E-Commerce Customer Spending Predictor

Key Result

R²=97.8% · predicts within $10.48 of actual yearly spend

The Challenge

What needed solving

NovaBuy had 500 customers and no way to predict how much any individual would spend in the next year. Marketing budgets were allocated equally across all customers — high-value long-term members received the same spend as brand-new signups. Without a spending model, the team couldn't prioritise retention campaigns, personalise offers, or identify which behavioural signals actually drove revenue.

The Solution

How we solved it

I built a full ML pipeline in scikit-learn: four regression models (OLS, Ridge, Lasso, ElasticNet) trained with GridSearchCV hyperparameter tuning and 5-fold cross-validation. A performance threshold gate (R²≥0.95, RMSE≤$15) ensures only production-quality models are deployed. Feature analysis revealed that mobile app engagement and membership length are the dominant revenue drivers — website time adds near-zero predictive value (coefficient $0.31, confirmed by Lasso zeroing it out). The model is served via a FastAPI endpoint and visualised in an interactive Streamlit dashboard with live sliders, batch CSV upload, and a confidence interval on every prediction.

97.8%

R² score on held-out test set

$10.48

RMSE — average prediction error

1.79%

MAPE — mean absolute % error

4

Models compared, best selected automatically