Build a Demand Forecaster You Can Order Stock From

• ✓Build a seasonal-naive baseline and measure every forecast against it
• ✓Evaluate a forecast by walking forward in time, and show a random split overstating the score
• ✓Build forecast-time features and find the leaked feature that would not exist when you order
• ✓Forecast the 10th, 50th and 90th percentiles, check that the interval covers what it claims, and repair it from past errors
• ✓Turn a forecast range into an order quantity from the cost of running short and the cost of waste

Forecast demand the way a planning team does — beat a baseline, test by walking forward, keep the future out of your features, report a checked range, and choose how much to order from the cost of being wrong.