Retail Promotion Decision Engine

−$481contribution margin from the revenue-optimal plan

29.1%of rows under strict propensity overlap

0actions surviving the downside-margin screen

The question

Which products should a retailer promote, at what discount, and in which stores, when the objective is incremental contribution margin rather than unit lift or revenue alone?

The data

University of Chicago Booth Kilts Center Dominick's Finer Foods scanner data, built into validated store-product-week panels with explicit checks on keys, prices, quantities, source quality, and bundle-price semantics. Promotion coding is incomplete, historical pricing was not randomized, and stockouts are not cleanly observed — limits that shape every conclusion below.

How it was done

A validated DuckDB mart establishes store-product-week grain and provenance.
Fixed-effects and hierarchical Bayesian models estimate conditional price response with temporal holdouts and partial pooling.
A doubly robust promotion analysis faces overlap, balance, placebo, and negative-control tests before it can influence optimization.
A mixed-integer optimizer compares revenue and contribution-margin scenarios under capacity and downside-risk constraints.
Release gates block recommendations when causal identification, calibration, monitoring, or economic inputs are inadequate.

What the evidence showed

Evidence and what each result implies for the decision

Evidence	Result	Decision implication
Bayesian holdout uncertainty	81.9% coverage for a nominal 90% interval	Model is overconfident
Split-conformal recalibration	85.4% coverage on a later window	Improvement, still below target
Historical shadow replay	Twelve of thirteen weeks trigger at least one block	Aggregate performance hides unstable weeks
Strict propensity overlap	29.1% of rows	Historical promotion effect is not decision-grade
Post-weighting balance	Maximum absolute SMD 0.224	Residual imbalance remains material
Revenue scenario	+$1,159 revenue and −$481 contribution margin	Unit and revenue lift do not imply profitability
Downside-screened margin scenario	Zero actions at all tested capacities	No promotion is the robust base-case decision

What I recommend

Do not deploy the observational promotion estimate or use it to automate promotion selection. The large adjusted effect is withheld because overlap, balance, a future-treatment placebo, and a past-outcome negative control all fail. Numerical convergence does not override failed identification.
Keep the engine as a refusal mechanism and experiment-design tool: it names the missing evidence, stops an invalid causal estimate from reaching the optimizer, and keeps "no action" on the table when margin risk is unfavorable.
Run a stratified randomized promotion test next — explicit discount depths, stratified by store, product and baseline velocity, with unit sales and incremental contribution margin pre-registered as co-primary outcomes. Capture display and feature status, offered price on zero-sales weeks, inventory and stockouts, supplier funding, and current replacement cost.
Supplier funding is the commercial lever to negotiate: no candidate clears the downside screen unfunded, while $0.10 per promoted unit allows 12 modeled actions with roughly $14 aggregate funded margin. A planning calculation for 20% unit lift over 13 weeks at intraclass correlation 0.05 needs about 43 stores per arm.

What this does not prove

Price-response coefficients are conditional associations, not causal elasticities.
The recorded promotion flag is incomplete, and unlabeled weeks are not proven controls.
The demand model conditions on positive sales and cannot identify censored demand or stockouts.
Optimizer economics use historical accounting-cost proxies, not a current replacement-cost feed.
Substitution, basket effects, residual demand shocks, and future regime changes sit outside the scenario model.
A containerized, monitored service is not evidence that the policy is safe to deploy.