If the demand is too random, there is sometimes nothing better than a flat line of zeroes, or ones. However, most clients/stakeholders do not like flat lines for forecasts.
How this file is connected to other assets
Discover other files like this one
The Wasserstein or Earth Moving Metric looks at how much energy is required to reshape the predicted forecast into the actual data. For a lot of demand forecasting use cases, like inventory management, being close, say within a day, is usually pretty good and within the tolerances of the system, making Wasserstein distance a good choice.
They are challenging to model, difficult to scale, and painful to integrate. Credit https://www.nousot.com/resources/using-genai-to-completely-disrupt-traditional-platform-migrations-to-databricks-2/
Statistical, machine learning, and deep learning forecasting approaches each have their own unique pros and cons