superdkj/retail-world-model-scripts

Retail World Model

A transformer-based world model for retail sales prediction, trained on the Favorita store sales dataset.

Architecture

• Encoder: T5 Encoder processes historical retail time-series
• Latent Dynamics: LSTM-based world model that imagines future states
• Decoder: Predicts future sales with probabilistic forecasting (mean + variance)
• Fine-tuning: LoRA adapters on encoder attention layers

Features

• Context Window: 60 days of historical sales data
• Prediction Horizon: 14 days of future sales forecasts
• Inputs: Sales history, promotion flags, day-of-week, month, product family
• Output: Probabilistic predictions with 90% confidence intervals

Usage

Training

python train_v4.py

Inference

python inference.py --checkpoint /path/to/model --history history.csv --output predictions.csv

Dataset

• Source: t4tiana/store-sales-time-series-forecasting
• Stores: 54 Favorita stores in Ecuador
• Product Families: 33 product categories
• Records: ~3M daily sales records (2013-2017)

Model

• Base: T5 Encoder (tiny)
• Parameters: ~30M base + LoRA adapters
• Loss: Negative log-likelihood (probabilistic forecasting)

References

• Chronos-2: Universal Time Series Forecasting (Amazon Science)
• iTransformer: Inverted Transformers for Time Series (NeurIPS 2023)
• TabFormer: Tabular Transformers for Sequential Data (IBM Research)