Analyse statistical data

Produce the monthly statistical analysis for the demand forecasting team: clean the sales time series, run the seasonality decomposition, and deliver summary tables and charts to the forecasting lead, Daniel Reed, by Wednesday. Ensure the dataset excludes promo weeks flagged as outliers and include a note on any missing weeks; confirm the model inputs before sharing.

Before I hand the analysis to Daniel, make it actionable for planners. Put the forecast accuracy metric and the largest sources of error on the first sheet, show the seasonal indices where they materially shift demand, and flag weeks that deviate more than two sigma. Add a one-paragraph implication for each flagged week so planners can decide capacity changes without re-running models.

I ran the ARIMA and found a sudden autocorrelation spike in residuals around the holiday weeks, which makes the confidence intervals unreliable. Daniel wants to know if the model still works for allocation. I don’t know whether the spikes reflect true structural change in demand or simply promotional mislabeling in input data. What should I test next and what’s the least risky recommendation to operations right now?

Across several forecasts we keep getting caught by promo weeks that weren’t labeled in the input and the models then miss capacity needs. This costs planners time and leads to over- or under-ordering. Where are we losing accuracy and how should I change our data habits so statistical models stop breaking on predictable promotions?