Congress Asks Airlines How Algorithmic Pricing Strategies Are Decided

What Happens to the Price if Nobody Looks at It?

A question about the review process admits three architectures, and an affirmative answer is true under all of them.

The distinction underneath is an old one. Engineers have argued about humans in, on, and out of the loop for decades, and nothing here improves on that vocabulary. What is new is that a congressional questionnaire and a federal appellate opinion now both turn on which rung a company occupies, and neither uses the word.

There are three rungs, and they are separated by one question: what happens if nobody acts. On the first, a person opens each recommendation, accepts or rejects it, and records why. Nothing publishes unless somebody acts. This is the only rung on which a human decides a price.

The Third Circuit has already tested what that authority is worth. It was undisputed in that case that the casino-hotels retained final pricing authority and could change any suggested price. The claims were reinstated anyway, on the reasoning that prices are fixed when they are agreed upon, whether or not the parties always adhere to them. Keeping the power to say no did not answer the question.

On the second rung, recommendations publish by default and a person may intervene. The ability exists and may go unused. Prices publish. On the third, nothing human happens at any point. Prices publish.

Underneath all three sits a configuration layer that is not a rung at all. Before the model runs, somebody sets the floor, the ceiling, how far a price may move in a day, and which competitors count as the comparison set. Those choices shape every price the system produces, and a company can make them once and never look again. All three rungs support the sentence "we review our prices."

The Justice Department drew the same line in its amended complaint against RealPage, a rental-housing pricing vendor, describing an auto-acceptanceA setting under which a recommended price publishes on its own unless a person intervenes before a deadline. feature whose parameters defaulted to three percent daily and eight percent weekly, and a rejection path that required specific business commentary routed to a vendor advisor who could escalate the matter to the landlord's regional manager, the person who supervises the property manager. On that account, accepting cost nothing and refusing cost a memo, a review, and a conversation with your own boss. RealPage itself recognizes, the complaint says, that acceptance rates are not necessarily the best measure of its influence.

Pallone drew the same ladder without naming it. Question 8(c) asks about review after the fact. Question 8(b) asks whether the model weighs data elements differently, and if so which ones and how. Question 8(e) asks what parameters regulate the scope and scale of the algorithm, and gives as its example whether there is an upward limit to how high a price may be set. That is a ceiling, and a ceiling is the configuration layer. Question 8(f) asks who decides whether a price takes effect, and offers "automatically" as an available answer.

One number shows why the distinction is not academic, and it means less than it appears to. Both appellate opinions mention a figure of ninety percent. Both are reciting the same allegation: that the vendor said its clients charge the recommended rate that often. The Third Circuit recorded it as an allegation and treated it as supporting an inference of agreement at the pleading stage. The Ninth Circuit, in a case where the horizontal claim had been abandoned on appeal, noted that the precise rate at which the hotels actually implemented the recommendations was not in the record.

The same alleged marketing claim, in two courts, put to two uses. And an acceptance rate is not a fact about behavior in any case. It is a function of the tolerance bandThe width of the gap from a recommended price that still counts as accepting it, chosen before any acceptance rate can be measured. chosen to define acceptance. The Justice Department's own filing shows the range: measured within one percent of the recommendation, the rate was forty to fifty percent; within two and a half percent, nearly sixty; within five percent, more than eighty-five. Same decisions, three numbers.

The version of algorithmic pricing most people can actually see works differently, and the evidence is thinner than the attention. Electronic shelf labels have drawn state bills and Senate letters on the theory that a store can raise a price the moment demand spikes. The only empirical study located for this piece, an unpublished working paper, finds no such effect. Temporary price increases were running at 0.0042 percent of products on an average store-date before the labels arrived, and the change afterward was six ten-thousandths of a percentage point, which is not statistically distinguishable from zero. A shelf label is a broadcast device. It shows one price to everyone in the aisle and cannot price a person.

The regulator has been working the personal side instead, as disclosure rather than prohibition. On August 19 the Federal Trade Commission proposed for comment a policy statement on personalized pricing, with a worked example of a grocery chain charging a delivery customer more for milk because data showed children in the household. The Commission says in the same document that it lacks authority to ban the practice in all circumstances. Four months earlier it had issued an advance notice of proposed rulemaking on food delivery fees, asking among other things about failing to disclose whether prices differ from those offered to other consumers on the platform.