When to Forecast?

When you forecast is not just a question of buckets (such as days, weeks or months), cycles (how frequently you forecast) and horizons (how far in time you forecast) though they are, indeed significant factors, but also includes a number of other time-related elements that will influence your forecast activity and accuracy. The various factors can be grouped as follows:

  1. Time Bucket: The lowest level of time that you forecast at.
  2. Calendar: The groupings of Time Buckets for allocation, aggregation and reporting.
  3. Forecast Horizon: How far into the future to forecast.
  4. Type of Forecast Length: Fixed or Rolling
  5. Forecast Cycle: How often you generate, tune, review, adjust, approve & publish the forecast.
  6. Forecast Term Focus: Period of Time that receives most attention or tuning.
  7. Frozen Periods: Time where changes are prevented or restricted.
  8. Historical Horizon: How much historical time to retain.
  9. Past Forecasts: Which old forecasts to store for trend analysis and accuracy.
  10. Forecast KPI's: The time selection (& methods & datasets) to measure performance
  11. Business: Your Culture & Tradition, Systems, Data Streams, S&OP, Resources etc. all drive the selections of the previous 9 points.

Buckets of Time

In Forecasting, Time is a Dimension with Hierarchies of Calendar. Calendars have different ways of grouping time and the choice of Calendar can seriously impact (or restrict) how you manage your forecasts. The 3 most common calendar approaches (at least in the Western World) are Fiscal, Gregorian & Manufacturing. The most significant impact is whether you need to have weeks, and if you do, how they are collated.

Example of Calendars

Systems Solutions will often organise the use of 445, 454 or 544 weeks in into Months.

A fourth Calendar option that can sometimes be utilised is the creation a Composite or Business Calendar.

The Composite Calendar can be useful when trying to get the best of Gregorian and Manufacturing buckets of time but er... it can be time consuming to create and maintain.

Forecasting at Day level (especially in a Demand Planning solution) can be extremely data intensive especially when aligned with other very low levels such as SKU and Customer Ship-To. The lowest level of granularity is not always desirable in Demand Planning, learning to raise levels up and simplify can smooth forecasts and speed up cycle times.

How far into the future do you see?

The extent of your forecast horizon should depend on the purpose of the plan being created. It could be anything from seconds to decades. OK, predicting in seconds would be rather odd but I expect people out there in space exploration, nuclear fusion or pharmaceuticals are.

Forecasts will sometimes contain a horizon that is longer than the planners want or need either because the source data had it and/or because the person or system they publish to needs it. Control of this is sometimes termed the Time-Fence. Inside Time-Fence is where the Planners efforts on data will have an impact while Outside Time Fence is for reference.

inside and outside time fence

Ensure the horizons are appropriate for the maturity level of your planners, systems and processes. Flexibility is probably the most useful recommendation here - businesses should aim to quickly switch from short to mid and long forecasting and back in reaction to things like IT implementations or dramatic Customer behaviour.

A mature forecasting solution will be able mix and match the horizon and hierarchy levels appropriately for the forecast purpose: Perhaps days for the immediate forecast, weeks for the short forecast, months for the mid-term and quarters for the far forecast. How flexible are your forecast levels and lengths?

A stickler for time or rolling with it?

Another elemental factor in Forecast Horizons is to consider whether your forecast rolls continuously or ends at a fixed point. Budgets will typically be fixed from one point to another (such as Jan to Dec) while Demand Plans would normally roll forward (dropping the earliest bucket and adding a new one to the end of each cycle).

Mixing and comparing fixed and rolling often creates problems and tension. I wonder if these sound familiar: Budgeting in Excel at grouped level that doesn't exist in the Demand Plan but needs to be imported and compared... the levels and totals don't match. Or the annual Sales Promotion Plan that needs to be created afresh at a set point each year using an extract from Demand Planning and then modified with items & customer that don't yet exist and therefore can't be imported back into Planning.

Excel solutions are very time flexible since they allow you to you create unlimited plans and alter them at any time, but they create data inconsistency and control problems. On-Premise and Cloud solutions can help by standardising Master Data (Calendars, New Products, Conversions, Prices, Customers etc.) and enable multiple users in the same time, on the same combination.

Automated and Integrated. Sounds ideal but carefully consider the impact of all your forecasting and planning needs before building your solution. Perhaps now is a good time to evaluate them all and pave the way for an Integrated Business Plan?

How often do you forecast?

Forecast Cycle is the term used to define how frequently a forecast is generated, reviewed, approved and published. Forecast Cycle activities should respect the start and end points primarily and then fit everything else in-between. The starting time is arrival of source data (perhaps a full month's actual sales) and the end-time should be when the forecast needs to be submitted for onward activities in Supply Chain and Finance.

