Understanding Propensity Modeling: A Non-Technical Guide

If you've ever received a product recommendation that felt uncannily accurate, or a retention offer that arrived at exactly the right moment, you've experienced propensity modeling in action.

What Is Propensity Modeling?

At its core, a propensity model answers one question: *How likely is this person to do this thing?*

• How likely is this browser to make a purchase in the next 30 days?
• How likely is this subscriber to cancel their subscription?
• How likely is this prospect to respond to a specific offer?
• How likely is this customer to add a second product line?

The output is a score — typically between 0 and 1 (or 0-100) — that represents probability. A purchase propensity score of 0.72 means the model estimates a 72% chance that this person will buy within the specified time window.

The Inputs That Matter Most

Propensity models are only as good as their training data. The most predictive models typically combine:

1. Behavioral signals (most predictive): Website visits, app sessions, email opens, content consumption, search history. *These capture active intent in real time.*

2. Transactional history: Past purchases, average order value, return rate, purchase frequency, product category preferences. *These reveal patterns and preferences.*

3. Demographic and firmographic data: Age, income, location, job title, company size, industry. *These provide context but are rarely sufficient alone.*

4. Third-party intent signals: Job change events, funding announcements, technology installs, content consumption on external sites. *These catch signals before they reach your own properties.*

5. Engagement recency and frequency: How recently and how often someone has interacted with your brand. *The most reliable single predictor in most models.*

A Practical Example

Imagine you're a B2B SaaS company selling to mid-market technology firms. Your propensity model for "will purchase Enterprise tier within 60 days" might weigh signals like this:

• 40% weight: Product usage patterns (active users, feature adoption, API call volume)
• 25% weight: Engagement signals (support ticket frequency, NPS score, CSM meeting attendance)
• 15% weight: Third-party intent (recent funding round, job postings for relevant roles, tech stack changes)
• 10% weight: Firmographic fit (company size, industry, revenue band)
• 10% weight: Contract signals (days until renewal, current plan tier, expansion history)

How to Get Started

1. Start with a clear use case. The best first project is one with a measurable outcome and a short feedback loop — like predicting 30-day conversion on a high-traffic landing page.

2. Audit your data. Propensity models need outcome data to train on. Do you have clean, labeled historical data showing who converted and who didn't?

3. Begin with 5-10 signals. Don't boil the ocean. Start with the signals your team already trusts, validate the model, then expand.

4. Run a holdout experiment. Split your audience into a model-guided group and a control group. Measure the actual lift — did the model-guided group convert at a higher rate?

5. Retrain regularly. Consumer behavior shifts. Models trained on Q1 data may degrade by Q3. Schedule quarterly retraining and monitor performance drift monthly.

A well-built propensity model doesn't replace marketing judgment — it sharpens it. You still decide the strategy, the creative, the offer. The model just tells you who's most likely to care.