Unlocking Better Forecasts: Useful Concepts for Data Scientists to consider for their next project

Over the past few months, I’ve been working on various forecasting projects and found myself asking what I assumed were pretty standard questions for anyone in the field. As it turns out a lot of work has already been done developing approaches and techniques to address these challenges.

Thanks to my Data Science Team and the internet writ large, I was able to find the answers I needed, all of which will be invaluable in my future forecasting work. Since I wasn’t initially aware of these solutions, I thought it might be helpful to share them in a blog for anyone else who has similar questions.

How do I know what forecast metrics to present without overwhelming the audience while still showcasing an accurate view of how the solution is performing

Say each week, you’re forecasting demand for 50 SKUs over the next 10 weeks. That means the first time forecast was run, it was run to predict demand each week for weeks 1–10. Then it was run to predict demand each week for weeks 2–11. We have decided to categorize the forecasts into short-term (Weeks 1–4), medium-term (Weeks 5–7), and long-term (Weeks 8–10) forecasts.

Lets say the business asks to review accuracy for short term forecasts, you have enough data to validate the first model run (Weeks 1–4) and second model run (weeks 2–6). We have two options here: