I’m a cyclist. But I have a confession to make. I’m a fair-weather cyclist. While riding my bike helps keep me healthy and gives me time to think, when I get caught in a rain shower, I know I’ll be spending a significant chunk of my afternoon cleaning mud and gunk off my bike. And so, based on my willingness to accept risk that day, I make a judgement call. Before I leave the house, I always consult the weather app on my phone. Anything more than a 30% chance of rain and I stay inside.  I estimate that has saved me hours of de-greasing in the past year alone.  

I find it interesting that concepts that we readily accept with our personal devices don’t always seem to resonate in the workplace.  When I’m considering whether to risk cleaning my bike after a ride, I rely on a weather forecast that may or may not materialise. I don’t question whether the forecast is right, I accept that it is uncertain; sometimes it rains, sometimes it doesn’t. I make my decision based on a preference for staying dry.  

Demand Forecasts are Uncertain 

The same principles of risk and uncertainty that we accept in weather forecasts apply equally to demand forecasts. Order too much and you lock up capital in slow moving inventory or eat into margins for perishables Order too little and you run out of stock and disappoint customers. Planning inventory relies on an accurate demand forecast. Understanding that forecasts always include a level of uncertainty is the critical first step toward resilient supply chain automation.  

So how do you build a resilient demand forecast that understands uncertainty? Traditional approaches to forecasting are fundamentally flawed. Basing a demand forecast on sales history alone doesn’t measure true demand. Local out-of-stocks are written into the sales record alongside historical weather patterns and events, creating future manual work for demand planners, such as holiday and cultural events like Easter, or trying to “correct” last year’s wet-and-washed-out summer. This problem is compounded when daily profiles are applied to a baseline forecast before manual adjustments are made. 

Weather only repeats itself 15% of the time, but it’s often used as an influencing factor for demand.  We shouldn’t be surprised that demand planners spend a considerable amount of their time tuning algorithms then adjusting the output to something that looks more “correct,” based on their experience. The problem with experience, especially in today’s fast changing market, is that no-one really has relevant experience that they can draw on.  

McKinsey research tells us that the past is no longer a guide to future behaviour, as consumers increasingly shift channels and are looking for new brands and more convenient experiences. Forecasting capabilities that look backward to predict the future are fundamentally flawed. They struggle to adapt to changing shopper behaviour and cannot measure uncertainty without making assumptions. They leave demand planners exposed to inventory risk. 

Embrace Uncertainty in Demand Forecasting 

Blue Yonder corrects these flaws by forecasting via a unique and more accurate approach. Our demand forecasts are built on machine learning analysis of the relationships between many different historical data sources, such as weather, special events and price. We don’t 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 customer behaviour changes, the model is able to self-correct. Configurations like 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 advisor. 

Our unique demand forecast predicts the full spectrum of demand, not just the mean, and it calculates the probability of every unit of demand at the item, store, day level rather than assume a shape or profile. This information becomes valuable when you later want to make inventory decisions, recommend price changes or make assortment changes. Irrespective of how close your mean might be to historical sales, there is always a chance that customers 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. Fresh produce demand can vary widely across the week as the weather changes. Promotional demand varies depending on pay day, local competition and the season. And, of course, seasonal products like BBQ accessories and clothing can depend disproportionality on weather. Chasing volatility manually cannot be accomplished with the precision required to make a difference at the point of purchase. Planners might plan at a cluster level, but shoppers act locally.  

Accurately understanding this complexity at the local level is extremely useful for creating better demand predictions but automation is not possible without trust between humans and the machines that serve them. The complex reality captured by our forecast engine is presented to humans in a meaningful way, with data grouped in buckets that make sense in plain language. Weather, promotion or trend rather than temperature or shelf position, breaking the black box problem that plagues most machine learning solutions.  

By adopting an automated, self-learning demand forecast like Blue Yonder’s, planners no longer need to tune algorithms or adjust forecasts to account for changing weather or promotional calendars. The forecast can be largely automated, leaving planners to focus on system-generated exceptions, and collaborating with others on how to best use more intelligent and accurate forecast to drive better results, like getting the right amount of inventory where it is shoppers are likely to want it.  

This re-imagined approach is all very simple to the user. A lot like my Sunday afternoon when I can skip giving the bike a bath and focus on more productive things.