The Demand Planner's Hope:
Give me peace of mind in knowing that I cannot predict the future, the insight to manage the products that matter, intelligent systems to take care of the products that don't, and the analytics to help me know the difference.
Demand Planners all over the world are hard at work trying to improve forecast accuracy by fine tune statistical models and using a lot of manual intervention to change forecasts. The sales people are being asked to enter their guess on what the customer will buy in addition to doing their primary job of selling product, while marketing is asked to predict the distant future. Once all that is done, everyone comes together for a meeting to come to a consensus number that will be used to drive the supply plan. The key question is what out of all that effort is adding the most value (or even what of that effort is making things worse), and where should the next hour that all those people have available be spent to help the company the most. There are two analysis techniques that can be put in place in order to help figure out how well the organization is performing in the demand planning process, and more help coming in the area of machine intelligence so that the planners can get more from the system with less effort.
The best way to know where the biggest contribution to forecast accuracy is coming from and machines to learn, is to use the principle of "Value Added". This is not just the measurement of Forecast Accuracy for each step in the process, but the measure of how much better the forecast gets during each step in the process. This way you can stop or correct the effort for specific products, where value is destroyed, or little value is added. Value in the process is created when the forecast improves during a given step in the process. The baseline measure is to do nothing and just use the actual from the last forecasting period for this period. If the statistical forecast is the first step in the process, that step is adding value if the statistical forecast is more accurate than the last cycles actual. Each step is measured the same way, comparing it to the number given by the prior step and seeing if the number is getting better or worse and by how much. If the next step is the planner update of the statistical forecast, a measure is taken to see if the planner is getting closer or farther away from what ends up being the actual shipment. Proper statistical measures should be used to create the measures. This same "Value Added" measurement should be done for any step included in the process.
The insight to manage products that matter, can best be done with an A, B, C analysis that segregates products not just based on revenue or unit volume, but by making the analysis 2 dimensional and including a second A, B, C analysis on predictability. This ABC x ABC matrix of categories, show where effort by the planners will pay off the most. The most effort should be put into the products that are the highest value products, and the hardest to predict with just intelligent machines. These products may change over time, but at any given point in time even in a company that has 10,000 products to manage the highest value, least predictable products will number only around 30 products. The most predictable products can be managed using automation and machine learning (see next paragraph). Unpredictable products with low value, can be managed by intelligent systems with human assistance when problems arise, but time should be allocated based on the value of the category.
Intelligent systems and machine learning to take care of the product that do not need human intervention is growing in sophistication and power. It used to be that a planner would spend time tuning time series statistical models in order to increase forecast accuracy. Previously, Forecasting systems would focus on how the planner can use the system to fine tune those models. Now intelligent systems are fast enough and smart enough and learn from their forecast failures and successes. The world of AI computing and genetic algorithms have become a reality when it comes to measuring how well forecasting algorithms are performing, tuning them, and recognizing patterns in the data that give clues as to where the market place is going. What takes a person 10 minutes to analyze and decide a computer can simulate, measure and adjust hundreds of thousands of times. It does not make sense to have one forecast anymore internal to the system, but many can be stored and monitored to see which ones are giving the best results for each product. It is survival of the fittest, with weaker methods, being beaten by stronger ones, and learning taking place with each iteration. Only the toughest of problems require human intervention.
In summary, the Demand Planner should have hope that the understanding, intelligent systems, and analytics are out there and they really are helping make life easier.