You are correct. The last two years of historical data has a lot of noise in it. I'm sure others will have more, but here are some things I've seen:
1. Build a model with all that data and see what happens. This can be a baseline to see how bad your errors are.
2. Shorten the historical data you feed into the model. This misses seasonality, but may fit the business better.
3. Use less historical data (like above), but adjust with some seasonal factors from 2017-2019- ish data. This may give you some insight to feed into the model.
I was a little sloppy on the forecast error. It is the standard deviation of the errors. The errors are the difference between the forecast and what happened. It can be confusing because different forecasting systems will use different terms for this-- or not provide it.
Another way to think about it is that the standard deviation of historical demand can be an OK approximation. I worked with an executive that insisted that we use this measure--and I liked the simplicity.
How do you take into account the last two years worth of "abnormal" peaks/valley of demand due to covid?
Wouldn't that significantly impact inventory levels?
You are correct. The last two years of historical data has a lot of noise in it. I'm sure others will have more, but here are some things I've seen:
1. Build a model with all that data and see what happens. This can be a baseline to see how bad your errors are.
2. Shorten the historical data you feed into the model. This misses seasonality, but may fit the business better.
3. Use less historical data (like above), but adjust with some seasonal factors from 2017-2019- ish data. This may give you some insight to feed into the model.
When you say forecast error, do you mean residuals from the fit of whatever forecast model you use?
I was a little sloppy on the forecast error. It is the standard deviation of the errors. The errors are the difference between the forecast and what happened. It can be confusing because different forecasting systems will use different terms for this-- or not provide it.
Another way to think about it is that the standard deviation of historical demand can be an OK approximation. I worked with an executive that insisted that we use this measure--and I liked the simplicity.