In the last two posts in the Retail Forecasting Playbook series, I have talked about measuring forecast accuracy. Now, let’s turn our attention to improve your forecasts.
Below are five steps to help improve the accuracy of your forecasts. This is a good process to use when first implementing workforce management software. It’s also a good process to use periodically to fine tune your forecasts.
Step #1: Clean Bad Data
The principal of Garbage In, Garbage Out applies to forecasting as it does with many things. In this case, the “Garbage In” is the historical data used by the forecasting engine to produce the forecast. In fact, bad historical data is by far and away the most common cause of forecasting errors. Only by ensuring that you have a nice clean supply of data going into the forecasting engine can you expect to get decent results out.
Cleaning involves correcting errors with your historical data. Errors get introduced into your data in a many ways. You need to find a way to wash the dirty data (wrong values) out of your data and replace them with the correct values. This often involves finding a second set of data to compare against the data you’re using.
When cleaning dirty data, you should not make up data, estimate what the result should have been or delete the bad values. If you know what the right value should be, then you should correct the error. If you don’t have the right answer then leave the wrong error alone. Take care of the using Special Days (see Step #2 below) or, if possible, let the forecasting method take care of it.
The best way to avoid dirty data is to use a source of audited data. Unfortunately, except for sales, audited data is often not readily available. Data for certain drivers you need will not be audited (I see this with traffic data, for example); it may not be available at the granularity you want; and/or it may take time to produce those audited numbers and the data not be available when you need it for forecasting purposes.
One last note on audited data: just because you’re using audited data, it doesn’t mean that errors won’t exist. Instead, it usually means that there will be fewer errors and that the errors have been accepted by someone in authority.
Step #2: Use Special Days
Data cleaning corrects errors in the data, but what do you do with data anomalies? What happens when sales spike because of a huge sale in the off-season or when stores close because a snow storm buries the northeast?
These anomalies aren’t errors, and can’t be “cleaned”. Instead, you should use a feature found in virtually every WFM system called “Special Days” or “Special Events”. This feature gives you the ability to mark certain calendar days as “special”.
Marking a day (or days) as special prompts one of two responses from the forecasting engine. Either the special day is ignored by the forecasting engine or the forecasting engine uses other, related special days to properly forecast the unique event.
Special Days is a very powerful tool that is often misused, and misuse of this powerful feature usually hurts accuracy more than helps it. I’ll examine the proper use of special days in an upcoming blog post in this series.
Step #3: Change when Forecasts are Created
The timing of the forecast – when the forecast is created relative to the date of the forecast – has a significant impact on the accuracy of the forecast. More specifically, creating a forecast for one week from now will produce a more accurate forecast than one created for four weeks from now. Therefore, forecasts should be created as close to the dates they represent as possible.
Ideally, you would create your forecast the day before you need it. Unfortunately, that is not practical. With WFM, the timing of your forecasts is driven by when you need to post your schedules for your associates relative to when the schedule starts and how much time you want to allow for reviewing forecast and editing the schedule.
While the timeline varies from retailer to retailer, most retailers post schedule 3 to 10 days before the schedule starts. Add-in another 2 to 3 days for schedule editing, and you quickly see that the shortest window between when the forecast needs to be finalized and the dates it represents is 6 to 13 days.
If you’re generating forecasts more than two weeks in advance and your running into forecast accuracy problems, you owe it to yourself to do some analysis to determine if a more current forecast will improve the situation.
Step #4. Change the Granularity of the Forecast
The Law of Large Numbers tells us that with statistics bigger numbers produce more accurate results than smaller numbers. Applying this law to forecasting, this means that larger numbers will produce more accurate forecasts than smaller numbers. Given that, if you’re still struggling with forecast accuracy by Step #4, you need to create larger numbers for your forecasting engine to work with.
You create larger numbers by changing the granularity of what you’re forecasting. Rather than forecast by product line, consider forecasting by product category or department. Rather than forecast by quarter hour, consider forecasting by half hour or by hour.
Changing the granularity of the forecast can be a complex undertaking, especially if made to a production environment. It often involves changing the data source and/or import process. It may involve rolling up labor standards. It could also require changes to drivers and/or the store hierarchy. As such, changes associated with this step should be well thought out and well planned. However, improvement to your forecast accuracy can be well worth it.
Step #5. Consider Changing the Forecasting Method
Until this point, we improved the quality of the data, accounted for data anomalies, improved the timeliness of the data, and increased the size of the data. In short, we’ve focused solely on the data. The last step to improve forecast accuracy is to look beyond the data and look at the forecasting method.
For many people, this feels backwards. The natural reaction is to look at the forecasting method first – to ask how the system can solve the problem. Unfortunately, that Garbage In, Garbage Out principal that I mentioned in Step 1 makes any such focus fruitless. Until you have clean data upon which to base your forecasts, you’ll never have accurate forecasts regardless of which forecasting method you use.
Also, notice that this step is not called “Change the Forecasting Method.” Instead, I simply suggest that you consider changing the forecasting method. Chances are you’ve already done some analysis to determine that a certain forecasting method is the right one for a particular driver, and at this point, you should validate that your initial analysis of the various forecasting methods available to use was correct.
If it was correct, the only reason why another forecasting method would be better is that something has significantly change with your data. This change could be due to a change in your business, or if your results in the previous four steps were widespread enough, it could be due to your efforts to clean and prepare your data for the forecasting process. To determine if a change to your forecasting method is warranted, you’ll need to redo your analysis using a current set of clean data.
After completing these five steps, your forecasts will be as accurate as possible given the limits of your data, business processes and available forecasts. If you have any questions about these five steps and how to improve your forecasts, feel free to drop me an email.
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.
