Sales Pipeline Management with Machine Learning: A Lightweight Two-Layer Ensemble Classifier Framework

Haipeng Gao4 Jan 2022

Problem

A sales team has limited reach-out capacity and every day must pick which open opportunities to work on. Opportunities stay in the pipeline for days before they are won or lost, and new information arrives daily: days in the pipeline, stage, number of unsuccessful contacts. A classic binary classifier gives one propensity score (likelihood of winning) when the opportunity is created and cannot react to these changes. Working the list by that static score spends outreach on deals that have gone stale.

Example from the article: opportunity A scored 90% and B scored 60% at creation. Today A has been open for 2 weeks and B for 2 days. It is unclear whether to prioritize A for its higher score or de-prioritize it because it is old and harder to close.

Solution in brief

Split the model in two. A first-layer classifier scores an opportunity once, at creation, from static attributes. A second-layer classifier runs every day and combines that initial score with dynamic features such as duration, producing the propensity score for that day.

Questions this case answers
  • How to produce a daily updated win probability for items that stay in a funnel for days, when static scoring ignores time already spent
  • How to combine slowly changing and fast changing features in one prediction while keeping the daily model cheap and interpretable
  • How to prioritize the daily work-on list of sales opportunities when scores must change as a deal ages?
  • How to keep a propensity score current when features such as days in the pipeline change every day?

Tools and methods

  • Gradient boosting machine (GBM) — first layer, static attributes known at creation
  • Logistic regression — second layer, initial score plus dynamic features, readable coefficients
  • Cumulative gains curve — back-test against the classic single classifier, by days in pipeline
Continue reading — it's free

Google or a one-time email code. No password.