A retailer needs 30-day demand forecasts for thousands of items using years of daily data. Stockouts are much more costly than overstock. Item-level features (category, brand, safety stock) are available, and a binary promotion flag exists but only has known future values for 5 days. Which Forecast algorithm and evaluation metric should be selected to maximize profit under these constraints?

The correct choice is CNN-QR with wQL at P75. CNN-QR is suitable for this problem because it can handle thousands of items (scalability), leverage item-level features, and incorporate the promotion flag. Its ability to learn complex patterns from time series data makes it robust. Since stockouts are more costly than overstock, a higher quantile (P75) is appropriate. This over-forecasts demand, reducing stockout risk. Weighted Quantile Loss (wQL) directly optimizes for this asymmetric cost function, making it the ideal evaluation metric. ARIMA models are less suitable as they are typically univariate and struggle with thousands of items and incorporating multiple features efficiently. WAPE (Weighted Absolute Percentage Error) is a symmetric error metric, meaning it treats over-forecasting and under-forecasting equally, which is not aligned with the problem's asymmetric cost of stockouts.