[D] Building a demand forecasting system for multi-location retail with no POS integration, architecture feedback wanted

A retail tech team is developing a novel demand forecasting system that deliberately operates without POS integration or external data feeds. The architecture relies on operators manually logging just 4-5 key signals daily: revenue, customer covers, waste, category mix, and contextual flags like weather or local events. This constrained design aims to create a lightweight, accessible tool that outputs a clear weekly directive—telling managers what to expect, prep, and order—alongside a stated confidence level. The system is built for scalability across multiple retail locations where complex integrations are impractical.

The technical approach is phased and pragmatic. For the first 30 days, it uses only statistical baselines (day-of-week decomposition and trend analysis) without machine learning. After this period, it implements a "light global model" where similar venues train together but predict individually, improving accuracy through shared patterns. A key innovation is handling outlier flagging before model training, not after, with corrupted signal days excluded entirely. The team is actively seeking architecture feedback on three core challenges: determining whether a global model outperforms local statistical models with under 10 venues and 90 days of history, best practices for flagging anomalous days without external validation, and implementing lightweight, interpretable confidence intervals (like conformal prediction) that operators can trust as simple "high" or "low" signals.