Dual-Band Supply Chain Lead Time Predictor
Model Overview
This repository contains two parallel Gradient Boosting Regressors designed to automate supply chain scheduling by predicting fulfillment timelines from raw Purchase Order parameters.
Instead of relying on static, human-assigned lead times, this architecture utilizes a Dual-Band approach to balance operational accuracy with financial risk management:
- Model A (Expected Mean): Optimizes for
squared_error to predict the statistically most likely arrival date.
- Model B (Safety Buffer): Optimizes for
quantile (alpha=0.95) to predict the 95th-percentile worst-case scenario, providing an impenetrable safety buffer to prevent assembly line stockouts without holding excess inventory.
Architecture & Features
- High-Cardinality Target Encoding: Rather than standard One-Hot Encoding, this pipeline uses dynamic Volatility Target Encoding. It extracts both the historical
mean (average delay) and std (standard deviation) for every Vendor, City, and Item. The algorithm mathematically distinguishes between consistently slow vendors and highly volatile/unpredictable vendors.
- Data Leakage Prevention: Human-assigned scheduled dates are explicitly dropped during training to force the model to learn actual physical transit times.
Intended Use
This model is built for supply chain operations teams and ERP integration. It ingests unscheduled bulk Purchase Orders and outputs two calendar dates: the Expected Arrival Date, and the Safest Scheduling Date (for manufacturing floor planning).
Training Data
Note: To comply with corporate confidentiality, these public models were trained on a 10,000-row synthetic dataset mathematically engineered to mirror the high-cardinality constraints, noise, and vendor volatility profiles of a proprietary enterprise ERP environment.
