How demand planners augment Machine Learning Forecasting Models
Best Outcome = Human Plus Machines (NOT Human Vs Machines)
Demand planning starts with a reliable, automated machine forecast. However, the forecast is complete with the addition of human insights, layering on inventory strategies, and reconciliation with consensus forecast .
With the continuing evolution of machine forecasts with AI/ML, the reliance on demand planners for manual forecasts will come down. This should free up time for demand planners to focus on strategic decisions. However, there will be two primary areas that will continue to primarily depend on the demand planner.
- Educate and augment the machine forecast for unpredictable events and to align with the chosen inventory strategies for those situations.
- Where machine forecast is not accurate or not confident, take control. Forecasting errors can cost companies millions of dollars in losses either by ordering too high or ordering too low. Between the machine and the human, it is important to catch these errors early before they lead to adverse impacts.
#1. Examples that humans “Educate and augment the machine forecast”
- When a major virus shuts down a strawberry farm, the strawberry demand forecast for that region might have to be transferred over to an alternative product. It is unlikely for a forecast model to be able to know a local farm’s problems and hence need human intervention. Additionally, forecasts in upcoming years need to be able to account for zero sales due to unavailability of strawberries. The same is true for large sales events that affect supply availability like Suez Canal blockage or a Covid-like pandemic.
- Region-specific differences in hurricane recovery are the norm. For folks in Florida, products required for home builders are different from those in Louisiana. Also, the timing of relief packages in these local geographies (available in the news, etc) indicates when demand is going to pick up.
- Knowing what is to be managed in a forecast vs what is outside of a forecast. Ex. During natural calamities, emergency need items such as water, torches, and batteries are manually allocated to specific locations- This is typically referred to as “push”. The aggressiveness and recovery of the natural calamities need humans to have control over the response to get products to the customers in need.
- Many calamities and large supply chain disruptions do not repeat and even so do not repeat the same time next year. While machines are trying to catch up on what is repeat vs a one-time behavior, it is still a space that has not peaked on the maturity curve. And hence, there is continuing reliability on the planner to “tag” and/or “clean “ these one-off events from history, to make the model aware of the data context.
- There are scenarios where large-scale model changes are not tested enough for wide scale implementation. This is when planners resort to switching forecasts manually until models are corrected. In order to protect business risks, negative impacts to inventory availability or inventory cost risks, it is important to have that fall back control in the hands of humans.
#2. Examples of human intervention “Where machine forecast is not accurate or not confident, take control”
- Some forecast errors can be highly expensive. I have seen a single outlier sale (found to be bad data later) in history drive the forecast of a $1200 item by 10x. This item was just 1 of the 100 million forecasts. Imagine what happens if we ship 100 of shower doors (that each cost $1200) and to a store instead of the regular 10 units — the store backspace is clogged, the cost of reshipping those back to the DC or another store, unavailability of products to the other stores, the cost of this error is huge.
- Considerations of cost of error in terms of impact to customer promise, impact to cost of execution and impact to handling costs, highly strategic products (traffic drivers, Key Value Items), locations, time periods, and highly strategic vendors need additional review when the forecast is not confident or has not had reliable forecast accuracy. The same is true for high retail, high cube and high cost items.
Summary
With the advent of AI/ML and neural networks, forecasts are going to be more and more machine driven. The role of a demand planner evolves with the leaps in data science that led to forecast automation. However, until the models are fully mature, the demand planner will need to “coach” the models and be the final “guard rail” in reviewing forecasts that are outliers and/or impacted by never-seen situations.
Related Medium Stories
