Best Propects Case Study Header

Best Prospects Scoring Model (B2B Sales Targeting)

Built a weekly top-K prospect scoring system aligned to rep capacity (125 calls/week) using industry hierarchy, penetration signals, and proximity features—validated with time-based evaluation and productionized for repeatable scoring.

Which prospects should sales reps call this week?

  • Sales reps can only act on ~125 calls per week
  • Conversions are delayed from appearing in the system up to ~2 weeks
  • Base rates are low and vary by region/market/territory
  • Business needs consistent outputs
  • The top 125 need to have precision/conversion rate of 3% or higher

What types of data and labels were used?

  • Prospects + customer + outcomes
  • NAICS (industry codes) at the 2, 4, 6 levels
  • Basic firmographic data (business demographic data)
  • Geographic data
  • The scoring of a prospect occurred within a 7-day window
  • The outcome used was based on a prospect converting within 60 days after scoring
  • Note: To prevent double-counting of a prospect via a “seen” flag, which would remove the prospect after conversion during scoring

Success Metrics

Precision@125

How many predicted to convert to a customer became a customer.

Lift@125

How many times the conversion rate or precision was greater than the base conversion rate for the entire sample for the week.

PR-AUC

A metric used when the target is imbalanced where it measures the area under the precision/recall curve.

Model Choice

XGBoost classifier was chosen based on the imbalanced target, better explainability to the business, and its performance exceeded that of other models.

Feature Engineering

Market penetration

What percentage of the market Cintas penetrated by global/region/market by naics level 2, 4, or 6 depending on whether the business count met the required threshold.

Current Customer Proximity to Prospects

The haversine distance to the nearest customer. Also, the counts of customers within a distance radius to a prospect.

Guardrails

  • Time-based splits/Rolling validation windows
  • No “future” features leaking post-score signals
  • Added a 15-day buffer for the lag between sales and the data entering the CRM system

Evaluation

Trained on data through 12/31/2023; evaluated via a rolling monthly backtest across 2024, starting 4/1/2024, shifting the evaluation window forward every 30 days, and aggregating performance metrics across all windows.