Three-in-One World Model: Energy-Based Consistency, Prediction, and Counterfactual Inference for Marketing Intervention

Researchers propose a Three-in-One world-model architecture using Deep Boltzmann Machines to unify marketing decision-making by simultaneously capturing consumer heterogeneity, predicting outcomes, and enabling counterfactual reasoning about interventions. The approach outperforms existing causal inference baselines in recovering treatment effects, particularly for confounded price-promotion scenarios.

This research addresses a fundamental challenge in marketing analytics: existing models typically optimize for either prediction accuracy or causal inference, but rarely integrate both within a coherent framework that captures latent consumer dynamics. The proposed DBM-based architecture solves this by learning a frozen belief representation from demographic and behavioral data, then attaching lightweight adapters for specific tasks. This separation of concerns—stable belief layer versus task-specific modules—enables the same underlying model to simultaneously perform consistency checks, make predictions, and generate counterfactual scenarios.

The work builds on decades of machine learning research, particularly the resurgence of energy-based models and the growing recognition that causal reasoning requires structured representations of latent variables. Marketing applications have increasingly demanded tools that understand not just what consumers will do, but why they do it and how interventions might change behavior differently across segments. Traditional meta-learners (S-learner, T-learner, X-learner) and causal forests have dominated this space, but they treat prediction and causal inference as separate pipelines.

For practitioners, the framework's superior recovery of heterogeneous treatment effects has immediate value in optimizing promotional spending and pricing strategies. The free-energy penalization mechanism provides interpretable signals about which counterfactual scenarios violate learned consumer patterns, reducing the risk of deploying interventions that contradict observed behavior. The ability to query counterfactuals while maintaining belief consistency suggests potential applications beyond marketing, including personalized recommendations and dynamic pricing.

Future work should examine scalability to real-world datasets and whether the approach generalizes across industries with different intervention structures.

• →DBM-based world models unify prediction and counterfactual reasoning in a single framework, outperforming specialized causal inference baselines on treatment effect estimation.
• →The frozen belief layer with task-specific adapters enables consistent answers to prediction, consistency, and counterfactual queries without retraining the core model.
• →Free-energy penalties identify counterfactual scenarios that violate learned consumer patterns, providing interpretability and reducing risk in intervention deployment.
• →Performance gaps were largest on confounded price-promotion interventions, suggesting the approach handles complex real-world scenarios better than existing methods.
• →The architecture disentangles latent consumer traits like base preference and promotion sensitivity in a form that survives counterfactual analysis.