how-to-sales-forecastinglisted

# how-to-sales-forecasting Data-Science-layer skill (layer 3 of the phoebe-data-skills 4-layer roadmap). The job is not "fit a model" - it is to put a number in front of a planner and a P&L owner that both can sign. Most forecasts fail one of three ways: they measure a data defect as demand, they hide a once-observed event inside a confidence interval, or they deliver units when the meeting decides dollars. This skill exists to avoid all three. Showcase walkthrough (Everrest retail case, real executed charts): https://github.com/phoebefu6/phoebe-data-skills - `docs/how-to-sales-forecasting/` ## Where this sits in the lineage `raw dump -> lake -> warehouse -> marts` **-> demand forecast -> inventory buy**. Reads the clean marts (the output of `how-to-schema-and-warehouse`), and STILL runs its own data-quality gate - clean-looking marts carried a duplicate-row defect that manufactured a whole phantom demand spike. ## The three rules that separate it from every forecasting tutorial 1. **DQ gate before any measurement.** Duplicate join keys fan out and create phantom demand. In the showcase, a retry bug (600 duplicate order rows) manufactured a 1.49x "March spike" that a naive forecast would have bought ~$109k of stock for. Dedupe first, report what was removed. 2. **Count observations per seasonal cycle.** Weekly seasonality: ~48 obs per weekday - measurable. Annual seasonality: 1 obs per month - NOT measurable. A once-observed peak is a parameter, not a find