Demand forecasting augmented by machine learning helps you better meet customer expectations with reduced inventory investment.
Most companies view seasonality as a pattern of demand that has regular and predictable change occurring every calendar year, like beer sales rising in the midsummer months, and cold medicine in the winter. This foundation has been the basis for automated replenishment systems for many years, but like everything in supply chain today, it’s more complex than that, and the traditional replenishment approach has a lot of “noise” in how it’s calculated.
One difficulty of using a demand pattern from year to year is that the demand is a sum of variables occurring daily. Another difficulty is that, of course, not all seasons start and end on specific dates. Demand can shift over time to trend higher or lower based on other items, markets, and media events. Finding the basis of a proper seasonality index is further hampered by other causal effects such as patterns of weather. The ultimate impact of a good–or poor–understanding of seasonality is on finding optimal inventory levels to balance cost and service.
The root cause of this anomaly is that most demand forecasting systems look at the historical aggregate demand instead of smoothed demand potential as their basis. Aggregate demand has the noise of causal events. One example is weather. Weather-dependent seasons really are about how far from the normal temperatures each day is. Rain can impact consumer demand at retailers. These deviations from the norm can cause a lot of noise that makes its way into the seasonality index. And to further complicate the approach, seasonality is not a global value, but applies to very specific locations and each location can have differing indexes for individual items or groups of items. Other causal factors are such things as media events that might focus attention on a specific product or category.
So how can the day-to-day demand anomalies be removed from the seasonality index to create a better baseline for future seasonality forecasting? The answer comes with being able to identify and track each causal effect. Data such as how hot or cold, the variance in precipitation, how strong the wind was, and other metrics play a part in why consumers shopped more or less on any given day and what would have happened if the causal events were at their normal metric.
Once a smoothed baseline of potential demand is calculated, then future seasonality forecasts better meet customer expectations with reduced inventory investment and better confidence levels. Using this method provides better service, less disruption, and even reduces inventory needs during a normally highly volatile period.
It’s true that identifying and analyzing all these causals and their effect on future demand would be impossible for planners to do on spreadsheets. Fortunately, “next-generation” forecasting, augmented with machine learning, helps us humans do what was previously impossible.
Machine learning is one of the most powerful technologies being applied to achieve market-driven forecasting. Seasonality is one of the most popular use cases we’ve seen, as well as promotion forecasting, new product introduction forecasting, and others. Planning systems that apply machine learning do indeed “learn,” which makes them grow better at predicting demand over time. Not only do these systems learn from a wide range of demand and historical data, but they also assimilate the knowledge and experience of demand planners and others involved in the planning process. These smart systems elevate planners, allowing them to focus on the most productive and the highest value-added activities.
Whether it’s heated hand grips for motorcycles, flu medicine, furnace parts, or Christmas sweaters, the pre-holiday and winter season is the perfect time to gear up to make sure your products are available to customers when and where they need them. This year, why not give your planners the gift of AI-enhanced forecasting?