Direct-to-Consumer Brand
Model how likely customers are to purchase, given data across multiple engagement channels. Uncover pain points of competing products in the marketplace, and understand latent needs of consumers to guide product development.
Challenge
Outdoor furniture is not an impulse purchase. Customers research extensively over weeks or months, interact across multiple devices and channels, and may not be recognized as the same person across touchpoints.
- Can high-purchase-probability customers be identified early?
- Can probabilistic household matching be established across devices?
- What pain points appear in competitor reviews?
- What unmet consumer needs exist in available data?
Solution
We built a machine learning pipeline using XGBoost combined with geolocation-linked census data and SHAP interpretability to score purchase probability. We employed zero-shot topic modeling to filter 600+ GB of text review data, then used language models for product description ideation and need discovery.
Results
One top predictive variable was actionable in production, driving immediate lift in conversions before full deployment. The analysis identified both pain points and desired product features to inform design and roadmap decisions.