Machine Learning Forecasting Process
Machine Learning Forecast Manager was built to make it easy for a Demand Planner to generate a machine learning influenced forecasts alongside the standard DemandCaster forecasting process.
The process requires only 5 steps: requiring only 2 main process steps:
- Selecting the PCL's to be forecasted
- Determining if the selected PCL's to be forecasted should be retrained
- Selecting how many periods to forecast
- Determining if you would like to include Price as Related Time Series
- Running the ML forecast
- Viewing the results
The current forecast method is necessary to generate forecasts beyond the ML forecast horizon. Though this is an optional process, we do recommend it in order to drive a long term view of demand.
To run a new demand plan forecast, follow the steps described in the article Forecasting Process.
The standard forecasting process may be run in parallel with the ML Forecasting process.
Determine which PCL's to run an ML Forecast against
As stated in the article Machine Learning Introduction, the number of forecasts that can be run is determined by your subscription tier. This means that in many cases, you must determine which entities to forecast. To learn how to select the entities to run, please review the article Contextual Filtering.
Running the ML Forecast
In the image below, we have selected the customer "Greensburg" and all it's PCL's to forecast.
Once the PCL's are selected, click “Run ML Forecast” to open the pop-up to commence the forecasting process.
(1) Select Initiate New Training
Checking this option will initiate a new training upon initiating the ML Forecast process. Training update the model metrics, based on which Machine Learning will execute forecast. We recommend to initiate training if significant event happened with PCL or there is new PCL selected, since the learning could take a number of hours.
(2) Enter the Number of Weeks to Forecast
Enter the number of periods to forecast. Please note that this
It is the number of time steps being forecast . If you have less than 2 years of history data we are highly recommend to use not more than 26 weeks, to have more accurate forecast.
(3) Select Include Price as RTS
Checking this option will incorporate the price as a related time series, so the Machine Learning model can capture the impact of price changes on the item's demand and adjust the forecast accordingly. Related Time Series functionality in details described here: Related Data
(4) Aggregation Type
Aggregation Type allows choosing between PCL level or aggregated level forecasting. For aggregated level forecasting, simply define the desired level from your forecast hierarchy. Example above shows aggregated forecast at 'Location – Item' level minimizes demand noise from customers. Forecasts are disaggregated using a Top-Down approach based on historical data.
ML Forecast line
The results from Machine Learning forecast will be visible in ML Forecast line and will be the basis of the Blended forecast line as shown below and described here: Blended Forecast Concept

0 Comments
Add your comment