Everline Attrition Brief

05

Overtime is the lever. Satisfaction only modulates it.

Overtime and low satisfaction each look bad alone. Do they compound?

vw_overtime_satisfaction_attrition

Attrition rate by overtime × job satisfaction

Cell shows attrition rate, lift against the 16.1% company base rate, and segment size.

Adding work-life balance

Three conditions stacked, worst to best.

06

Engineering doesn't have a retention problem. It has a tenure mix.

Which departments lose more people than their workforce age explains?

vw_department_attrition_controlled · indirect standardisation

Standardised Attrition Ratio (observed ÷ expected leavers)

Expected leavers = the company-wide rate for each tenure cohort applied to that department's own tenure mix. 1.00 means “exactly what its workforce age predicts”.

Crude rate vs tenure-adjusted rate

Where the two diverge, the crude rate was measuring workforce age.

02

The highest rate and the biggest problem are different cohorts.

Do we lose people early, or after they're established?

vw_tenure_cohort_attrition

Attrition rate by tenure cohort

Tenure measured at termination for leavers, at snapshot for active employees.

Share of everyone who left

A high rate on a small cohort is a smaller problem than a moderate rate on a large one.

03

Pay position matters — but it isn't monotonic.

Are people underpaid relative to their peers leaving more?

vw_compensation_percentile · PERCENT_RANK() partitioned by role

Attrition by pay quartile within job role

Comparing salary across roles is meaningless; the percentile is computed inside each role.

04

Span of control explains nothing. Reported anyway.

Do managers with larger teams lose more people?

vw_manager_span_attrition

Attrition rate by manager span-of-control band

Aggregates people, not per-manager rates — averaging rates would weight a 4-person team like a 20-person one.

01

Quarterly rate and the rolling window.

Which departments are worst, and is one bad quarter distorting it?

vw_attrition_by_department · ROWS BETWEEN 3 PRECEDING

Company-wide quarterly attrition rate

Terminations over average headcount, by quarter.

Read this one as a technique demo, not a finding. The source dataset is a single dateless snapshot, so termination dates had to be synthesised. The upward slope here is a property of that construction. The rolling-window SQL is the deliverable; the trend is not evidence.

●

Attrition risk watchlist

Who is at elevated risk right now, and what do I say to their manager?

vw_attrition_risk_watchlist · logistic regression, scored 2025-12-31

Model output, not a verdict. These are associations, not causes, and carry roughly ±0.06 PR-AUC of uncertainty. Use as a prompt for a conversation — never as an input to a performance or employment decision. Names are synthetic.

■

The model, and why it's the boring one

Logistic regression vs random forest on a 16% minority class.

scikit-learn · stratified 75/25 split · 5-fold CV · balanced class weights

Held-out performance

Accuracy is excluded on purpose — predicting “nobody leaves” scores 83.9%.

What the model relies on

Permutation importance: drop in PR-AUC when a column is shuffled.

⁂

What's real and what's constructed

Stated up front, because it changes how every number above should be read.

Built on the public IBM HR Analytics Employee Attrition & Performance dataset — 1,470 real rows. Every driver variable and the attrition label are used unmodified: income, satisfaction scores, overtime, work-life balance, performance rating, distance from home, tenure, and the Attrition flag.

The source is a single dateless snapshot, so anything time-shaped was constructed to make SCD modelling and time-intelligence DAX possible: employee names, hire and termination dates, the per-year compensation and review rows, the manager hierarchy, and the six-department overlay.

Consequence: findings 2–6 are about drivers and are real signal. Finding 1's quarter-over-quarter slope is partly by construction and is labelled as such rather than presented as an insight.