Blogs

Axsium Blog Post
How to Use Special Days to Improve Forecast Accuracy
Bob Clements
Wednesday, July 6, 2011

There is a lot of activity in my household right now as my family gets ready for the Nordstrom Anniversary Sale which starts in a mere nine days.  It’s funny how much time my household gets planning for a department store sale, but its become an annual ritual. 

As crazy as I think we are, I’m always reassured (and relieved) when I show up at Nordstrom on the first day of the sale to see that we’re not the only family that does this.  Inevitably, I gawk at the crowds that fill the store in the middle of July when the main competition is great, summer weather.  I watch the customers bee-line for what they planned to buy and then buy on impulse that accessory that they didn’t know was going to be on sale.

As a retail geek, the traffic and volume make me think about the spike in revenue that Nordstrom gets from this sale and I wonder how it compares to last year.  I think about how Nordstrom tweaked the sale from last year by changing the available merchandise or by the way it was promoted to make it more effective.  As a workforce management geek, the Anniversary Sale also gets me thinking about Special Days.

Special Days (with capital letters) is a feature in virtually every workforce management forecasting system.  Depending upon the system, this feature may be called “Special Events”, “Holidays and Events” or the like.  Regardless of the name, Special Days is intended to help you improve the accuracy of your forecasts.  How does Special Days do this?

As you know, a forecasting engine is going to use historic results to predict future business volume.  Therefore, accuracy depends largely on how consistent your business is over time.  For most retailers, their business is fairly consistent from year to year.  Generally, Friday in one fiscal week in 2011 is going to feel very similar to the Friday in same fiscal week in 2010.  However, if this year’s Friday falls on a holiday or during a major sale, or if the store was closed last year due to a hurricane or blizzard, then this Friday and other Fridays cannot be compared. This is where Special Days comes in.

The forecasting engine isn’t intelligent.  It doesn’t know when these things happen unless you tell it.  Special Days gives you the ability to tell forecasting engine that something unusual happened and tells it how to account for that unique event when it occurs again.  This is done via configuration in most WFM systems.

There are typically two ways to deal with a special day.  First, you can tell the system to ignore the special day and not use it as part of the historical data used to generate a forecast.  This is useful for situations like emergency (or at least unexpected) store closures or system problems in which the data for a particular day is not representative of a normal day.  Second, you can tell the system that the special day is an unusual recurring event such as a holiday or sale, and that when encountered, to look for previous occurrences of the special day to build the forecast.

The problem is that in most WFM systems Special Days must be manually identified and configured.  And, unfortunately, most retailers use Special Days incorrectly and forecast accuracy suffers rather than improves.  The biggest mistake retailers make is that they over use Special Days.

Retailers have a tendency to think of any unusual day as special.  In fact, a typical store will often have fewer than two dozen special days (and often closer to a dozen) per year.  The challenge is understanding which days are truly special and which ones are not.  There are two basic steps to making this determination.

First, the event needs to have a measurable impact on business compared to a normal day.  That measurable impact is just not an increase or decrease in volume.  Rather, it needs to be an unusual jump or drop in volume.  This could be determined, for example, by finding the standard deviation of volume by weekday and looking for outliers in the last couple of years.  These outliers provide a clue to days that may be special.

Second, the event needs to occur on a different fiscal day than it did last year.  Just because sales or traffic falls outside of the norm for a particular weekday, does not mean that it is a special day.  If the same thing happened last year on the same fiscal day, the forecasting engine will likely pick up the trend and apply it to future forecasts automatically.

Going back to the Nordstrom Anniversary Sale, there is no doubt that when the event launches on Friday, July 15, it will generate significantly more sales than the previous Fridays in July.  Heck, it will probably generate more sales than most Fridays in the year!  However, the sale also starts on the same fiscal day as it did last year.  This means that Nordstrom would not have to configure the Anniversary Sale using Special Days.

Next year, if Nordstrom were to change the dates of their Anniversary Sale (I’m not suggesting they will.  This is only an example.  I have no inside knowledge or anything like that.), then they would need to go back and mark the Anniversary Sale using Special Days so that the forecasting engine would use the correct days when predicting volume for the Anniversary Sale dates as well as those dates that no longer fall on the Anniversary Sale.

A significant sale is not the only time to use Special Days.  Holidays are also great Special Days as they tend to generate similar traffic year to year but because they fall on calendar days, tend to jump around the fiscal calendar.  Again, the key is to ensure that the holiday has a meaningful impact on volume.

The last thing to note about Special Days is that their specific behavior will vary based on the forecasting algorithm being and the WFM system used (not all WFM system implement the same algorithm in the same way).  As such, check your documentation to ensure you understand the specifics of your situation.  I also suggest that you test Special Days on a handful of representative stores before you decide to deploy the feature to ensure that you get the results you expect.

This post is part of the Axsium Retail Forecasting Playbook, a series of articles designed to give retailers insight and techniques into forecasting as it relates to the weekly labor scheduling process.  For the introduction to the series and other posts in the series, please click here.