---
license: apache-2.0
language:
- en
pipeline_tag: tabular-regression
tags:
- tubular-regression
- demand-forecasting
---

# Shelfy: Tomorrow's Best Sellers
**AI Demand Prediction** - Rank products by tomorrow's conversion rate. Top 30% products captures **50% of actual sales**.

## Usage
1. Enter today's product stats
2. See tomorrow's predicted conversion + revenue
3. **Stock Top 10% products → 2.5x ROI**
   
## Business Results (42M events test)
- Top 10% → **25.4%** sales - ROI: **2.5x** <br>
- Top 30% → **52.1%** sales - ROI: **1.7x** <br>
- **Test:** Oct 25-29 (blind)  **Dataset:** 42M ecommerce events

## Features
- **Live prediction** - Tomorrow's conversion rate per product
- **Demand signals** - High / Medium / Low
- **production features** - Views, carts, momentum, intent ratios

## Model Card
- Architecture: XGBoost Regressor
- Features (8): views(29%) + carts(29%) + cart momentum(13%) + user metrics(17%) + other(2%)
- Target: Tomorrow's purchase rate (actual purchases / views)
- Stability: 82% day-to-day Top 10% products overlap

## Production Stack
- Data: Clickstream (views, carts, purchases)
- Stack: Python + XGBoost + FastAPI + PostgreSQL
- Retraining: Weekly with shelfy's real data

## Tech Stack
- Model: shelfy_purchase_rate_v2.joblib (XGBoost)
- UI: Gradio 4.44.0
- Data: Ecommerce behavior (Oct 2019)

---
**Built for datastorm hackathon** | **Apache 2.0** | **Deployed 1/16/2026**