A Promotion Spike Is Not a Demand Curve

The only price data most companies have

Most companies have one source of evidence on how customers respond to price: their promotions. Prices sit still for months, then drop 20% or 30% for a week, and volume jumps. It is the only time the data shows a large price move beside a large volume move.

So when a permanent pricing decision comes up, such as passing on a supplier cost increase, the promotional history gets reused. Someone divides the lift by the discount and calls the result an elasticity. The number is large, it comes from real transactions, and it looks rigorous.

It answers a different question. A promotion measures the response to a temporary event, while a permanent price change asks about a new normal. Treating one as the other biases pricing in a consistent direction: too timid on increases, too eager on cuts.

J.C. Penney ran this experiment at full scale. On February 1, 2012, it replaced nearly 600 sales a year and most coupons with a simpler scheme: everyday prices about 40% below the old list prices, plus a few month-long sales. If the old promotional spikes had measured price sensitivity, permanently lower prices should have held the volume.

They didn't. Comparable store sales fell 21.7% in the second quarter, and store traffic fell about 12% even though conversion rose. Other changes were under way at the same time, but the pattern is hard to miss: customers had been responding to the events, not to the price level.

What a promotion actually measures

A promotional spike has 4 sources, and only 1 of them says anything about a permanent price.

• Timing. Customers who would have bought anyway buy earlier, or buy more and stock up. The weeks after the promotion dip. A permanent price change gives nobody a reason to shift timing.
• Switching. Customers move from your other products to the promoted one. Units move between lines, but the business has not gained a customer.
• The event. Display, advertising, a deadline, the word "sale". J.C. Penney showed how much of the response this piece can carry.
• Price. The genuine response to a lower price level. This is the only component a permanent pricing decision needs.

Timing and switching can be measured from your own data: look at the weeks after the promotion, and at sibling products during it. The event and the price effect usually cannot be separated, because they always arrive together. That limit turns out to be useful, as the rule below shows.

A worked example

In this example, the raw promotional spike says hold price, and the evidence that actually applies says raise it. The figures are illustrative.

A product sells 1,000 units a week at $10.00, with a unit cost of $7.00. A 2-week promotion at 20% off lifts sales to 1,800 units a week, or 1,600 extra units in total. Divide the 80% lift by the 20% cut and you get an elasticity of −4.0.

Now suppose you could see where the 1,600 units came from:

The decision on the table is a permanent 5% increase, from $10.00 to $10.50. The break-even volume loss is the share of units you can lose before the increase stops paying, the method set out in Elasticity Without Econometrics:

Here is the same decision under 3 readings of the promotion:

The raw spike predicts that the increase cuts weekly margin from $3,000 to $2,800. The price response alone predicts a rise to about $3,240. Same product, same data, opposite decisions.

The Ceiling Rule

Once timing and switching are removed, what remains is a ceiling on permanent price sensitivity, not an estimate of it. You still cannot separate the event from the price, so the true figure is somewhere below.

A ceiling is only useful in one direction, so use it that way:

• If even the ceiling implies a loss below break-even, the increase is safe on the evidence you have. In the example, the ceiling of −1.75 predicts 8.75% against a 14.3% break-even. Raise.
• If the ceiling implies a loss above break-even, the promotional data cannot answer the question. It does not say hold price. It says you need different evidence.
• For price cuts, the logic reverses. Promotional data overstates the volume a permanent cut will buy. A cut that only pays at the promotional elasticity probably does not pay.

This turns a number nobody can defend into a test people can agree on. No one has to claim to know the elasticity. They only have to accept that the true response is no larger than the ceiling.

Why managers get this wrong

The mistake is rarely careless. The promotional elasticity is the most available number, and several things make it look more trustworthy than it is.

• It is the only data with large price moves. Regular prices barely change, so promotions feel like the only option.
• Promoted products are not a random sample. They are chosen because they are expected to respond. Even a clean estimate describes the most responsive items, not the range.
• Precision reads as accuracy. A regression reports an elasticity to 2 decimal places, whatever it is actually measuring.
• The dip hides in the noise. Pulled-forward demand comes back out spread thinly over several weeks, inside normal week-to-week variation. Nobody sees it unless they look for it.
• Estimating got cheap. With an AI assistant, an analyst can fit 50 elasticities in an afternoon. Knowing which ones the data can support takes the same judgment it always did.

Economists have measured the size of the timing problem alone. In a study of storable household goods, Igal Hendel and Aviv Nevo found that models ignoring stockpiling overestimated price sensitivity by 30%. They also overestimated how many customers would stop buying altogether by more than 200%. That second figure is exactly the fear that stalls a price increase.

If not promotions, then what

The better evidence is regular price changes: moves in the everyday price that customers read as permanent. They are rarer and smaller, and they only mean something against a comparison group.

The standard design is a difference-in-differences. Compare items whose regular price changed with similar items whose price did not, before and after the change. The comparison group absorbs seasonality and trend, so what remains should be the price effect.

This design has a quiet failure mode, and a placebo test is how you catch it. Run the same comparison on a period when nothing changed. If it "finds" a price effect anyway, the two groups were drifting apart for other reasons, and your real estimate carries that drift too.

A common way it fails: a business reprices most of its range at once. The items left unrepriced, which become the comparison group, are the ones nobody bothered to reprice, often slow sellers on their own trajectory. The comparison then mixes that trajectory into the price effect.

When the placebo fails, the honest response is usually not a more sophisticated model. It is a different question. Go back to the break-even and ask whether the volume loss you can tolerate is plausible, which needs judgment rather than an estimate.

The halo check

Promotional data also flatters promotions themselves, through the "halo": the claim that promoted items pull full-price items into the basket. Check how those full-price items were actually priced before you credit it.

Promotions often run alongside a basket-wide discount, such as a member event, a sitewide coupon or a store-wide sale. When they do, the items riding along were discounted too. Their lift belongs to the other discount, not to the promotion.

The fix is to classify every discounted line by the mechanism that produced it: the coupon, the offer code, the discount field. Do not infer it from a proxy such as discount depth. A proxy that is right most of the time is wrong exactly on the lines that matter. Once each line carries its real mechanism, the halo is usually much smaller, and sometimes it disappears.

Know the limits

The Ceiling Rule holds in most categories, but not everywhere.

• Increases and cuts are not symmetric. Customers notice a price rise more than a price cut of the same size. On products they watch closely, the reaction to a visible increase can exceed anything a promotion suggested, so the ceiling is weakest exactly there.
• Some products cannot be stockpiled. For perishables and services, the timing component is small and promotional lift sits closer to the real price response.
• Event effects vary. A quiet promotion carries little event lift and a heavily advertised one carries a lot. The ceiling is loosest after the loudest events.
• Competitors respond to permanent moves. A one-week promotion rarely draws a reaction, while a permanent price change can. Neither number includes it.
• Straight-line extrapolation is an approximation. The example applies an elasticity measured on a 20% cut to a 5% increase. For small moves that is usually close enough; for large ones it is not.

None of this makes promotional data useless. It is the right evidence for promotional decisions, and whether a promotion pays is answered directly by its break-even lift, covered in The Promotion That Grew Sales and Shrank Profit.

The principle

Match the evidence to the decision. Promotional data tells you what happens when you run a promotion. It cannot tell you what happens when you change the price.

Sources

• J.C. Penney Company, Inc., Form 10-Q for the quarter ended July 28, 2012.
• Igal Hendel and Aviv Nevo, "Measuring the Implications of Sales and Consumer Inventory Behavior", Econometrica 74(6), 2006.

Related analysis and tools