Picnic developed CARP as a specialized XGBoost-based machine learning model designed specifically for predicting repeat purchases in grocery retail. The model leverages handcrafted features tailored to capture grocery shopping patterns, including average rebuying frequency of items and customer shopping periodicity. Unlike complex sequential models that attempt to model entire purchase sequences, CARP focuses exclusively on the repeat purchase problem where abundant historical data exists.

The system operates by analyzing each customer's purchase history and ranking previously bought items by their likelihood of being repurchased. CARP powers multiple touchpoints within the Picnic app, most notably the previous purchases page where customers can quickly rebuild their shopping lists. The model processes millions of recommendations daily, serving as the backbone of Picnic's personalized shopping experience.

Performance analysis reveals CARP's superiority over state-of-the-art sequential recommendation models including RNNs, Transformers, and Graph Neural Networks. CARP achieves 25.2% improvement in HitRate@1, 24.0% improvement in Precision@10, and 23.5% improvement in MAP@30 compared to sequential model averages. The model demonstrates statistical significance with p-value of 0.000943 and large effect size, confirming its practical superiority for repeat purchase prediction in grocery retail contexts.