How do you Forecast?
In a previous article I posed the question "Why Forecast?" and hopefully provided some answers as to the advantages of making and measuring predictions. The next logical step must then be to ask the question: How do you Forecast?
I am not going to evaluate the various Technological Solutions, or how they need to be configured or how different Corporate Cultures and Department Structures can influence decisions but instead focus on the most basic options that are available to anyone who needs to place a number in spreadsheet - what methods can you use?
Fundamental Forecasting Methods
There are two basic approaches that can be used:
- Statistical Forecast
- Previous Approved Forecast
These two approaches can be quite entrenched in corporate culture. I have worked with clients who could never envisage using the alternative method. There is a third way, which is to combine the two approaches, but more of that later.
Opposite sides of the same coin.
It is entirely feasible to mix these two approaches (Statistical and Previous Approved Forecast) up a bit. A Customer Forecast that is received each cycle can be treated as if it were a brand-new statistical forecast where you need to validate each combination for the full horizon each cycle or it could be treated like a Previous Forecast where you only review the changes.
The same can be said for Sales, Marketing, Finance and any Third-Party Forecasts you might use. Whether you treat them as Statistical or Previous Approved depends on the nature of your business and the systems and processes that are in place. Anyway, enough semantics - let's look at the two methods in a little more detail:
Statistical Forecast
This prediction creation method automatically generates a forecast using anything from a simple moving average formula in Excel to a comprehensive system solution or, 'PHD in a Box'. Statistical Forecasts require data, and the traditional approach is to collect history such as Orders, Shipments or Invoices and then using that data, generate a forecast of expected Orders, Shipments or Invoices.
The advantage of Statistical Forecasting is that you don't need to manually type a number in a spreadsheet (or rather thousands or even millions of cells in a spreadsheet). The system will perform the historical analysis and then create a forecast for every combination and for your full horizon which can save a huge amount of time and effort.
System Solutions that create statistical forecasts will typically offer a list of models (Regression, Intermittent, Holt Winters, Croston etc.) that can be selected automatically or manually and this is known as the 'Best Fit' approach. The Best Fit can be a one-size fits all or it can be more granular and applied per segment or even per combination though of course the specific capabilities vary across System offerings.
The Best Fit approach has been superseded in recent years by System Solutions that offer 'Bayesian Forecasting'. Bayesian forecasting blends statistical methods within individual combinations. So, an organisation, product & customer could have a mix of Croston and Holt Winters rather than just one or the other. This is meant to provide a superior end forecast especially where data is changeable due to casual factors.
The disadvantages of Statistical Forecasting is that 'The System' can become very complex to configure and maintain. 'The System' might create the forecast but it is the Demand Planners who needs to check, re-run, validate and override the forecast and this procedure will be repeated with each cycle. Lack of trust in 'The System' can cause huge problems and even result in external databases and spreadsheets to manage overrides and alternate forecasts.
Previous Forecast
This forecast creation method starts each cycle with the Forecast approved at the end of the last cycle. In its most basic format this forecast is created manually (so the first time takes a long time!) but thereafter the Forecast only requires adjustment.
Previous Forecast methodology can use the same source data as Statistical Forecasts (Orders, Shipments or Invoices) but these data streams are used for evaluation not generation. Previous Forecast methodology does not need historical data though - it can use imported data and is frequently used in businesses where forecast data already exists such as from Sales & Marketing, Finance (Budget) or perhaps collected from the Customer.
Prepare to think creatively with your data streams. Historical Data can be lagged into the Future to become a Forecast Stream. If you then add a percentage adjustment calculation you will have a forecast using last year's trends with a factor change function.
The advantages of the Previous Forecast approach are that there is no need for a time-consuming 'statistical engine step' nor does the forecast need to be checked everywhere. This approach is more about refinement leading to an evolving and improving forecast. A forecast created this way reduces system volatility and the risks associated with automatic generation.
There are a number of disadvantages of the Previous Forecast approach though. Changes in historical trend will not be automatically captured, new combinations will need to be manually created and old ones removed. Forecast values will need to be added for every combination in the increased horizon (a new cycle adds a new time period - that new time will be entirely blank for everything). Additionally, any source forecast data will contain bias that can be difficult for Demand Planners to remove. This is perhaps most especially true where the Demand Forecast is owned by Sales but managed by Operations.
Note - If a Sales Forecast is imported afresh each cycle it can result in a process very similar to a Statistical Forecast in that the entire forecast need to be maintained afresh. The Best Practice here is to feed the Final Approved Forecast back to Sales so that it becomes their starting point too.
A Third Way
Best Practice is to view these two approaches not as mutually exclusive options but as useful methods to be applied in tandem.
Segmentation analysis of your data can determine which combinations can be statistically forecasted easily, moderately, occasionally or not at all. Anything that falls into the latter two segments should not be statistically forecasted but instead should use a Previous Forecast. Likewise, any Previously Forecasted combinations that can be defined as Smooth or Repeatable data types should be automated so as to reduce Demand Planning maintenance.
There is more segmentation and refinement that can be done: Instead of Statistical and Previous use Statistical and Previous and Customer Forecast (where Customer is better than Statistical or Previous) and Sales (where Sales is better than Statistical or Previous or Customer) and Budget (where Budget is better than Statistical or Previous or Customer or Sales).
The lesson to be learned here is that you should consider embracing both approaches because when successfully employed this will reduce overrides, improve accuracy, reduce cycle time, increase planning maturity and create happier planners.
__________________________________________________________________________
Articles in this series are:
If you have any questions, thoughts, additions, queries or contentions about Demand Planning, Oracle Solutions or even grammar reflections please get in touch.