Out of stocks have always plagued the grocery industry, but the rise of Covid driven digital shopping last year exposed some new causes for concern.
By ShopAbility co-directors Peter Huskins and John Day, and Blue Yonder Solutions Director Sanjay Prakash.
Embrace uncertainty in demand forecasting
Demand forecasts are now increasingly built on machine learning analysis of the relationships between many different historical data sources, such as weather, special events and price. The trick is not to use artificial intelligence to layer influencing factors on top of a baseline, but rather look at how strong each influencing factor is at any given point
in time and use this as the basis for forecasting into the future. Missed sales are accounted for and special events are automatically moved as calendar dates shift.
Once trained, the forecast engine tests itself against actual sales every day to ensure that, as shopper behaviour changes, the model can self-correct. Configurations such as weekly profiles and assumed statistical distributions are made redundant as the forecast is created every day from the ground up, using the most recent data to improve accuracy. This changes the role of demand planner from algorithm tuner into data custodian and strategic adviser.
The demand forecast needs to predict the full spectrum of demand, not just the mean, and calculate the probability of every unit of demand at the item, store and day level, rather than assume a shape or profile. This information becomes valuable when you later want to decide on inventory, capacity for planograms, and price or assortment changes. Irrespective of how close your mean might be to historical sales, there is always a chance that shoppers will want to buy more or less of what you predict.
A demand forecast that intuitively understands what factors drive shopper behaviour is useful for predicting demand for all retail assortments.
Read more on supply chain management in the latest issue of Retail World.