The Challenge
RetailEdge's merchandising team was relying on a single spreadsheet with a manually tuned seasonality factor to plan stock for 200+ SKUs. Forecast errors were causing both overstock (tied-up capital) and stockouts (lost sales). The finance team had no confidence in the numbers and demanded a rebuild before the peak trading season.
The Solution
I built a Prophet-based time-series forecasting engine with automatic seasonality decomposition, holiday effects for 12 markets, and a Bayesian uncertainty band that gives the team a confidence interval — not just a point estimate. The engine runs in a scheduled Python job, writes results to Supabase, and surfaces the forecasts in an augmented Power BI report that highlights SKUs with high uncertainty so buyers know where to focus attention.
Results
23%
Improvement in 12-week MAPE
200+
SKUs forecasted automatically
18%
Reduction in overstock carrying cost
12
Markets with localised holiday calendars
Tech Stack
“Our buyers finally trust the forecast. The confidence intervals were the key — they tell us not just what to expect but where the risk is. We went into peak season with 23% better accuracy and it showed in our margins.”
Sophie Darko
Head of Merchandising, RetailEdge Co.