The COVID-19 pandemic has impacted retailers in different ways. Some stores were deemed essential businesses and remained open, while some were forced to close. Additionally, many stores that stayed open saw dramatic changes in consumers’ buying patterns, including an increase in online purchases.
Regardless, these changes impact future volume forecasts — predictions of how much business is going to be done at a future date. Volume drivers are the source of the volume forecast. In fact, there are multiple factors to consider when calculating the volume forecast, including historical POS data, forecast rules, and adjustments to predicted volumes for special events.
With that in mind, I sat down with two forecasting experts on our product management team, Nancy Lord and Andrea Davis. Their insight and knowledge will provide you with guidance as they share their recommended forecast model as well as best practices for how to treat special events during and after the pandemic. Ensuring that your forecast gets back on track will depend on far more than merely treating COVID-19 as a special event.
Guidance when using special events
Let’s start with what constitutes a special event. Special events are typically used to identify unusual past or future business activity and are grouped into the following categories:
- Known major holidays — Valentine’s Day, Memorial Day, Thanksgiving, etc.
- Milestone-type events that are recurring and annual — Graduations, summer vacations, back- to-school promotions, etc.
- Unexpected special events — Bad weather, closed stores, etc.
Past activity
Please note that the unusual business activity or store closures resulting from the pandemic should not be marked as special events. Our forecasting experts recommend that you do not discard the entire period that your stores are closed or mark any day during the pandemic as a special event. Doing so will result in traditional algorithms, like daily trend and exponential smoothing, not having enough data to create a forecast.
Future activity
If you want to create an event for an upcoming holiday, such as Father’s Day, you should create a new special event instead of tagging the day with an existing event. For example, consider creating Father’s Day 2020 as a new event so you can distinguish it from existing Father’s Day special events. These are not normal times and using the existing event could skew your data for Father’s Day events overall.
An introduction to your available forecasting algorithms
Depending on your Kronos solution, there are a few options when it comes to choosing your preferred forecasting algorithms — daily trend, adaptive forecasting, exponential smoothing, and machine learning. To start with, both Workforce Central® Forecast Manager and Workforce Dimensions™ Forecasting include the daily trend and exponential smoothing algorithms. There is an adaptive forecasting model within Workforce Central Forecast Manager that was replaced in Workforce Dimensions Forecasting with an enhanced and improved machine learning model. The machine learning model in Workforce Dimensions Forecasting creates a significantly better forecasting model — the result of the distributed computing power that exists within the Kronos D5™ platform. Simply put, the Workforce Dimensions D5 platform makes doing awesomely complex algorithms an everyday reality.
Let’s dive into the forecasting algorithms.
Daily trend
Daily trend is a mathematical model that incorporates historical data from both current and prior years to capture cyclical variations that are common in the retail environment. This algorithm is built into both Workforce Central Forecast Manager and Workforce Dimensions Forecasting products. The concept with the daily trend algorithm is that retailers can typically expect that the day of this year will look very similar to an equivalent day from last year. Moreover, when developing a forecast in anticipation of future volume, you can expect that the same day of this week will look like the same day for next week, barring any special events falling on that day. Put another way, the daily trend algorithm is an adjustment of the same day of the week last year based on a linear trend. This algorithm is helpful to use when trends in historical data are steady year over year.
Our forecasting experts do not recommend using the daily trend method for forecasting during COVID-19 for the obvious reason that we were not experiencing a global pandemic at this time last year.
Adaptive forecasting
Adaptive forecasting identifies recurring seasonal trends in the historical data and considers how business volume fluctuates over a year. Within the Workforce Central product suite, this algorithm is often used in conjunction with the daily trend and exponential smoothing for a more dynamic volume forecasting method. However, it’s important to note, that the adaptive algorithm may not find a significant trend in the data when there are many outliers. In fact, this model is best used when seasonal patterns repeat year over year.
Therefore, the product team does not recommend using adaptive forecasting for forecasting during COVID-19 because the extremity of this pandemic does not equate to any other seasonal activity.
Exponential smoothing
Exponential smoothing is a statistical model that analyzes the current year volume history to determine upcoming forecasting values. Additionally, it can be used with as little as three weeks of historical data. This model is less adept at anticipating cyclical fluctuations, such as higher volume in December due to the holidays and slower activity in January.
This model, available within Workforce Central Forecast Manager and Workforce Dimensions Forecasting, is best leveraged for situations when there is little available data. For example, a recent implementation limits the amount of historical data to just a few weeks. In this case, using the exponential smoothing algorithm to derive a forecast using complex statistical analysis is actually more accurate than using the average of historical data for that short timeframe. Exponential smoothing applies a “smoothed” or weighted average to the past N weeks (same day of the week). There are benefits to using this model when there is little data available or when the past year’s data is unrelated to current events.
Machine learning
Then, there’s machine learning within Workforce Dimensions Forecasting which uses many features of historical data and business structure to create a finely tuned formula. This is valuable for complex situations where the past is prologue — meaning that you can look to what has already happened as an indication of what will happen next. The beauty of this algorithm is that you define the feature configurations that control the behavior of the forecasting engine during a volume run.
Exponential smoothing and machine learning are the recommended algorithms to use during COVID-19
In collaboration with the Kronos data science team and members of Kronos engineering, our product experts, Nancy and Andrea, recommend using the exponential smoothing algorithm during this time. The data science team analyzed customer data and determined that exponential smoothing follows the current business behavior more accurately than daily trend forecasting does. In fact, Workforce Central and Workforce Dimensions customers should continue to use exponential smoothing until your volume has stabilized.
Additionally, the data science team conducted studies using the machine learning algorithm in the April and May time periods. The results showed that in cases of high volatility, the machine learning algorithm forecasts were similar to the forecasts produced using exponential smoothing (utilizing recent averages and trends). Moreover, machine learning performed slightly better in many cases, so customers that are using this algorithm should continue to do so. However, it’s important to note that retraining the machine learning algorithm is not recommended at this time, since there is not enough data for capturing what “normal” looks like.
In other words, here’s an analogy from Adria:
Within my Pandora account, I have created multiple (and musically diverse) stations based on my moods. In fact, it’s taken me years and thousands of thumbs up and thumbs down to “train” my stations so that only the songs that fit my preferred mood are played on the selected station. Using the machine learning algorithm during a global pandemic would be like trying to get my Cher station to start playing Bjork within the hour — a near impossible task because solid modeling takes time and accurate data. Plus, you don’t mess with Cher.
The product management team continues to work with our data science engineers to run tests and determine the applicability of when training the machine learning algorithm will be necessary.
Where to learn more – Workforce Central users
- Workforce Central Forecasting Guidance to Manage the Impact of COVID-19
- Workforce Forecast Manager Configuration and User Guide