## 1. OverTime (nominal)

**Pattern**

- OverTime = No: score ≈ **‑0.4** → noticeably **less likely** to leave.

- OverTime = Yes: score ≈ **+1.2** → **substantially more likely** to leave.

- This is one of the largest single‑feature jumps in the model.

**Business insight**

Heavy overtime is one of the clearest signals of flight risk. People working overtime regularly are much more prone to attrition than those who don’t.

**Actions**

1. **Overtime monitoring & early‑warning rule**

  - Flag employees with consistent overtime as “high‑risk”.  

  - Feed this directly into HR dashboards as a leading indicator.

2. **Workload & staffing interventions**

  - Review staffing in teams with structurally high overtime; consider hiring or rebalancing work.

  - Cap overtime hours where feasible; require manager justification beyond a threshold.

3. **Wellbeing initiatives**

  - Offer recovery time / comp days after peak periods.

  - Provide stress‑management and burnout‑prevention programs, targeted first at high‑overtime teams.

---

## 2. NumCompaniesWorked (continuous)

**Pattern**

- 0–3 companies: scores slightly below zero (around **‑0.2 to 0**): **lower to average risk**.

- 4 companies: score crosses above zero (~**+0.1**).

- 5+ companies: scores rise steadily towards **+0.8–0.9**.

- The increase from 0/1 to 8/9 companies is almost 1 logit, so the effect is strong and monotonic.

- Histogram shows most people are in the 0–4 range; very high counts worked in only 1 company.

**Business insight**

Employees who have already moved between many employers are systematically more likely to leave again. This is a **behavioral trait** (higher external mobility / “job hopper” tendency), not something you can change easily.

**Actions**

1. **Hiring stage**

  - Treat high NumCompaniesWorked as a **retention risk factor** when recruiting for roles where continuity is critical.

  - For such hires, probe more deeply into reasons for moves and expectations for tenure.

2. **Onboarding & career planning**

  - For employees with 4+ prior companies, develop a clear 12–24‑month development path early, so they see a reason to stay.

  - Pair them with mentors; schedule earlier career conversations (e.g., within first 6 months instead of 12).

3. **Retention segmentation**

  - Use NumCompaniesWorked together with other features to define “structurally mobile” segments and design **shorter, more intensive engagement cycles** (more frequent recognition, early promotions where justified, etc.).

---

## 3. EnvironmentSatisfaction (1–4)

**Pattern**

- 1 (“Low”): score around **+0.7 to +0.8** → **high attrition risk**.

- 2 (“Medium”): near **0** (baseline).

- 3 and 4 (“High/Very High”): around **‑0.2 to ‑0.3** → **lower attrition risk**.

- Histogram: many employees are at levels 3 and 4, fewer at 1 and 2.

**Business insight**

Perceived work environment (physical conditions, climate, tools, culture) strongly influences retention, especially at the **low end**. Dissatisfied employees are clearly more likely to leave.

**Actions**

1. **Targeted environment improvements**

  - Identify units with a high share of “1” ratings; conduct focus groups to pinpoint issues (workplace safety, noise, equipment, facilities, manager behavior, etc.).

  - Prioritize quick visible fixes to signal responsiveness.

2. **Pulse surveys & local dashboards**

  - Run quarterly environment‑satisfaction pulses; track distribution of 1–4 by team.

  - Set thresholds: e.g., more than X% of “1” ratings triggers an action plan and a follow‑up survey after 3–6 months.

3. **Manager accountability**

  - Include environment satisfaction in manager scorecards.

  - Provide training for managers in creating psychologically safe, supportive environments.

---

## 4. Age (continuous)

**Pattern**

- Ages ~18–33: score declines from about **+0.6** down towards **0** at 33. Younger employees are **more likely** to leave.

- 34–50: mostly slightly **negative** scores (~‑0.2 to ‑0.1) → **more stable** population.

- 50–57: score drifts back towards **0** (marginally higher risk than 40s but still moderate).

- 58–60: sharp **drop to ~‑1.6**. Note the confidence band is wide here (few employees), but the model associates late‑50s with **very low attrition**.

**Business insight**

- Your **highest attrition risk by age is in early‑career staff** (late teens to early 30s).

- Mid‑career employees (mid‑30s–40s) are relatively stable.

- Employees very close to retirement leave less (at least in your historical data), perhaps because those who planned to change jobs already did so earlier.

**Actions**

1. **Early‑career retention strategy**

  - Offer structured career paths, rotations, and learning programs targeted at <30–33.

  - Emphasize progression milestones in years 1–3 (title evolution, responsibilities, skill badges).

2. **Age‑sensitive interventions**

  - For younger staff with other risk factors (e.g., overtime + low satisfaction), treat them as **priority retention candidates**.

  - For older staff, focus on knowledge transfer, flexible work, and phased retirement planning rather than attrition prevention alone.

