dreamprice-cso

Model Card: DreamPrice

Model Details

• Model name: DreamPrice
• Model type: Learned world model (RSSM with Mamba-2 backbone)
• Architecture: DreamerV3-style three-phase training with DRAMA decoupled posteriors
• Parameters: ~22M (full model)
• Framework: PyTorch
• License: CC-BY-NC-4.0
• Trained checkpoint: step_0100000.pt (100K gradient steps)

Training Results (100K Steps, DGX Spark)

World Model Quality (Evaluation on Validation Set)

Policy Comparison

Intended Use

DreamPrice is a learned dynamics model for retail pricing environments. It is intended for:

• Counterfactual demand estimation: "What would demand be if price were cut by 5%?"
• Pricing policy optimization: Learning pricing strategies via imagination-based offline RL
• Research: Studying learned world models in economic domains

DreamPrice is NOT intended for:

• Direct deployment in production pricing systems without human oversight
• Real-time autonomous pricing decisions
• Categories or retail environments significantly different from the training data

Training Data

• Dataset: Dominick's Finer Foods scanner data (Kilts Center for Marketing, University of Chicago)
• Period: September 1989 to May 1997 (400 weeks)
• Scope: 93 stores, ~18,000 UPCs, 29 product categories
• Primary category: Canned soup (cso), ~581K training tuples
• Temporal split: Train weeks 1-280, Validation weeks 281-340, Test weeks 341-400
• HuggingFace dataset: qbz506/dreamprice-dominicks-cso

Architecture Details

Causal Identification

Price elasticities are estimated via Hausman IV + DML-PLIV and frozen into the decoder:

• Instrument: Leave-one-out mean log(price) across other stores for same UPC-week
• Method: DoubleML PLIV with random forest nuisance learners, 5-fold cross-fitting
• DML-PLIV elasticity: -0.940 (SE=0.006, 95% CI: [-0.952, -0.928])
• First-stage F-stat: 23,381 (>> 10 Stock-Yogo threshold)

Resources

• Code: github.com/SharathSPhD/dreamprice
• Demo: huggingface.co/spaces/qbz506/dreamprice-demo
• Wandb: wandb.ai/qbz506-technektar/dreamprice

Limitations

• Trained on 1989-1997 data; modern retail dynamics may differ substantially
• Single category evaluation (canned soup); cross-category transfer not validated
• Offline learning only; no online fine-tuning or real-time deployment tested
• Causal identification relies on Hausman IV panel structure

Citation

@article{sathish2026dreamprice,
  title  = {DreamPrice: A Learned World Model for Retail Pricing via Mamba-2 Recurrence and Causal Demand Identification},
  author = {Sathish, Sharath},
  year   = {2026},
  url    = {https://github.com/SharathSPhD/dreamprice}
}