HCFT Retraining — Labour Forecast Model Optimization

Context

The Headcount Forecasting Tool (HCFT) supports labour planning across Coles’ 1,800+ stores, predicting staffing needs to balance operational efficiency with service quality. As part of my internship in the Data & AI Command Centre, I worked on improving the model retraining pipeline and streamlining how UAT (User Acceptance Testing) packs are created and reviewed.

Problem

The existing HCFT retraining workflow involved pulling plots from multiple platforms, manually assembling slides, and cross-referencing accuracy metrics — taking approximately 7-8 minutes per slide. The fragmented process made it difficult to ensure consistency and slowed down the feedback loop between the data team and store managers.

What I Built

• Refined Smoothing Logic — Tested and validated improved smoothing parameters in Jupyter/Python to enhance model stability across diverse store profiles
• Consolidated Palantir Dashboard — Combined multiple scattered views into a single Palantir interface for reviewing model outputs, reducing context-switching during UAT pack creation
• Accuracy Validation Pipeline — Conducted systematic accuracy checks across stores to ensure retraining met performance benchmarks before deployment
• Workflow Documentation — Created comprehensive retraining documentation for future cycles, ensuring knowledge transfer to the permanent team

Impact

UAT pack creation time dropped from ~7-8 minutes per slide to ~3-4 minutes — approximately 50% faster. The consolidated dashboard and documented workflows ensured consistency across retraining cycles and reduced dependency on institutional knowledge.