3. **Messaging**

  - Employer branding and internal communication for young employees should highlight growth, learning, and internal mobility opportunities.

---

## 5. JobSatisfaction (1–4)

**Pattern**

- 1 (“Low”): score ≈ **+0.5–0.6** → **higher attrition**.

- 2: slightly above 0.

- 3: near baseline.

- 4 (“Very satisfied”): ≈ **‑0.3 to ‑0.4** → **lower attrition**.

- The shape mirrors EnvironmentSatisfaction: **strong penalty for low values, benefit for high values**.

**Business insight**

Overall job satisfaction matters a lot. Very dissatisfied people leave; very satisfied ones stay. It’s not linear: the biggest jump is between 1 and 2/3, and another benefit at 4.

**Actions**

1. **Individual risk alerts**

  - Treat JobSatisfaction=1 as a **red flag**. Pair with manager outreach, stay interviews, and career discussions.

  - Move quickly: your data show high risk at this level.

2. **Job design & growth**

  - Address classic drivers of satisfaction: task variety, autonomy, recognition, role clarity, and career path.

  - Use internal mobility to move dissatisfied employees into better‑fit roles where possible.

3. **Link to performance and rewards**

  - Watch for high performers with low job satisfaction—this combo is especially dangerous. Offer differentiated development and retention packages.

---

## 6. DistanceFromHome (continuous)

**Pattern**

- 1–8 km: scores slightly negative (about **‑0.3** at the low end rising towards 0) → **lower risk** for people living close by.

- Around 10–12 km, score crosses zero and becomes **positive** (~+0.2–0.4) and continues rising gradually up to ~0.5 toward the furthest distances.

- Histogram: majority live within ~10 km; fewer at the long‑distance tail.

**Business insight**

Longer commutes are associated with higher likelihood of leaving, with a threshold around 10 km.

**Actions**

1. **Flexible work and commute support**

  - For employees with long commutes, consider:

    - Flexible hours / remote or hybrid days.

    - Transportation subsidies or parking solutions.

  - Prioritize interventions for those with other risk factors (e.g., overtime + long commute).

2. **Location‑aware workforce planning**

  - When opening new roles, where possible prefer candidates within a reasonable commute radius (especially for roles with limited flexibility).

3. **Targeted surveys**

  - Ask long‑distance employees about commute strain & potential mitigations; test pilot schemes (compressed work weeks, remote days).

---

## 7. MonthlyIncome (continuous)

**Pattern**

- Very low income (around 1–2.5k): **high positive scores** (~+0.7–0.9) → **strong attrition risk**.

- 3–7k: scores gradually drop below zero (down to about **‑0.3** at ~5–6k) → **lower risk**.

- 8–12k: scores hover around zero to slightly positive.

- 13–18k: scores become **negative again** (~‑0.3 to ‑0.5) → lower risk at some upper‑mid ranges.

- Extreme high end (approx. 19–20k): score spikes **strongly positive** (> +1). Confidence intervals are wide and histogram is thin, indicating **very few cases** here.

**Business insight**

- **Low pay is a clear attrition driver.** Once employees cross a mid‑range threshold, income ceases to be a strong risk factor.

- The spike at the very top likely reflects a **tiny group** with idiosyncratic behavior (e.g., senior specialists with external offers); treat with caution due to low sample size.

**Actions**

1. **Compensation benchmarking**

  - Identify employees in the lowest income deciles, especially when combined with high performance or long tenure.

  - Review pay equity: adjust salaries that lag market/peers significantly.

2. **Structured progression**

  - Ensure transparent pay bands and clear steps employees can take to progress from the at‑risk low‑income band into more stable ranges.

3. **High‑income niche group**

  - Even though sample size is small, monitor senior/high‑income roles closely for market poaching; consider tailored retention agreements or long‑term incentives.

---

## 8. YearsSinceLastPromotion (continuous)

**Pattern**

- 0–2 years since last promotion: scores around **0 or slightly negative** → baseline to **lower risk**.

- 3–5 years: scores become **positive** (~+0.2–0.4).

- 6+ years: risk steadily increases, reaching around **+1.4–1.5** at 15 years.

- Histogram: many employees have 0–3 years since last promotion; fewer at the extreme high values, but still non‑trivial.

**Business insight**

The longer someone goes without a promotion, the more likely they are to leave. This is **one of the strongest continuous drivers** of attrition in your model.

**Actions**

1. **Promotion‑stagnation alerts**

  - Flag employees with **>3 years since last promotion** as increasing risk, and those with **>5–6 years** as high risk.

  - For each flagged employee, require a manager review:

    - Is promotion justified but blocked?

    - Are lateral moves / special projects possible?

    - Or is there a mismatch that should be addressed openly?

2. **Career frameworks**

  - Implement clearer progression steps, including lateral and “expert” tracks so employees can advance without always changing job titles.

