Price elasticity in FMCG: weekly SKU-level models

September 13, 2026 · 5 min · Free, with full transcript

Price elasticity has been calculated in consumer goods since the 1980s, so why does it matter again now? The answer given here is cadence and granularity: quarterly, category-level estimates are historical artifacts, while pooled models that borrow signal across similar products and stores can produce weekly estimates across roughly 40,000 SKUs per category. Vendor throughput claims from OpenAI and Google AI are worth scrutiny, since O'Reilly Radar engineering write-ups put real production costs higher once data plumbing is included. You leave knowing how to separate price cuts from display promotions in point of sale data, model the post-promotion dip that MIT Technology Review flagged, validate against holdout stores, and bring SKU-level numbers into trade negotiation with retailers.

Listen to the podcast

5 min

Chapters

Key takeaways

  • Pool data across similar products and stores so slow-moving SKUs inherit patterns instead of producing noise.
  • Move elasticity refreshes from every 13 weeks to weekly so numbers survive contact with a live trade negotiation.
  • Clean your point of sale feeds so a price cut can be separated from a display promotion sitting on top of it.
  • Model the post-promotion dip explicitly and validate against holdout stores where you deliberately skip the price change.
  • Treat a wide confidence range as the model admitting it does not know, and override it with human judgment.
Read the full transcript

Host:Welcome to the Leaders Insights podcast. Today's episode, Building AI Elasticity Models for FMCG Assortment and Price Optimization. Let's start with the thing everyone in consumer goods is whispering about this week. Price elasticity models have existed since the 1980s. Why is this suddenly a hot story in September 2026?

Expert:Because the old models were built to answer one question a quarter. And the business now needs answers on thousands of products every week. Elasticity. How much sales move when you change the price? Was always calculated at a level so broad it was basically useless. You'd get detergent is price sensitive and then have to guess which pack size, which retailer, which region.

Host:But the math hasn't changed. A price coefficient is a price coefficient? What's actually new?

Expert:The math hasn't. The scale has. The old regressions choke on sparse data. Most individual products barely sell enough units in a given store week to estimate anything reliable. What's new is pooling. You borrow signal across similar products and stores. So a slow moving SQ. Stock keeping unit, one specific product and size, C, inherits patterns from its neighbors. That's the difference between an estimate you'd stake a negotiation on and noise.

Host:The article anchors this on a specific number.

Expert:The playbook sites teams moving from elasticity refreshes every 13 weeks down to weekly across roughly 40,000 SQ's per category. That cadence is the whole point. A quarterly number is a historical artifact by the time you use it. A weekly one survives contact with an actual trade negotiation.

Host:You mentioned vendors are loud about this. Whose figures should I distrust?

Expert:Distrust is strong, but scrutinize. Yes. OpenAI and Google AI have both published throughput claims about running these models cheaply at scale. Very impressive per inference costs. Worth remembering they sell the compute those models run on, so the economics are conveniently rosy. Cross check against O'Reilly radar's engineering write-ups, which put real production costs higher once you include the data plumbing nobody demos.

Host:The data plumbing. That's the part that always gets skipped in the sales deck.

Expert:It's 80% of the job. Your point of sale data, the scanner records from checkout, arrives late in different formats per retailer with promotions logged in consistently. If you can't cleanly separate a price cut from a display promotion sitting on top of it, your model attributes the sales lift to the wrong lever. Garbage estimate, confident decimal places.

Host:Give me a concrete failure. Where does this go wrong in the real world?

Expert:Classic one. A brand sees demand jump when they discount. Model says very elastic, cut price permanently. But the jump was pantry loading. Shoppers stockpiling because it's cheap, then buying nothing for six weeks. The model saw a spike, missed the crater. MIT Technology Review flagged exactly this pattern earlier in 2026. Models that don't account for forward buying overstate elasticity and quietly destroy margin.

Host:So how do you catch that?

Expert:You model the post-promotion dip explicitly, and you validate against holdout stores. Locations where you deliberately don't run the price change, so you have a real comparison. If your model can't predict what happened in stores, it never saw. It's memorizing, not learning.

Host:Let's talk power. This connects to trade negotiation. How does a better elasticity number change what happens in the room with a retailer?

Expert:Enormously, because retailers have had the better data for years. The retailer knows your true elasticity from their own scanners. When a brand walks in with a course quarterly estimate, they're negotiating half-blind. A weekly, skew-level model means you can say, this promotion depth loses us both money. Here's the depth that grows the category. And back it with the same granularity they have.

Host:Isn't there a risk you over-trust the model and stop using judgment?

Expert:Yes. Discipline teams treat outputs as a decision input, not a verdict. The model tells you elasticity with a confidence range. If that range is wide, that's the model admitting it doesn't know. And that's when a human should override.

Host:Give me the one thing a listener can do Monday morning.

Expert:Pull your last four promotions and check whether anyone measured the post-promotion dip. If nobody has that number, every elasticity figure in your building is inflated, and you're leaving margin on the table before you touch a single model.

Read and go deeper