As everyone in the consumer goods industry knows, demand forecasting is a demanding business. A serious attention to detail is required to achieve best-in-class accuracy levels.
In The No-Nonsense Guide to Forecasting, we discuss data preparation as the foundation for building a successful forecasting model. This time-consuming but indispensable process necessitates exactitude. Messy data sources like sell-through units and out-of-stocks must be collected, cleaned, and harmonized at the most granular level possible.
Once the most up-to-date data has been prepared, you’ll need to develop a baseline demand forecast, at the heart of which is seasonality. It will give you insights into how sales vary over the year, allowing you to adjust your decision-making process for both the long- and short-term.
While seasonality is key to baselining, implementing this important factor for predicting demand is a tricky proposition. In this post, we’ll examine the various types of modeling methodologies, their potential limitations, and proven solutions for meeting the seasonality challenge.
There are three common seasonality types: yearly, weekly, and monthly.
- Yearly seasonality encompasses predictable changes in demand month over month and are consistent on an annual basis. For example, the purchase of swimsuits and sunscreen prior to the summer months and notebooks and pens leading up to the new school year.
- Monthly seasonality covers variations in demand over the course of a month, like the purchasing of items biweekly when paychecks come in or at the end of the month when there’s extra money in the budget.
- Weekly seasonality is a characteristic of more general product consumption and reflects a host of variables. You may find that consumers buy more (or less) of different products on different days of the week.
It’s not always obvious whether a specific time series exhibits seasonality. If that’s the case, you can perform hypothesis testing on your data to determine if there’s seasonality. Another option is to take a more empirical approach and check whether models with or without a seasonal factor are more representative of your data by using a measure like the Akaike information criterion.
Seasonality will typically vary from region to region, depending on local calendars and weather, so you may need to segment your data geographically to identify seasonality patterns. At the same time, aggregating forecasts to a less granular level—product category instead of product, for example—may make it easier to distinguish seasonal patterns from random noise.
Most forecasting methodologies allow for explicit modeling of a seasonal term. SARIMA (ARIMA with seasonality) allows for forecasts based solely on the past values of the forecast variable. And the Holt-Winters seasonal method comprises a forecast equation and three smoothing components for the level, trend, and seasonal components.
However, these models do have their limitations, too.
Some models don’t allow for multiple seasonality, like both weekly and yearly. If you're using one of these models, depending on the nature of your data, you can choose to include only the most critical type of seasonality or go with seasonality as a categorical variable. Other options range from modeling the cyclic nature of seasonality by using Fourier terms as regressors to utilizing a model that can handle multi-seasonal time series.
Then there are some models that assume integer seasonality, which is not always accurate. There are 365.25 days in a 12-month period due to leap years, and the length of a month fluctuates throughout the year.
What’s more, some seasonal events may not be tied to a calendar date (Lunar New Year, Thanksgiving, etc.). Therefore, instead of using a seasonal term, you could model these as binary regressors using holiday calendars.
There are also times that basic seasonality is not actually the most relevant factor in the forecast, which may be covering up the true underlying driver. Because of that, it would be more beneficial to incorporate regressor variables in the model instead.
Here’s an example:
Purchases of antifreeze may have a more direct relationship with the temperature than the precise time of year, as weather patterns waver on an annual basis. With that in mind, using temperature as a regressor variable, instead of modeling seasonality, would generate better short-term forecasts.
However, for longer-term forecasts, seasonality would still be an easier path to follow given the difficulty of getting accurate weather forecasts on longer timescales.
Meeting the Seasonality Challenge
As you can see, there are many moving parts in the building of a successful forecasting model—from data preparation to baselining—to help you better comprehend consumer behavior.
It’s important to note that once you have a solid baseline, the seasonality picture will become much clearer, so it’s worth the investment to get it right. A good rule of thumb: Go back as many years as possible in order to pinpoint reliable seasonality patterns. When it comes to newer products or product categories, you may need to extract those patterns from similar products by leveraging syndicated market data.
Without a doubt, seasonality can be challenging. But intelligent forecasting solutions, like Alloy, can help you make the right methodology moves and avoid the pitfalls.