Learning demand from prices without mistaking correlation for cause

When prices already contain information

How much would sales change if a retailer changed a price? Historical records do not necessarily answer that question directly. Prices reflect decisions, and those decisions may incorporate information about demand that a researcher never observes.

In Causal Multi-Task Demand Learning, Varun Gupta and Vijay Kamble illustrate the problem with stores whose managers know their local demand. Stores with stronger demand charge higher prices and sell more. Pooling their records can suggest that higher prices increase sales, even though every store’s actual demand curve slopes downward.

This article uses the paper’s third arXiv version, revised on 7 November 2026. Its arXiv record reports acceptance at NeurIPS 2026, although the supplied PDF is labelled “Preprint.”

Learning separately for each store is not an easy alternative. A store or product may have been offered at only two or three prices. Sales fluctuations can then make its estimated price response extremely noisy, particularly when the observed prices are close together.

The researchers investigate how to borrow information across many such contexts while avoiding misleading comparisons between them.

Keep the prices, hide selected outcomes

Their method assumes that demand within each context follows a straight line: an intercept sets its level, a slope describes its response to price, and random shocks add noise. Contexts can have different intercepts and slopes.

The essential requirement is at least two distinct, locally exogenous prices per context. Roughly, the demand shock at either selected observation must still average to zero after accounting for the full price history, context characteristics and other demand observations. This can hold when prices are committed before demand shocks occur, even if managers choose those prices using private knowledge about underlying demand.

The proposed method, Decision-Conditioned Masked-Outcome Meta-Learning, or DCMOML, gives the learner all historical prices, available context characteristics and any demand observations outside the selected pair. It hides both selected demand outcomes from the inputs used to predict the demand parameters.

Those hidden outcomes still supervise training. The theoretical construction randomly chooses one of the two as the prediction target without telling the parameter-predicting learner which was chosen. In the experiments, the authors instead average the prediction losses at both prices, which has the same population objective.

Why this arrangement? Including all prices lets the model use information encoded in previous pricing decisions. But supervising at just one known price cannot uniquely determine both a line’s intercept and slope: many lines pass through the same point. Requiring the same predicted line to work at two distinct prices resolves that ambiguity. Hiding both outcomes prevents the missing-outcome pattern from revealing which price was selected for supervision.

What the guarantee actually establishes

Under the stated assumptions, the training objective uniquely selects the conditional mean of the causal demand parameters given the permitted inputs. This is an average over underlying demand curves compatible with that information—not exact recovery of every individual store’s curve.

The authors also prove a worst-case boundary for this particular objective. Revealing additional information about the hidden outcomes or selected query can make it fail for some otherwise admissible data-generating process. This does not mean extra information is always harmful, or that other methods cannot use it under stronger assumptions.

Simulations show a benefit—and a cost

The synthetic experiments give managers noisy estimates of revenue-maximizing prices, around which they make small price variations. The researchers vary how strongly prices reflect hidden demand parameters and test contexts with two to five observations.

With strong confounding and two observations, DCMOML’s mean squared slope error was 0.0077, compared with 0.0783 for standard support–query meta-learning. Each configuration used 20,000 training contexts and 10,000 new test contexts, repeated across ten random seeds.

The advantage was not universal. With weak confounding, standard meta-learning estimated intercepts more accurately at every tested observation count. Masking outcomes sacrifices useful information; when confounding is less consequential, that sacrifice can outweigh the benefit.

Retail prediction is not causal validation

The retail experiments used transaction records from a UK gift retailer. Models learned from two observations per product and predicted a third withheld observation. One task withheld a less frequently used price; another withheld a later pricing episode.

For the later-episode task, DCMOML’s exposure-weighted root mean squared prediction error was 35.08, compared with 39.26 for standard meta-learning. But its fitted slopes varied substantially across training seeds and were not interpretable as causal price responses.

Neither retail task verified local exogeneity. Transaction prices could vary with customer type and order size, and days without sales were absent from the records.

The theoretical guarantees also depend on linear demand and a suitable price-setting process. Closely spaced prices remain weak evidence about slopes, and flexible learning requires many contexts. Because the target conditions on historical prices, changing the pricing policy may require refitting. The contribution is a conditional way to improve learning from sparse, confounded histories—not proof that observational retail data reveal the consequences of changing prices.