What is the Scale of Your Forecasting Process?:
Trends 30-Second Survey Results

The survey on scale of the forecasting process asked six questions; twenty-three surveys were completed. The questions were:

1) The lowest level of detail at which I forecast is (choose from list):
        -Product Group
        -Item (i.e., SKU or Part#)
        -Item by location
        -Item by customer
        -Item by customer location (i.e., "ship to")
        -Other (specify)

2) At the lowest level of detail, the total number of forecasts I make is (choose from list) (Note: For     instance, if you forecast at item level how many items do you have? Or if you forecast at item-location     level, how many item-location combinations do you have?, etc.)
        -Fewer than 500
        -Between 500 and 5,000
        -5,001 - 10,000
        -10,001 - 25,000
        -50,001 - 200,000
        -200,001 - 500,000
        -500,000+

3) What periodicity (time buckets) do you forecast in primarily, and what is your typical forecast horizon?

4) How much history do you typically use?

5) Are you using a forecasting system (e.g., dedicated forecasting software, demand planning system,     ERP system, etc.) to generate baseline statistical forecasts at the detail level? If yes, how long does it     typically take (in minutes) for the system to generate the forecasts?

6) Which version of Windows is your company using?

7) If your forecasting software runs on a server, is it (choose from list)
        -Client Server Environment
        -Terminal Server
        -Citrix
        -Remote Login
        -Web Login
        -Don’t Know

Additional open-ended comments were also solicited. Click here to see the original survey.

Survey Results

For 52% of the respondents, the lowest level of detail they work at is at the item level. Collectively, those who work at a more granular level make up 40% of the respondents, and are mixed in terms of the dimensions that drive granularity (i.e., customer, location, customer location).

Two-thirds of the participants in the survey report making 5,000 or fewer forecasts, while roughly 25% report making 50,000 or more forecasts. Interestingly, the survey shows that forecasting at a more granular level does not automatically equate with higher scale. Of those who reported that they generate 50,000 or more forecasts, half forecast at item level. Further, the single respondent who reported making greater than 500,000 forecasts, does so at the item level.

While the number of items is one way to define scale in forecasting, another is to consider the total number of data points involved. For illustration, the table below shows the impact of periodicity on total data points. Each example has 100 items with three years history and one year of forecasts, the only difference being the time buckets.

In the survey, almost three-quarters report that they forecast primarily in monthly time buckets. Weekly time buckets are the next most prevalent, followed by days and then quarters.

Of those using monthly data, two-thirds used at least 24 months of history. One-third used less than 24 months of history. This is interesting, especially given the fact that two full cycles (24 months with monthly data; 104 weeks with weekly data, etc.) is an important data threshold when trying to adequately capture seasonal patterns. Surprisingly, of those who respond that they use weekly data, the maximum number of weeks of history used is only 36 weeks.

For those using monthly data, the most typical forecast horizon is 12 months. This makes sense; many companies maintain a rolling 12-month forecast which is updated each month. All respondents who forecast in weeks use a 4-week horizon.

By combining the breadth (i.e., number of items, number of item/customer combinations, etc.), the number of periods of history, and the length of the forecast horizon, the full scale of the data forecasters work with can be ascertained. The table below shows calculations of scale for each of the respondents in the survey and ranks them from largest to smallest.

An easy way to conceptualize the scale of data is to think in terms of a spreadsheet view. For the top-ranked respondent above, the dimensions of a spreadsheet containing their full history and forecasts would be 500,000 rows by 66 columns—33 million data points.

Most (70%) of those responding do use a forecasting system to generate baseline statistical forecasts. When asked about their system’s processing time to generate the statistical baseline forecast, it is apparent that processing times vary widely. As shown in the table below, for one respondent, generating forecasts for 5,000 items using 36 months of history and a horizon of 12, the process takes just one minute. While, for another respondent with the same data parameters, the process takes 120 minutes.

The respondents were asked which version of Windows they used. One of the most interesting findings is that none of the respondents report using Vista. Another, is that Windows 7 appears to be becoming prevalent, with 48% saying they are using it. Whether the operating system has an impact on processing is entirely unclear. For instance, the two respondents referenced in the previous paragraph—each forecasting the same scale but with drastically different processing times—both report running XP. In addition, of those with the longest processing times, two are running XP and two are running Windows 7.

Although every participant is aware of which version of Windows they use, only half of those whose forecasting software runs on a server knows which server environment they are using. Of those who know there was an even split between Citrix, Client-Server and Web Login.

Click here to download the survey slides as a PowerPoint slideshow

Click here to download the survey slides as a PDF document