Choosing the Right Forecast Error Metric

You are comparing three forecasting models across 200 different retail SKUs, some high-volume (hundreds of units/day) and some very low-volume, intermittent items (0–2 units/day, frequently exactly 0). Your manager wants "one number" to rank the three models overall.

1. Why is MAPE a poor choice for this comparison, specifically because of the low-volume SKUs?
2. Why is raw MAE (averaged across all 200 SKUs) also misleading here?
3. Propose a metric (or combination) that would give a fair, one-number comparison, and justify it.