Over the last 18 months there has been a great deal of interest in the application of AI/ML Forecasting and as covid data has become embedded in companies’ sales history, the need for additional demand indicators and systematic methods to utilize them has accelerated.
In our latest Insight Hour: Using AI/ML in Forecasting, the aim10x Innovators Network members discussed roadblocks to forecast accuracy, data input quality & accountability and the changes machine learning is having on Demand Planner Skillsets and IBP. The session was presented and chaired by Nitin Goyal (o9) and Fazlur Rahman (Kraft Heinz).
For this peer session the topics covered ranged from how to improve the quality of source data and gain better forecast accuracy to the impact of implementing AI/ML Forecasting on traditional IBP and how planner skills are changing in the face of new forecasting solutions.
Using AI to Understand a True Demand Signal
Nitin Goyal prefaced the session by explaining how history is no longer considered the only indicator for predicting the future. He expanded further that historical data represents what did happen, not what could have happened! How much Market Share do you have? Did a competitor have a successful marketing event? Was there a stock out? These events can distort and disguise the true demand signal.
Traditional statistical forecasting processes are only supported by lagging indicators, and this creates a lot of manual work to account for planned changes in future. Connecting Leading Indicators of demand with internal enterprise data is challenging and typical planning solutions lack the capability to run scenarios. The resulting forecast accuracy issues cause a lot of friction between Finance, Sales, and Supply chain.
This is where AI/ML Forecasting, using not just history but additional internal and external drivers as lagging and leading indicators, can really transform the accuracy of your forecasts. o9’s flexible, open source, machine learning forecasting solution is able to consider any driver of demand such as promotions, pricing, competitor activity, holidays, local events, cannibalization, policy changes, macroeconomic factors, industry data, weather and more.
Moving Beyond History as a Forecast Indicator
To entice discussions, the participants were invited to answer two poll questions: “How would you rate your forecasting accuracy?” and “What inputs are you currently using for forecasting?”. The results showed that most participants struggle to get above average accuracy and are moving beyond pure history as a forecast indicator. The resulting discourses revealed that many were conducting small scale experiments with additional driver data to increase forecast accuracy and were hoping to ramp up driver usage over time.
Roadblocks to Improving Forecasting Accuracy
The group discussed challenges to improving forecast accuracy and members articulated a range of issues that they were experiencing in the pursuit of forecast improvement including business cycles not being uniform, disjointed systems, incorrect historical weighting, Excel spreadsheets outside the system of truth, and planner capability”.
Three suggestions that really resonated were “Lack of confidence, adoption & collaboration”. The member offering these examples explained that planning insight not being believed or shared was resulting in failures to make good decisions.
Others brought up the significant problem of lack of trust in the forecasts produced by a ‘Black Box’. How do you get the business to buy into what is seen as almost magic?”
Another member explained the belief that one of the biggest challenges they were facing was also the answer to the issue of an untrusted black box, obtaining insight and using it effectively.
Quality of Inputs and Accountability
The challenge of gaining insight nicely segued into the next topic, which was the quality of inputs and accountability for the forecast. Some questions asked included, “How do you reconcile different assumptions and reach consensus?” and “Has anyone overcome the struggle with data quality?”
Members answered that data across multiple systems in different granularity and states of cleanliness were significant challenges that they were working hard to overcome. The rationalisation of source systems, defining the correct data on which to base their forecasts and data cleansing were spoken of as activities to improve data quality and facilitate clearer accountability. One member stated that his company was looking to bring in more ordering automation including utilising orders for short term demand sensing.
Changing Traditional IBP Process and Planner Skills
Considering the increased use of automation and the advances of AI/ML, the moderator asked the members what changes they thought were likely to happen to the traditional Integrated Business Planning processes and the skills required from planners.
There was agreement amongst the members that they were seeing changes happening. One common observation being the shift towards selling as compared to forecasting by sales. Another typical theme was the shift of forecast horizon emphasis as algorithms resolve short-term forecasts allowing planners to focus on the mid- and long-term.
As one member pointed out, people “like their numbers” and that they are encountering a huge shift of planning emphasis “as we change our IBP process to be action oriented and forward looking, because the AI machine learning being able to do most of the historical autopsy.”
Embracing Planning Transformation
July’s Peer Session included some fascinating insights about how members are tackling their planning transformations by creating solutions and processes that can collect additional drivers to create Machine Learning forecasts that will reform their company’s supply chains. To conclude the discussion, a member suggested that the topic for the next meeting should centre on how to collect additional drivers and assess if they should supersede existing ones.