Customer Case Study · FLS Attrition Prediction
Front-line sales is where life insurers lose people fastest, often in the first six months. HMC built a Python machine-learning model on a leading life insurer's HR, sales and activity data that scores every front-line seller each month, explains why each one is at risk, and tells managers what to do about it.
Figures indicative; validated metrics available on request.
Front-line sales attrition was measured after the fact, explained by exit interviews and treated with one policy for everyone.
Most leavers went in their first year, before recovering their hiring and licensing cost.
Monthly attrition reports told leaders what had already happened.
HRMS, sales MIS, incentive files and CRM logins were never joined at seller level.
Retention spend went to everyone, or to whoever shouted loudest.
Built in Python with HR, sales operations and channel heads; reviewed every month.
Select a step to see what happened.
HRMS, sales MIS, incentive payouts and CRM activity at seller-month level.
Deliverable: Analytical dataset and data dictionary.
Every feature had to be something a manager could act on or explain.
Deliverable: Feature catalogue and leakage checks.
scikit-learn and LightGBM, explained with SHAP.
Deliverable: Scoring pipeline and risk desk.
Agency and bancassurance first.
Deliverable: Manager playbook.
Direct, tele-sales and partnerships added.
Deliverable: Roll-out plan.
A synthetic population of 1,800 past and 420 current front-line sellers. The machine-learning models really train in your browser, mirroring the Python pipeline shown alongside. Drag the 3D charts to turn them.
Optional. The model and every chart work without a key; retention plans and the CHRO summary come from built-in rules. Add your OpenAI key to have them written for each seller and manager.
Each dot is a past seller placed by tenure, target achievement and engagement. Red dots resigned within 90 days. Survival curves show how long each group stays; the hexagon map shows where leavers cluster.
Kaplan–Meier: share still employed by months since joining
Hexbin: tenure × target achievement · colour = share who left · size = sellers
Synthetic data generated to resemble a life insurer's front-line sales force. No real employee data is used.
A logistic-regression baseline and a gradient-boosted tree model, trained on 70% of sellers and tested on the other 30%. Change the features or settings and retrain; the Python tab shows the same pipeline as delivered.
True positives against false positives, test set
Gradient boosting log-loss by tree, train and test
Predicted risk against actual leaver rate, 10 bins
Test set, gradient boosting
Gradient boosting gain share against logistic-regression standardised coefficient
Models train in JavaScript on the synthetic data so the page runs anywhere; the delivered pipeline runs the Python shown, with LightGBM and SHAP.
Pick a seller to see why the model rates them at risk, try what-if changes, and see their risk landscape in 3D. Then draft the retention plan for their manager.
| Seller | Channel | Zone | Tenure | Target | Risk |
|---|
Force plot: each reason pushes risk up (red) or down (teal) from the average seller
Move a lever and the model re-scores this seller
The plan uses this seller's top reasons and the what-if result.
Expected leavers by zone and tenure as a 3D city; at-risk sellers grouped into personas by why they are at risk; and the threshold that saves the most money.
Net saving = leavers kept × replacement cost − cost of acting on everyone flagged
Dendrogram: sellers above the threshold, clustered on their reasons for risk
The summary combines the forecast, personas and the best threshold.
With the riskiest sellers reached every month, early exits fell and fewer hiring and licensing costs were written off.
Share of each joining month still employed, by months since joining
Combo chart: bars = exits within six months of joining (% of joiners); line = cost written off, ₹ lakh
Figures indicative; validated metrics available on request.
A risk score without reasons gets ignored; reason codes turn it into a manager's to-do.
Onboarding, early targets and a stable manager matter more than pay for new sellers.
Acting on everyone wastes budget; the cost curve shows where retention pays.