Summary
Deciding which actions to take in retail from data can be tricky because different methods sometimes give conflicting advice. The authors created a system called F-DACE that checks multiple ways of estimating cause-and-effect, and only makes a recommendation if these agree and pass several quality checks. If the evidence is shaky or contradictory, the system wisely chooses not to recommend any action. When tested with simulated data and real Walmart store info, F-DACE made careful decisions most of the time and avoided many wrong suggestions. It also powers a conversational assistant that reliably answers user questions about retail decisions.
What this means in practice
- •For retail data teams: Provide safe causal recommendations by fusing multiple estimates and abstaining when evidence conflicts or is weak using retail panel data.
- •For conversational ai developers: Build dialogue agents that engage users with trustworthy causal explanations and optimization suggestions grounded in multiple causal analyses.
Abstract
Observational decision-support systems often expose one causal estimate as a recommendation even when plausible estimators disagree. The inherent engine of the proposed system is causal machine learning: a conditional-average-treatment-effect estimand identified by backdoor adjustment, estimated by an EconML DML causal forest and DoWhy linear regression, checked by two-way fixed effects, and converted into candidate levers by constrained optimisation. F-DACE is the decision layer on that engine. It represents precision, propensity overlap, placebo-refutation stability, interval overlap, and directional agreement as fuzzy memberships. Hard vetoes force abstention after estimand mismatch, failed diagnostics, informative sign conflict, or weak evidence. In 180 panel simulations spanning six identification conditions, F-DACE made a decision in 67.2% of runs and limited false recommendations to 17.2%; the corresponding rates were 33.3% for the causal forest and 35.6% for backdoor regression, matching deterministic unanimity rather than dominating it. Nearly all (30 of 31) false recommendations occurred under shared unmeasured confounding, which no fusion rule can diagnose when every component shares the omitted variable. The retail application aggregates a public Walmart panel to 6,435 store-weeks across 45 stores. F-DACE abstains for all five markdown indicators: some estimates are imprecise, one refutation fails, and MarkDown5 has a direct sign conflict. A LangGraph conversational agent exposes impact, what-if, and lever-optimization tools while a deterministic verifier preserves causal-layer status. On 24 live questions it achieved 100.0% tool-routing accuracy, 100.0% status fidelity, and 0.983 mean groundedness. On ten adversarial questions it resisted all injected instructions.