DreamPrice: A Causal DreamerV3 World Model for Offline Retail Pricing

Retail pricing is a high-stakes sequential decision problem in which every price change triggers a cascade of demand responses: consumers substitute across products, stockpile during promotions, and competitors react, with effects that propagate over weeks and quarters. Despite decades of work on demand estimation and price optimization, the systems used in practice rely on either structural econometric models with manually specified demand curves or model-free reinforcement learning methods that cannot perform counterfactual reasoning. No learned world model has been applied to economic environments where transition dynamics are endogenous to the agent’s actions. This paper introduces DreamPrice, to our knowledge the first Dreamer-style world model for retail pricing. The system combines DreamerV3’s three-phase training recipe with a Mamba-2 selective state-space backbone, DRAMA-style decoupled posteriors, and a causally-constrained demand decoder. Causal price elasticities are pre-estimated via Hausman instrumental variables combined with double machine learning (DML-PLIV) and frozen into the decoder, preventing the model from learning confounded price-demand relationships from the observational data. The architecture replaces the gated recurrent unit (GRU) backbone of standard recurrent state-space models with Mamba-2’s structured state-space duality (SSD), enabling O(n) parallel training and O(1)-per-step recurrent imagination. MOPO-style pessimism via a 5-head reward ensemble guards against distributional shift inherent in offline learning. Trained on the Dominick’s Finer Foods scanner dataset (1989–1997, 93 stores, canned soup category with ∼25 stock-keeping units), DreamPrice achieves a world model ELBO loss of 22.44 and an actor return of 193.7 in the symlog-transformed reward space after 100K training steps. An ablation study across nine configurations confirms the contribution of each component: removing offline pessimism degrades return by 85.6%, removing symlog and twohot representations increases world model loss by 25×, and the Mamba-2 backbone yields 28.9% higher return than a GRU alternative.