Causally informed multimodal offline reinforcement learning for constrained promotional pricing in omnichannel convenience retail

Promotional pricing in omnichannel convenience retail requires retailers to balance contribution profit, sales retention, inventory risk, and operational feasibility under non-random historical promotion decisions. This study develops a causal multimodal offline reinforcement learning approach for SKU–store–day promotional pricing in an omnichannel convenience-store setting. The proposed method integrates product images, textual descriptions, customer review signals, transaction records, inventory information, store characteristics, and external conditions to represent heterogeneous demand contexts. Discount levels are treated as multi-level treatments rather than ordinary predictive features, and a doubly robust estimator with cross-fitting is used to estimate counterfactual demand responses. These estimates are incorporated into an uncertainty-aware conservative offline reinforcement learning model that penalises weakly supported actions and reduces extrapolation beyond historical promotion behaviour. A constrained portfolio optimisation layer then converts product-level discount values into operationally feasible store-day promotion plans. Using 19,684,527 SKU–store–day observations from 936 products and 128 stores, the proposed causal multimodal model achieved a demand WAPE of 23.1%, compared with 26.8% for the tabular neural benchmark. In the out-of-time model-based evaluation, the learned policy was estimated to improve contribution profit by 8.7% relative to the historical retailer policy, while retaining 99.1% of sales and reducing the stockout rate from 8.9% to 8.1%. These findings provide model-based evidence that combining causal demand estimation, multimodal context, uncertainty control, and operational constraints can improve the reliability of offline promotional pricing recommendations. Not applicable.