Developing an AI-based lead scoring system based on purchase likelihood
Purchasing likelihood scoring is a more accurate task than simply scoring based on ICP. The model predicts the likelihood of a specific purchase conversion based on behavioral signals and historical patterns.
Prediction model
Architecture: XGBoost / LightGBM gradient boosting for tabular features. For sequential data (dynamic behavior), we add an RNN/LSTM component.
Feature groups:
Demographics: job title, seniority, company size, revenue, industry, geography.
Behavioral: Pricing pages visited, case studies open, demo scheduled (strongest signal), email open rate, response time.
Temporal: time from first contact, rate of progression through stages, seasonality.
Historical: similar profiles in the past - with what outcome.
Minimum dataset: 500+ closed leads (won + lost) with activity history.
Model assessment
The calibration plot is key for predicting probabilities: if the model predicts 70%, there should be a conversion in 70% of cases. We use the Brier Score as the primary metric along with the AUC-ROC.
Model update
Weekly retraining on accumulated data. Concept drift monitoring (Population Stability Index) – as the market changes, the model degrades.