The most useful number in retail personalization this month was 0.2382. It is the Precision@5 of Walmart’s CASE model on Instacart with its temporal convolutional network removed; the full model returns 0.3989. The single component that captures when a user buys carries roughly forty percent of the discriminative weight. Most fashion retailers’ recommenders do not have that component at all.

CASE (Cadence-Aware Set Encoding) argues that for replenishment categories, calendar time is the dominant signal and most production recommenders are encoding the wrong thing. The paper’s framing is exact. Visit-order architectures — the BERT-flavored transformers that have become the industry default — treat baskets on days 1, 8, and 36 identically to baskets on days 1, 2, and 3. They know the order; they cannot see the gap. The gap is what matters. A weekly milk run and a quarterly fragrance refill leave the same trace under visit-order; only the calendar separates them.

Walmart Global Tech deployed CASE against the production stack and reported lifts of 8.63 percent on Precision@5, 9.90 percent on Recall@5, and 10.46 percent on NDCG@5 across a user base in the tens of millions. Those numbers measure what a deployed retailer was paying for a model that could not read a calendar; every retailer still running a visit-order recommender is paying it now.

The paper studies grocery, and the arithmetic carries to any vertical with an item that runs out. That covers more of the wardrobe than buyers admit. Mascara, cleansers, fragrance refills, hosiery, basics, sock multipacks — replenishment categories with cadences as stable as Walmart’s milk and detergent. A customer who buys a Charlotte Tilbury mascara every nine weeks is generating the strongest signal of next-basket intent. The collaborative-filtering model serves her a new lipstick because someone with similar embeddings recently bought one.

The personalization industry has spent a decade selling fashion retailers cross-sell and discovery models. Cross-sell is the part of the basket where the data is sparsest and the margins are lowest. Replenishment runs the other way: dense history, clean cadence, conversion that costs almost nothing to obtain. The replenishment customer is also the customer with the lowest acquisition cost; she is already buying, already on a clock. Walmart’s lift numbers are what that mismatch costs in the categories most retailers treat as beneath analytics.

The rebuke in the CASE paper is precise. The dominant component of next-basket performance is not the item embedding, not the cross-item attention, not the user tower. It is the calendar. Removing the temporal convolutions destroyed more than removing any other module. The basket has a clock, and most retailers are running models that cannot tell the time.

Fashion’s recommenders were built for a wardrobe that does not exist.

The customer who shops most often is the one the discovery model knows least. The margin sitting in her basket is not a research problem; it has been on the receipts for years. The question is which retailer reads them first.