My client is currently working on demand forecasting for their products. They have asked me for some advice on demand forecasting using JMP. I think Time Series Forecast platform will suit in their case because they have thousands of items to forecast. I would appreciate it if someone could share with me some advice when doing demand forecasting using Time Series Forecast.
Their situation is as follows:
- They want to forecast demand for the next few months for thousands of items using historical monthly sales data (few years).
- Currently they are using simple moving average approach for forecasting.
- There is variety of sales pattern. Some items have seasonal pattern. Some items can be forecasted using moving average approach.
Here are some questions:
- Does anyone know a customer who does demand forecasting using JMP? It would be helpful to learn from your/their experiences, understand any lessons they have learned, and explore the strategies they have used to improve the accuracy of their demand forecasts.
- How often should we update our demand forecasts to ensure accuracy? Are there best practices for monitoring and adjusting our forecasts as new data becomes available?
- Are there any potential challenges or pitfalls that we should be aware of when conducting demand forecasting?
Any other insights or recommendations you can provide regarding demand forecasting would be greatly appreciated.
Hi Yuichi-San,
The only time I faced something close was a customer who wanted to forecast the project time and necessary resource plan (manhours) in the different steps of the process from planning phase to ordering to producing, and delivery to better predict the timeline for the customer and better align the internal people resources. All classified on innovation level, importance level, and some other factors.
However, to provide at least some guidance, I try to answer some of your questions:
1. They really need to analyze the patterns of their products selling behavior. If there is a seasonal trend then it should be included. If there is some other regular trend involved one might want to find our what AR mode best matches the time series to have a lower risk in the forecast. This is work, yes, but if they want to get something more accurate than brute force moving average this might lead the way. Often a graph builder trend chart with a column switcher is enough to identify the few products it may makes most sense to become more specific.
2. Updating the model is often dependent on how frequent you notice changes in the selling behavior. Usually for time series modeling I would say once a quarter is a good check point. If I have a very new product it really gets more difficult as the model may has to adjust several buying stages. Initial launch with not that much knowledge about the product, then the hype phase, then the steady phase before the cooling down phase and finally the steady phase again. Here I would recommend to more often check if the model needs to be adjusted (also because you may just start to sea seasonal trends after many months or even years (unless it is clear from the product itself, like skiing boots 
) The mature and regular the product is the less often you will need to update the model.
3. Pitfalls are often if there are special events distracting your data and regular trend. Just an example you may can translate to this customers products:
- a telecommunication company had made a forecast for their SMS revenue (long long time ago
) and at some point they saw a really high peak in the data for one day. That was when WhatsApp hat an outage. If you leave such data points in the data for the forecast you may get a wrong or less accurate forecast.
- If you have underlying structure in the data like products depending on e.g. weekend/week days usage, you may want to split the data and build a model for both separately. Otherwise you again get less accurate.
- Another thing is thinking about how far into the past I want to look. Some technologies have been important but after some time they got less important (just think about dvd/blue rays, blue rays where said to be the big new thing, but before they really took off streaming was the big new thing).
- Also think about how far into the future you want to predict and count on the prediction. Basically for similar reasons as stated above
I hope this helps a bit though it may not answer all your questions to the detail you would need.
May @jiancao has something to add as guidance?
Hi @martindemel ,
Thank you for your reply. Your advice is very informative for me. I belive it will help my client.
Thank you very much for your kind help.
Yuichi
