Assortment Performance Attribution

The core mechanism that estimates what would have happened if the merchandising change had not been made. This involves constructing a synthetic control group using historical data from similar products, locations, or time periods that were not exposed to the change.

• Uses difference-in-differences analysis to compare treatment and control groups over time
• Employs synthetic control methods to create a weighted combination of untreated units
• Accounts for pre-existing trends to avoid attributing organic growth to the intervention

The statistical process of separating the signal of a merchandising change from external noise such as concurrent price reductions, email campaigns, or seasonal demand shifts. Without this isolation, attribution is inflated or deflated by unrelated factors.

• Applies double machine learning to orthogonalize treatment effects from confounders
• Uses propensity score matching to balance treated and untreated observations
• Incorporates marketing mix data to net out promotional lift from merchandising lift

The output layer that breaks down total observed revenue into baseline revenue (what would have occurred anyway) and incremental revenue (the true causal effect of the merchandising action). This decomposition enables precise ROI calculation.

• Separates cannibalization effects from genuine category growth
• Quantifies halo effects where a featured product drives sales of complementary items
• Provides per-product and per-location granularity for localized decision-making

A temporal modeling approach that recognizes the diminishing causal influence of a merchandising change over time. A homepage feature may have strong impact in hour one but negligible effect by day three.

• Uses distributed lag models to estimate the decay curve of treatment effects
• Applies time-varying coefficients that adapt as the intervention ages
• Prevents over-attribution by capping the lookback window based on observed effect persistence

The recognition that a merchandising change does not have a uniform impact. The same banner placement may drive high incremental lift for new customers but zero lift for loyal repeat buyers who navigate directly to their usual purchases.

• Employs causal forests to estimate conditional average treatment effects across segments
• Identifies which customer cohorts, geographies, or device types respond most strongly
• Enables personalized attribution rather than assuming a single average treatment effect

A falsification check that applies the same attribution methodology to time periods or locations where no intervention occurred. If the model detects a significant effect where none should exist, the methodology is biased and requires recalibration.

• Shifts the treatment date backward to verify no effect is detected pre-intervention
• Applies the model to untreated control stores to confirm zero incremental attribution
• Provides statistical confidence intervals around all incremental revenue estimates

A quasi-experimental causal inference method that compares the change in an outcome over time between a treatment group (exposed to a merchandising change) and a control group (not exposed). It removes biases from both permanent unobserved confounders and common time trends.

• Parallel Trends Assumption: Requires that the treatment and control groups would have followed similar trajectories absent the intervention.
• Application: Measuring the lift from a new recommendation carousel by comparing test stores to a holdout group of similar stores over the same period.

A causal inference technique that constructs a synthetic counterfactual by weighting a combination of untreated units to closely mimic the treated unit's pre-intervention characteristics. It is ideal when a single, well-matched control group does not exist.

• Weight Optimization: Weights are learned to minimize pre-intervention outcome differences.
• Use Case: Attributing a revenue shift to a new dynamic assortment algorithm in a flagship store by comparing it to a weighted blend of similar stores that did not receive the update.

The process of measuring the causal lift generated by a specific intervention, distinct from correlation. It answers: 'Would this conversion have happened anyway?'

• Ghost Ads: Serving a control group a blank or PSA instead of the merchandising treatment to measure true baseline conversion.
• Holdout Groups: Isolating a random subset of users or locations from a new assortment strategy to quantify the net incremental revenue per session.

A game-theoretic approach that fairly distributes the credit for a predicted outcome among multiple contributing features or touchpoints. In merchandising, it decomposes a sale into the marginal contribution of assortment change, price reduction, and marketing exposure.

• Additive Feature Attribution: Ensures the sum of individual contributions equals the total prediction difference from the baseline.
• Computational Cost: Exact calculation is exponential; SHAP (SHapley Additive exPlanations) provides efficient approximations for tree-based and deep models.

A causal inference technique used when unobserved confounders corrupt the treatment effect estimate. An instrument is a variable that influences the treatment but has no direct effect on the outcome except through the treatment.

• Relevance & Exogeneity: The instrument must strongly predict the merchandising change and be uncorrelated with the error term in the outcome equation.
• Example: Using a supply-chain delay as an instrument to measure the causal impact of stockout visibility on customer churn, independent of marketing campaigns.

The offline estimation of a model's performance using historical data by asking: 'What would have happened if we used a different policy?' It allows safe testing of new attribution or assortment strategies before A/B deployment.

• Inverse Propensity Scoring (IPS): Reweighs historical outcomes by the inverse probability of the logged action to correct for selection bias.
• Doubly Robust Estimation: Combines IPS with a direct outcome model to provide an unbiased estimate if either the propensity or outcome model is correctly specified.