Machine-learning research audit / 04
E-commerce Purchase Audit
Auditing the Evidence for Purchase Prediction
Research question
Do available customer, product and interaction attributes provide sufficient evidence for reliable purchase prediction?
Methodology
I compared logistic regression, random forest and gradient-boosting models with no-skill baselines, using cross-validation and an untouched test set.
Measurement and leakage
The audit identified post-outcome information leakage and limitations in how interactions and outcomes were measured. These findings changed the interpretation of what a modelling exercise could legitimately establish.
Negative result
The selected model achieved a test ROC–AUC of 0.468 and did not outperform a no-skill benchmark. The available evidence did not support deployment. The result is specific to this dataset and design; it is not a general claim that purchase behaviour cannot be predicted.
Next research step
I translated the audit into a proposed data-collection framework covering sessions, behavioural events, explicit purchase horizons and future time-based validation. The project demonstrates how negative findings can guide a more defensible research design.