A Machine Learning Framework for Predicting First-Hour Sales Surge from E-Commerce Metadata
Abstract
Relating product metadata to immediate post-launch behavior is crucial for enabling forward-looking decisions for e-commerce platforms including inventory management, pricing, and promotional efforts. Despite much attention to sales forecasting, existing methods primarily utilize post-launch behavioral signals such as clickstream and transaction data that are intrinsically not available before the actual product is launched, limiting their suitability for early-stage predictions. We present a supervised machine learning approach for predicting first hour sales spike events solely based on pre-launch product metadata. We build a dataset of 5,000 product listings using information from publicly accessible e-commerce sites and create features including price, brand, category, ratings, review counts, price tier and product visibility signals on the platform. Post-launch behavior variables are strictly forbidden to be in the feature space for prediction and only used to create ground truth labels. The proposed approach uses automated pre-processing and addresses the class imbalance problem with SMOTE, compares several classification models, and finds that ensemble based methods provide highly robust and successful performance with F1-scores of above 0.95 and ROC-AUCs above 0.97, showing that pre-launch metadata alone can carry strong predictive signal for first-hour sales spikes, without relying on any post-launch behavior.