qbz506/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)

Metric	Final Value
World Model ELBO	22.44
Reconstruction Loss	0.001
KL Divergence	19.20
Reward Prediction	3.17
Actor Return (mean)	124.33
Critic Loss	2.50
Training Time	~2.6 hours
Hardware	NVIDIA DGX Spark (GB10 GPU, 128 GB unified memory)

World Model Quality (Evaluation on Validation Set)

Metric	h=1	h=5	h=10	h=13	h=25
RMSE	5.001	5.167	5.286	5.283	5.243
MAE	2.130	2.168	2.207	2.205	2.197
WMAPE (%)	71.7	72.3	72.6	72.6	72.4
NDR(h)	1.00	1.033	1.057	1.056	1.049

Policy Comparison

Method	Mean Return	IQM
Cost-plus (25%)	54.8	54.8
Static XGBoost	87.2	85.6
Competitive Matching	42.1	41.5
DQN	68.9	65.2
PPO	76.4	72.8
SAC	82.3	79.6
DreamPrice	124.3	117.4

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

Component	Specification
Backbone	Mamba-2 SSM (d_model=512) with GRU fallback
Stochastic latent	32 categorical variables x 32 classes (z_dim=1024)
Posterior	DRAMA-style decoupled: q(z_t | x_t)
Observation decoder	3-layer MLP: cat(h_t, z_t) -> obs_dim
Demand decoder	Causal: theta * log(price) + MLP(z_t, store_features)
Reward ensemble	5 independent heads, twohot distributional (255 bins)
Continue head	Linear -> sigmoid
MOPO pessimism	r_pessimistic = r_mean - lambda_lcb * r_std

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}
}