One of the main reasons why companies forecast demand is to ultimately understand how their business will perform in the coming period and how to staff to ensure customer service levels. This way, they can avoid over- or understaffing when creating and filling shifts to cover that demand, saving cost and avoiding lost sales. The trick is: accurate demand forecasting isn’t easy. After all, each of your stores has its own context: local events and seasonal effects. A single store probably has multiple key metrics that drive demand. Each of those key metrics will have its own demand patterns and trends, and needs its own best-fit forecasting model to ensure the highest accuracy. Putting it all together, you can see where you quickly arrive at lots of metrics to forecast and lots of modeling approaches to manage. How it’s possible to achieve this level of specificity on a large scale, and why should you even try?
Example case: Patty’s Pizzas
Let’s take an imaginary pizza chain called Patty’s Pizzas as an example. Patty’s Pizzas has a thousand restaurants spread around the world. The difference in context between locations matters. For instance, demand patterns for Patty’s Pizzas New York will be different than Patty’s Pizzas Austin. The restaurants differ in size and might not offer the same range of products. Additionally, New York and Austin differ hugely in population, which translates to more deliveries in one location compared to the other. Weather and seasonality cause different patterns too; Austin rarely has snow in the winter while New York is snowed under at least once a year. Holidays and local events are two more variables that explain why labor demand forecasts for both restaurants will be very different, and the list of variables can go on.
This specific context makes it very difficult to use one forecasting method across the entire business. Historical averaging and other one-size-fits-all approaches are often not accurate enough to be useful and require managers to overlay their personal experience to get a reasonable estimate of demand over time.
Introducing: hyperlocal forecasting
Hyperlocal forecasting is an AI-solution that takes the context of each demand driver within each individual location into account to make accurate demand forecasts on the most granular level. If we think of Patty’s Pizzas, this approach generates a separate demand forecast for each individual data driver (e.g., transactions, total items sold, pizzas, drinks, deliveries and sales revenue) for each restaurant. This means that both New York and Austin locations have their specific contexts taken into account when generating the forecast so that the forecasts and resulting schedules take into account local events, seasonality, trends, weather patterns, etc.
One of the advantages of hyperlocal forecasting is that it allows your forecasts to be much more accurate. Instead of having a general idea of what your demand looks like in a given month, hyperlocal forecasting algorithms will identify, for example, that Patty is likely to sell more pizzas the first three days of November in her New York restaurant. It will also show that there will be a steep spike in foot traffic and less deliveries compared to other specific days. In this case, Patty can align her staff by knowing she’ll need less drivers and more cleaners on those days. Now where do these sudden spikes in demand come from? This is where event management comes into forecast accuracy. If Patty were to analyse those days, she would understand that the New York marathon is held that day and that’s why the forecasted peaks are expected. This might sound obvious, especially when such event is an annual occurrence. However, simplistic methods commonly overlook these types of events, and rely on manager overrides to take them into account. Using hyperlocal forecasting algorithms equipped with machine learning methods to crunch all existing historical data will significantly lead to higher accuracy results when compared to results from averaging sales of the last three months. Events are less likely to be overlooked and the impact of each event on every demand driver is better quantifiable.
How to realize hyperlocal forecasting
In order to make accurate forecasts per demand driver per store, different forecasting methods have to be applied. What could be the best method for one driver, doesn’t necessarily have to be the best method for another driver. This implies that in order to determine which method works best for a demand driver, results from all methods for that demand driver have to be compared and the best one has to be saved. That’s why hyperlocal forecasting algorithms are equipped with a mass-customization functionality. This feature automatically trains, selects and saves the best methods at huge scale and is essential for businesses that want accurate forecasts on the most granular level. That is: per demand driver per 15 minute buckets! Enterprise companies benefit from mass-customization because it allows them to easily scale, making specified and unique forecasts for each individual part of their organization, no matter the size.
All the algorithm requires is historical data. The more data, the better. If a store has a bunch of data, the algorithms learn from the data, and produce accurate hyperlocal demand forecasts that no traditional technique can compare to. Experience and gut feelings will be minimized and with hyperlocal forecasting at their disposal, new store managers will have highly accurate forecasts right away.
Hyperlocal forecasting: producing the most accurate demand forecasts for your retail and hospitality stores. It’s fast, it’s accurate, and with the right tools, it quickly scales across stores. It’s the foundation for cost-effective and efficient workforce management.