Understanding Seasonality in Demand Forecasting
A crucial aspect of understanding seasonality is recognizing that sales patterns often labeled as "seasonal" can actually originate from different sources. Having this distinction clear helps set realistic expectations about what a forecasting model will likely capture automatically and where additional adjustments may be beneficial.
1. What should be considered Seasonal?
Some demand patterns are genuinely seasonal, emerging from recurring events each year. These predictable patterns stem from well-defined, time-based sales events, such as:
- Events: Certain product categories often see demand spikes around specific events. For instance, Mother’s Day could consistently drive up sales in categories like handbags or blouses, while Father’s Day might increase demand for products like men’s shirts or accessories.
- Seasonal Clearance Events: Annual sales events, such as Summer or Winter markdowns, often follow consistent timing and discounting strategies. Since the types of products, discounts and timing of these events remain relatively stable year over year, they can be considered predictable.
2. What should not be considered Seasonal?
While some sales patterns might seem to recur annually, they don’t fall under true seasonality and should not be classified as such. Mislabeling these patterns as seasonal can lead to inaccurate expectations of the forecasting model's capabilities. Here are two common examples:
a) Aggregated Sales Increases due to Product Range and/or Store expansion
Increases in aggregate sales for a product category during a season don’t always indicate that each product within that category is experiencing higher demand. Often, this growth is due to more products being introduced to the category or a wider distribution of these products in stores.
For example, consider knitwear during the winter season. While total sales for knitwear might appear to rise between September and December, it may simply be because more styles were introduced, or because some of these styles started to be expanded to more locations as the season progresses:
In this case, the total sales increase experienced by the Family Knitwear would not be due to an increase in the actual demand for each knitwear item but to an extension on the availability of this type of products (more styles, more locations).
As a consequence, If we were to treat this as a seasonal trend, we would incorrectly apply upward demand trends to each knitwear product, when, in reality, the average sales per item may remain stable during September, October and November.
b) Sudden, unpredictable sales spikes due to circumstantial factors
Some sales patterns are irregular and non-recurring, making them inherently unpredictable and thus unsuitable for seasonal forecasting.
Let’s try to illustrate this with an example: if some of your stores had an appliances section, you could see how items like power banks may sell steadily on most days. However, if there’s an unexpected influx of tourists—such as from a large cruise ship arrival—sales might surge temporarily.
This type of demand spike is difficult to forecast, as it depends on random events that don’t follow a regular pattern. In these cases, forecasting models cannot be expected to reliably capture these sales. Instead, these items may be better managed by keeping a minimal level of stock across the network.
3. Identifying seasonal patterns: minimum requirements for effective forecasting
When considering cases that could be classified as seasonal, and therefore potentially predictable, it’s essential to understand the minimum requirements needed for forecasting models to better capture these trends automatically.
- Understanding the actual relevance of the event
The first step is to assess the significance of the event to the business. -
- If it impacts only a small group of locations, the signal is likely to be too weak for the model to detect it automatically. Bear in mind that low data density can translate into the event’s impact going unnoticed by the model.
- Similarly, if the event doesn’t affect an entire product family or category, but only a small subset, the signal may also be too weak for the model to identify reliably.
- Reviewing historical data consistency
Another important factor is whether historical data consistently shows an uplift across the same locations for the average sales of the product family each year.
If some years there is an uplift while others are not, the model may conclude during its training that the event is not seasonal and will likely not attempt to reproduce it in future forecasts.
Also, if every year there has been an event tagged the first week of March and in 2025 suddenly there is none, the model would have never seen what happens with the sales pattern during the first week of March when that event doesn’t exist… which means that the behaviour of the model can not be predicted in advance.
- Ensuring relevant product attributes are captured
In cases where it’s unclear if the model will be able to capture a specific product’s uplift, it’s also useful to validate whether the attributes shared via the Product Master can help the model identify that the product will behave in a given way.
For instance, if you have a New Year’s Eve collection, tagging those items as “New Year” in the Product Master every year could help the model recognize that the products associated with this tag tend to experience an uplift in November and December and therefore keep reproducing in the forecast of the upcoming years. However, remember that this uplift needs to be significant as noted in point one!
Here you can see an example of a product where, even if potentially considered as seasonal, neither its historical data nor the attributes reported could help the model forecast a robust demand for 2024:
In this other example you can see how some times, even if no particular attributes have been reported, the historical data is robust and consistent enough to make it possible for the model to capture its upcoming trends:
4. Validating your biggest events in Nextail
While these are valuable guidelines, it’s essential to remember that, due to the complexity of machine learning models, setting firm rules is challenging - and might lead to confusion or even fake expectations!
To avoid high levels of uncertainty when operating the system, we offer our customers the option of providing us with their top 4 or 5 key events. Once these events are uploaded to the platform, our data science team can be requested to analyse whether the model is able to capture them with existing variables or if certain events would be better managed through promotions or other operational adjustments.
