Problem
A delivery marketplace needs the right restaurants in every market, but the sales team can only approach a limited number of off-platform merchants. The company needs a value estimate per prospect to rank leads and to size the addressable market. New merchants ramp up slowly, so early sales say little about their true potential, and sales cycles are long.
Example from the article: DoorDash predicts, for each off-platform merchant, its success value and its rank within the local market, and the operations team uses both to prioritize sales leads.
Solution in brief
Train gradient-boosted regression models on merchants already on the platform. The label is sales over a fixed window after the merchant's activation has stabilized, and a separate model is trained per merchant segment. Predictions are refreshed daily and explained with Shapley values. Quality is checked with two decile-based metrics shaped around how sales allocates leads.
Questions this case answers
- How to rank prospective supply-side partners by expected value when their outcomes arrive only after a long onboarding period and training data is small
- How to evaluate a ranking and value-prediction model in business terms when the real outcome is delayed
- How to predict the value of merchants that are not on the platform yet, so sales can prioritize who to sign?
- How to define a fair success label for new merchants that ramp up at different speeds?
Tools and methods
- XGBoost — final regression model, one per merchant segment
- LightGBM — compared with linear and tree-based regressors during model selection
- SHAP (Shapley values) — feature importance with direction of impact, used for business explanations and model monitoring
- MLflow on Databricks — model management, weekly or monthly retraining, registered models pulled for daily prediction
- Feature store and Snowflake — daily merchant features and prediction-time feature processing
Google or a one-time email code. No password.