A common conundrum for the demand planning and sales planning professional is to find the correct forecast level at which to forecast and produce plans. Forecasts are often used at very granular levels of detail, sometimes down to the level of SKU, location and even day, but does this mean that forecasts should be generated at this detailed level? Is a detailed bottom up forecast better than an aggregate top down forecast? Or is it the other way around with a better result from top down forecasts? Which should be used?
The answer is neither (or perhaps more positively and practically, both!). The fact is, forecasts are used in many ways and it is extremely unlikely that a one size fits all approach will offer satisfactory results in all cases (check out an earlier post on planning perspective for more detail on this). Lets take a look at why this is, and more importantly, what we can do about it!
Fundamental Rules of Forecasting
It is generally accepted that there are two fundamental rules of forecasting:
- Forecasts are more accurate for closer periods of time. Forecasts are generally more accurate the closer they are to the current period. It also follows that forecasts are less accurate for farther periods of time. In other words, the accuracy or quality of the forecast decreases as we move further down the time horizon into the future.
- Forecasts are more accurate for groups. Forecasts are generally more accurate for groups (for example, grouped by products, customers, time or a combination). This means that forecasts are generally more accurate at aggregate forecast levels than they are at detailed forecast levels. For example it is easy to imagine producing a reasonably accurate product category forecast for 18 months in the future, but achieving a high level of accuracy on the underlying product and customer combinations is far more difficult.
These fundamental rules are fairly basic and it is usually intuitively clear that these hold true. On their own however, they do not allow us to understand which forecast level should be used in any given situation. For this, we need to delve a little deeper.
Forecast Level: Implications of the Fundamental Rules
The fundamental rules lead to some interesting implications to consider when assessing the forecast level to use to produce forecasts and plans. Lets look at 3 implications below:
- Forecast Level Implication #1: Detailed Granular (bottom up) Forecasts are More Effective in the Short Term. Considering the two fundamental rules together, it becomes clear that using detailed granular forecasts in the long-term horizon is in effect the worst of both worlds. Not only is the forecast less accurate due to the granularity and lack of grouping, it is also less accurate because of the distance into the future. It follows from this, that where detailed granular forecasts are needed, they will be more effective closer to the current period.
- Forecast Level Implication #2: Aggregate (top down) Forecasts are More Effective in the Longer Term. Of course, following from the first implication above, the reverse also holds true. Where forecasts are needed further out into the horizon, they will be more effective when components are grouped together into an aggregate forecast. In a practical sense, this means at an aggregate of product, customer, geography, time or any combination of these.
- Forecast Level Implication #3: The ROI for Detailed Granular Forecasts decreases Significantly Further down the Time Horizon. Detailed forecasts can often be resource intensive to produce, especially when utilizing judgmental qualitative forecasting and planning techniques that need human interaction. Given that forecasts further out into the time horizon are inherently less accurate, trying to increase accuracy of granular forecasts in the long-term is an uphill battle. Care should be taken to not over-capitalize in effort to produce detailed forecasts in the long-term horizon. Not only is it common for forecasts to be unnecessary at a detailed forecast level in this horizon, by producing detailed forecasts and rolling them up, it impacts the accuracy of the rolled up forecast as well! Combine this with the cost of producing the detailed forecast and it suddenly doesn’t make a great deal of sense at all!
We are starting to get a picture of how forecast level impacts the effectiveness of a forecast and why both bottom up and top down forecasts can be suitable depending on the circumstance. For a practical solution however, we need to delve even further.
A Practical Solution to Selecting the Perfect Forecast Level
So what does this mean? Does this help to select the right forecast level? In a way, the answer to this question is yes. However, it is necessary to come to terms with the fact that more than one forecast is often needed. As there are multiple uses for the forecast, it follows that the best level at which to produce forecasts for each use is not always the same. But trust me, this is ok! It is possible (in many cases I’d go so far as to say its necessary), to effectively use multiple forecasts within a larger planning process (for further insight check out this previous post on ‘one number’ forecasting). The 4 step process below will walk through how to decide where this is appropriate.
- Step #1: Identify the different uses (or use cases) for which forecasts are used in various planning processes (as a feed into the S&OP process, a feed into the master scheduling process, to drive distribution replenishment and so on). For each use case, note the frequency of update, the time horizon and the forecast level or level of granularity required.
- Step #2: Group these together based on the frequency of update, the length into the horizon and level of granularity. The trick is to strike the right balance between the number of forecasts, and what is sacrificed by not producing a forecast at a particular level. As an example, if you have two separate use cases for a weekly regenerated forecast at the same forecast level but with two different horizons, it makes sense to produce only one forecast for both (using the longer horizon). Similarly, if there are two use cases that need a similar (but slightly different) forecast level and the frequency of update is the same, it will probably make sense to produce single forecast at one level and aggregate or disaggregate to the other. However if two use cases require different frequency of update at different levels of granularity then usually this is a trigger for the use of multiple forecasts grouped separately.
- Step #3: Design forecasting processes around the previously identified groupings. At this stage, keep each grouping separate. Design in isolation the perfect forecasting process for the individual grouping of use cases. Design a methodology to build each forecast, including a process in which it is updated each period, a process to capture and incorporate relevant insight and a way to measure and review accuracy. If a forecast will be used at more than one level, show how aggregation and disaggregation is handled. Ensure that the generation and review timings line up with the equivalent downstream processes that will use the forecast.
- Step #4: Integrate each forecast into a broader forecasting framework. Critically, this step enables the use of multiple forecasts while maintaining alignment between them. The framework identifies where and when forecasts are reviewed against each other, acceptable tolerance bands for alignment, and what happens when there is a mis-alignment greater than the accepted tolerance. In general, less frequent, higher level, longer-term forecasts should be used as guide rails to sub-ordinate more frequent, lower level, shorter term forecasts. If a mis-alignment is identified the sub-ordinate forecast is reviewed ensuring its based on valid assumptions. If it isn’t valid then the sub-ordinate forecast should be corrected. If the sub-ordinate forecast is valid, feedback is provided to the higher level forecasting process for incorporation into the next cycle (the higher level forecast does not need to be changed immediately!). It’s vitally important to realize that this integration process is not the same as a reconciliation process. The forecasts do not need to be in complete alignment all of the time.
By using the above 4 step process, we can identify situations where more than one forecast is required to enable the selection of the perfect forecast level (or levels). In addition, we have taken the first steps towards designing an integrated multi-level forecasting process tailored to meet specific forecast use cases. Based on this, is the perfect forecast level a top down approach? Is it a bottom up approach? It is clear in most cases it is both!
Have you experienced using multiple forecasts in your organization? If so, were they integrated into an overall forecasting framework? Or perhaps you have seen a case where the use a single forecast has been sufficient and effective. In any case, please share your thoughts and perspectives in the comments below!
The post The Perfect Forecast Level: Bottom Up or Top Down? appeared first on The Planning Blog.