A Forecast Cycle should have a calendar set for the coming year with the date and time of each key activity (collection, meetings, approval, publish etc.) such as the example below:

Calendar of activities for a demand planning process

Modern Cloud Planning systems with their virtually zero down-time and multi combination access can transform Excel based solutions gaining valuable forecast accuracy and cycle time reduction. A mature forecast cycle will include the ability to re-forecast quickly and easily. Can you do that or does a significant change in circumstance have to wait for the next cycle?

Sharing some quality Forecast time?

There are many different Business Forecasts that run to different schedules with different horizons from the Annual Budget Plan with its 2-year horizon to the Daily Supply Plan for the next 3 months. See diagram below:

No alt text provided for this image

These various forecasts don't have to interact with each other, but it certainly helps increase accuracy and reduce maintenance if they do. If they cannot interact it is usually due to incompatible hierarchies, calendars and whether the horizon is rolling or fixed. S&OP or IBP will struggle to deliver change efficiently without synchronised structures through these forecasts.

Where do you focus your time?

A typical Demand Forecast might be created every Month, at the Monthly level for 24 Months but do you have a particular focus for planning activity? A common quote (or rather, complaint) you will see is "one thing you can be certain with a forecast is that it will be wrong", which is true enough but not particularly helpful. Your intent should be to make a forecast less wrong through science, art and of course, planner magic.

What then, is the most crucial part of the forecast to get as "less wrong" as possible? The near term, mid or long? The answer is down to your business and its needs: Lead-Times, Flexibility, Customer Delight, Profit, Inventory levels... any number of Company Strategy factors are important here, but I bet you Operations and your Planners are firefighting with the near-term forecast. They are probably matching orders to forecast and chasing outstanding orders.

matching the forecast to orders and chasing outstanding orders.  Fire-fighting in the near term.

If you want to improve - consider changing your Planning Focus. The better you get at longer term forecasting the less firefighting in the near term (probably).

Solutions that use a Statistical Forecast should be configured so that the engine is tuned to provide accuracy for the desired time focus. For example, it is possible to increase Statistical accuracy in the short term at the expense of mid and far. An engine tuned for equal accuracy across the entire horizon will likely not perform as well as one tuned for a particular time period.

Imagine... if you could increase accuracy in the near term your planners might do less firefighting. Alternatively, if planners are always going to adjust the near term regardless, maybe you should increase accuracy in the mid or far horizon to bring a better signal in over time. The point here is, evaluate & understand your time focus and configure your system and process accordingly to exploit any opportunities to provide more time to do value-add activities.

When Time is frozen.

A common method where forecasting suffers from severe bullwhip is to restrict the changes that are allowed to forecast in the near-term. This is called the Forecast Frozen Period. When Customer, Sales Forecasting or Supply Planning pressures are extreme this can be useful in stopping bad habits, but it is not always ideal as a permanent solution. A future article will look at how freeze periods work and the impact they can have on incoming supply.

Learn from your older mistakes

Forecast Accuracy KPI's usually select the most recent past forecast against actuals as a measure of predictive success. This is done because the most recent published forecast is naturally more accurate than the one from 2, 4 or 6 months ago. This is understandable when reporting the best news to The Board but it's really not the best way to evaluate and improve your forecasting.

If you can retain older forecasts, measure them. Find out how much value-add you are making and over what time periods. Comparing Final Forecast to Actuals is not enough. Compare Final Forecast to the raw Statistical Forecast and/or the Previous Approved Forecast without Overrides to find out how much value your changes make. It is entirely possible that not all of those adjustments improved things. Learn to let go the overrides that are not needed and focus on those that are and then find out what needs to be done to remove those too.

Fire up the Time Machine!

Here's a little trick that can be worth your time considering; why not become a Time Lord? If you run statistical forecast engines why not consider creating a dedicated tuning instance where you roll back the database time to a date in the past, then run your engine with setting changes. You will now be able to evaluate alternative forecasts using different models and parameters on a full forecast horizon with actuals in it. Voila! Instant and fulsome accuracy feedback. 

Time is money!

I have overlaid some linear forecast lines to reflect fiscal value (imagine Budget, Forecast and Last Year Actuals) across the birds-eye-view of time pictogram with some added mock-up dimension and hierarchy drop-downs to show how this kind of data could be used to create a scorecard. See the 'When to Forecast' diagram below.

No alt text provided for this image

Time is up

Just enough time left to revisit a few questions: Do you have too much data & not enough time? What time bucket do you predict? What Calendar are you using and is it the right one? Is your forecast tuned for short-, mid- or long-term accuracy? Do the Key Performance Indicators help drive better behaviour or just present the best possible %? Can you re-forecast at any time or are you stuck to the cycle? Do your Budget, Sales and Demand Forecasts share the same structures, or do they conflict?

If you don't know the answers to these questions, or you do and would like help in resolving them, please get in touch as I would love to help.

Now time really is up.

Thank you for your time!