Customer Case Study · Sales · Productivity Analytics
The lender's sales data was fragmented and inconsistent, so nobody could say with confidence who was performing. HMC cleaned it into one dataset, segmented every salesperson by region and product, modelled what separates performers from the rest, and turned it into insights sales leaders could act on.
Figures from the analysis shown below; business outcomes indicative.
The lending business sells Business Loans, Loans Against Property, K12 school finance and direct-to-consumer products across five regions. Performance lived in separate, inconsistent files, so every review started with an argument about the numbers.
No structured way to tell high performers from non-performers, or how close the middle was to the line.
Fragmented, unclean and inconsistent sales data across sources, with the same value spelt five ways.
No single source of truth and no visual dashboards, so decisions leaned on whoever had the latest file.
Region-level averages hid where productivity was really failing: specific products in specific regions.
Built with the lender's sales leadership, from raw extracts to an insight pack they could take into regional reviews.
Select a step to see what happened.
Before any analysis, the business agreed one definition.
Deliverable: Performance definitions and data inventory.
Excel and Power Query, then Python for the last mile.
Deliverable: Clean analysis dataset and data-quality log.
Python for segmentation, trend detection and outlier identification.
Deliverable: Insight pack and model notebook.
Insights went straight into regional reviews.
Deliverable: Intervention list by region and product.
The same pipeline reruns on new data.
Deliverable: Reusable pipeline and dashboard template.
The original analysis first, then four working modules: clean the data, explore region × product, convert potential performers, and compare the models.
Shown as delivered. Click any screen to enlarge it.
A sample of raw sales records with the problems the project faced. Press Clean or step through: pink cells are inconsistent, and the column-health bars on the right shrink to one value per real category.
The regional table from the analysis as a marimekko: column width is a region's headcount, block height is a product's share, colour is the performer rate. Click any block for detail; insights update with your view.
Counts from the region × product table in the analysis (L2–L3 population).
The tier flow from the analysis, live. Move the slider to convert potential performers into performers and watch the flow, the mix and the performer share change. Push further to see what converting some non-performers would take.
Both models were trained on the same features. Compare their scores, then move the decision threshold on either model to trade precision for recall. Test-set predictions are reconstructed to match the reported metrics.
With one clean dataset and a clear view of tiers, sales leaders focused effort where it would move people over the line.
Share of the L2–L3 sales force at or above 50% of target
Days per review cycle
Figures indicative; validated metrics available on request.
Once every label meant one thing, reviews moved from debating numbers to deciding actions.
Fifteen potential performers were the cheapest six points of performer share on the table.
Logistic regression beat the random forest here, and its drivers were easier for sales leaders to trust.