tool

• Integrate into HR workflows
• Retrain regularly with updated data

---

Tools & Technologies

• Python (Pandas, NumPy)
• Data Visualization (Matplotlib, Seaborn)
• Machine Learning (Scikit-learn, XGBoost)
• Model Tuning (GridSearchCV)
• Model Evaluation Metrics (Recall score, Precision score, F1 score, Accuracy score)

---

Project Assets

• attrition_analysis.ipynb → Full analysis and modeling
• dataset.csv → Dataset
• executive_presentation.pdf → Executive summary

---

Final Takeaway

Attrition at Salifort Motors is predictable, concentrated, and preventable.

With the right data and model, organizations can move from reacting to attrition → to actively preventing it.