3. **Manager expectations**

  - Coach managers to have regular career conversations (at least annually) and to avoid long unaddressed stagnation for strong performers.

---

## 9. StockOptionLevel (0–3)

**Pattern**

- Level 0: slight **positive** score (~+0.2) → mildly higher risk.

- Level 1 and 2: scores drop into **negative** region (~‑0.2 to ‑0.3) → **lower attrition risk**.

- Level 3: moves back towards 0 (but with few data points and wider intervals).

**Business insight**

Offering at least some stock/equity (levels 1–2) correlates with better retention relative to no stock options. The effect is moderate but consistent.

**Actions**

1. **Broaden equity participation**

  - For roles where feasible, move employees from Level 0 to Level 1–2 to improve attachment to the firm.

  - Consider **targeted stock grants** as part of retention packages for high‑risk / high‑value individuals.

2. **Communicate value**

  - Ensure employees understand the value and vesting conditions of stock options; poorly understood equity won’t retain.

---

## 10. JobRole – Research Director (binary)

**Pattern**

- Not Research Director: near **0** impact.

- Research Director = True: strong **negative score (~‑1.8)** → **much lower probability** of leaving.

- Histogram: Research Directors are a **small group**, but their behavior is very stable.

**Business insight**

Research Directors are historically very loyal / stable within your dataset. This might reflect high satisfaction, strong pay, meaningful work, or limited external alternatives.

**Actions**

1. **Protect & leverage**

  - Maintain the conditions that support this stability (compensation, autonomy, recognition).

  - Engage them as **mentors** to help improve satisfaction and career clarity for more at‑risk groups (e.g., younger staff).

2. **Don’t over‑rely**

  - Because they are few, losing even one could be painful despite low risk. Monitor individually rather than assuming zero risk.

---

## Cross‑feature takeaways (who is most at risk?)

High‑risk archetypes tend to combine multiple positive‑score conditions:

- Works **overtime**,  

- Is **young (18–33)**,  

- Has **low environment and job satisfaction (1–2)**,  

- Has worked at **many previous companies (≥4)**,  

- Lives **far from the office (>10 km)**,  

- Is at the **low end of the pay scale**,  

- Has had **no promotion for 3+ years**,  

- Has **no stock options**.

Low‑risk archetypes:

- No overtime,  

- Age mid‑30s to mid‑40s (or near‑retirement in your data),  

- High satisfaction with job and environment (3–4),  

- 0–3 prior companies,  

- Short commute,  

- Mid‑range or reasonably competitive income,  

- Recent promotion (<2–3 years),  

- Some stock options,  

- Research Directors.

---

## Using the model to improve business outcomes

1. **Individual‑level risk scoring**

  - Deploy the EBMs to produce a **monthly attrition risk score per employee**.

  - Combine with business rules from PDPs (e.g., “YearsSinceLastPromotion > 5” or “OverTime=Yes & JobSatisfaction=1”) to create interpretable flags.

2. **Tiered intervention playbook**

  - Define risk tiers (e.g., low/medium/high) based on predicted probability and key drivers.

  - For each tier, specify actions:

    - High: immediate manager outreach, stay interview, review of pay, progression options, workload, commute flexibility.

    - Medium: scheduled career conversation; monitor satisfaction over next survey cycles.

    - Low: standard engagement programs.

3. **Manager dashboards**

  - Provide managers with **team‑level explanations**:

    - “Top drivers of predicted attrition on your team are: overtime, years since last promotion, low job satisfaction.”

  - Show distributions (e.g. share of team with >3 years since last promotion) and benchmarks across the company.

4. **Policy & program design**

  - Use aggregated PDP insights to redesign policies:

    - Overtime policy & staffing models.

    - Promotion cadence and career frameworks.

    - Compensation adjustments for lowest‑paid groups.

    - Commute flexibility guidelines.

5. **Measure impact**

  - After implementing interventions (e.g., reducing overtime, raising low salaries, new promotion guidelines), **track changes over time**:

    - Compare predicted vs. realized attrition.

    - Re‑train the model periodically to see whether PDP shapes change, indicating successful mitigation (e.g., overtime becomes a weaker predictor).

---

### Prioritized recommendations

If you had to act on a short list first (balancing effect size and population size):

1. **Reduce structural overtime** and support wellbeing where overtime is unavoidable.  

2. **Address promotion stagnation**: identify employees with >3 years since last promotion and create individualized plans.  

3. **Tackle dissatisfaction**: systematically follow up with employees rating EnvironmentSatisfaction or JobSatisfaction = 1.  

4. **Review compensation for the lowest income bands**, especially for younger and high‑performing employees.  

5. **Increase flexibility for long‑distance commuters** (remote/hybrid options) and monitor their overtime and satisfaction closely.

Using this EBMs‑based explanation layer, you can both **predict** who is likely to leave and, more importantly, **understand why**, so HR and leaders can intervene in targeted, evidence‑based ways.

‍