Years of sales history sitting in spreadsheets, unused — turned into a forecast that holds up out of sample, not a curve fit to the past.
The model isn't picking the line that looks best on the chart it was fit to — it's validated against periods it never saw during training, the same discipline used to score out-of-sample edge in the risk models this is built on. 6.8% error on unseen weeks, not on the data it was trained on.