Elasticity estimates and the confidence they deserve
Price elasticity describes how much quantity responds to a change in price. It is estimated from historical variation, it is expressed as a single number, and it is one of the most confidently misused figures in commercial analysis.
The number is real. The confidence usually attached to it is not.
Where does the estimate come from?
From periods in which price varied and the response was observed, with a model separating price effects from everything else. The quality of the estimate depends entirely on how much genuine price variation exists and how well the other drivers were controlled.
A category with little price variation offers little to estimate from, and a model fitted to it will produce a number anyway.
What corrupts an elasticity estimate?
Unobserved promotional support, most commonly. A price change accompanied by display produces a response that the model attributes to price, which overstates sensitivity and then misleads every subsequent pricing decision.
Distribution changes, competitor moves and seasonality do the same in different directions. Each one that goes unmodelled is absorbed into the price coefficient.
Why is the range so wide?
Because the effect is estimated from a limited number of price changes, many of which were not independent of other activity. Confidence intervals on retail elasticities are typically wide enough that the commercially relevant decisions sit inside them.
That is the honest state of the art and it is rarely presented. An elasticity quoted to two decimal places implies a precision the estimation cannot support.
Does elasticity vary?
Considerably — by item, by store type, by price level, by season, and by whether the change is an increase or a decrease. A single portfolio elasticity applied everywhere assumes away all of that.
Asymmetry between increases and decreases is particularly important commercially and particularly hard to estimate, because historical data contains far more decreases.
How should it be used?
As a directional input with its range, inside a decision that also accounts for margin, competitor response and distribution. A price decision made on a point elasticity is a decision made on the midpoint of a wide distribution.
Scenario analysis across the range is more useful than a single calculation. If the decision flips within the confidence interval, the analysis has not settled it.
What about cross-elasticity?
Even harder to estimate and even more useful when done well, because it describes substitution within the set. The same caveats apply with more force, since cross-effects are smaller and therefore noisier.
What should be asked of any supplied elasticity?
The estimation window, the price variation available in it, what was controlled for, and the confidence interval. Four questions, and a provider who answers them clearly is giving you a number you can use.
Frequently asked questions
Can I estimate elasticity from my own data?
If you have sufficient genuine price variation and can control for support, yes, and the harder part is usually the control rather than the estimation. Without observed promotional support, the estimate will be biased and the bias will not be visible in the output.
How much price variation is enough?
Enough that the estimate is not driven by a handful of events. There is no fixed threshold, and the practical test is whether the estimate is stable when individual periods are removed, which is a straightforward check that is rarely run.
Does a high elasticity mean I should not raise price?
It means volume is more responsive, which is an input rather than an answer. The decision depends on margin: a high-margin item can lose more volume and still improve profit, and the calculation should be done rather than inferred.
Is elasticity stable over time?
Not especially. It shifts with competitive structure, with consumer conditions and with the category's own promotional intensity, so an estimate several years old should be treated as a prior rather than as a fact.
What about very low-priced items?
Rounding and price-point effects become significant at low absolute prices, where shoppers respond to crossing a threshold rather than to a proportional change. A continuous elasticity model fits that behaviour poorly.
Should elasticity differ by channel?
Frequently it does, because shopper missions and competitive alternatives differ. Applying a large-format estimate to independent retail is a common shortcut and an unsupported one.
How do I present an elasticity to a commercial audience?
With a range and a scenario table showing the volume and profit outcome across it. That format invites the right conversation about risk and avoids the false precision of a single figure.
Can machine learning fix the estimation problem?
It can fit more flexible relationships and it does not create price variation that never happened, nor does it observe support that was never recorded. The limits here are about the data rather than the method.
What is the safest way to use one?
As a boundary check on a decision rather than as its basis. If the decision holds across the plausible range, the elasticity has done its job; if it depends on the exact value, more evidence is needed.
### What should a brand build instead of a single number?
A scenario table: volume, revenue and profit at several price points across the plausible elasticity range. It takes the same inputs and it presents them in the form the decision actually needs.
The table also makes the asymmetry visible. A price move that is modestly good at the central estimate and badly damaging at the pessimistic end is a different proposition from one that holds up throughout.
Interested in NRS Insights data? See case studies or contact NRS Insights.