Forecasts are always wrong

This statement is true, but forecasting helps to be less wrong. Forecasts of future sales will never be ‘correct’, but they can be made more accurate by using ranges in forecasts and measuring ‘forecast error’.

The market demand (and supply) for an item is a ‘continuous variable’ consisting of multiple variables that interact dependently, independently and interdependently, creating emergent results (that is, you do not expect them). These results may be cumulative – what appear to be minor influences on a forecast can become major when linked with other small events. This ensures that a forecast of sales (and other supply chain factors) will be wrong.

If a forecast will be wrong, then developing a periodic forecast for an item with a single point number attached to each period is potentially dangerous. This is especially so if people in your organisation e.g. sales directors and chief accountants, view the forecasts as ‘accurate’.

Forecast a range

For this reason, instead of a single point forecast, use your influence with Sales to provide a range that identifies optimistic and pessimistic forecasts. Ask for the probability of attaining each forecast and an explanation that reflects the likelihood of the forecast occurring.

To improve accuracy, forecasts are done by product group or product line for the country or sales region and each data time series can be modified by applying four analysis to the data set – trend, cycle, seasonality and randomness.

Within the Operations Planning system, the forecasts by product group or line are multiplied by their probability; this is more likely to provide a believable indication of future demands. The forecast can then be de-constructed to the individual SKUs, based on percentages of prior sales within the product group or line and weighted by product features or options (such as for apparel: finish, size, colour and fit) and future promotions. The process now has the name of ‘proportional profile planning’ (PPP).

This approach can assist with planning in Logistics and Procurement. For Logistics, it allows the planning of warehouse space and transport requirements, which take account of a possible higher demand in the future for external or sub-contracted capacity to be negotiated.

For Procurement, the required items with long lead times, due to shipping times or manufacturing backlogs, can be ordered at the optimistic forecast level. This may result in periodic excess inventory of particular items, which is preferable to being out of stock. The items that are easy to obtain and on short lead times are ordered at the pessimistic forecast level. To provide cover for the potential of attaining the optimistic sales forecast, provisional orders are raised (with cancelation clauses), so the amount of an item can be increased at short notice.

Forecasts contain a ‘forecast error’

If forecasting in your IT system does not have the facility to incorporate range forecasts, use a spreadsheet. However it is done, range forecasts can improve forecast accuracy – especially if the ‘forecast error’ that occurs between forecast and actual sales is also measured.

The term ‘forecast error’ is NOT a measure of performance, but a statistical term to identify the difference between planned and actual situations. To compensate for ‘errors’ in the forecast, a level of ‘safety stock’ (or inventory) is calculated for each SKU.

Of the various measures of ‘forecast error’, the most common reference is the Mean Absolute Percentage Error (MAPE). This is the difference between the actual and forecast sales (change any minus signs to a plus) and divide by actual sales. Then sum the errors and divide by the number of observations. 

How accurate should forecasts be? Inventory should be identified by categories within the extended ABC inventory classification approach that uses the ‘Coefficient of Variation’ (CoV). This is another part of the analytical approach to inventory management (policy, planning and control) and sales forecasts that need to be implemented in organisations.

Using the extended ABC classification for inventory, SKUs in the high selling STEADY category could expect a MAPE of 10–15 percent. The tail end of variable, low selling SKUs can have a MAPE result of 20–30 percent and up to 60 percent. This is due to more unknowns associated with the sales of a SKU and the likely presence of bias in the forecast.

There is naturally a temptation to make the numbers look better, especially if ‘forecast error’ measures are used for the wrong reasons.. Even though the results may impress senior management, two approaches are not advised:

• only measure forecasts at the product group level and
• weighting the errors – called the Weighted Mean Absolute Percent Error (WMAPE)

An additional measure is the Tracking Signal, which assists inventory control at the SKU level. It is calculated by dividing the cumulative variation between forecast and actual sales for the number of periods under review by the standard deviation of the variances.

For each inventory category, a control limit is established. SKUs in the high selling STEADY category will have an acceptable tracking signal up to 4. For the more variable inventory categories, a review is triggered when the tracking signal is more than the control limit of about 7.0.

An objective for Operations Planning should be to work with Sales to improve the MAPE and Tracking Signal through investment in training, process improvement and systems. As part of the training, have the teams exposed to developments in planning applications that could have future use in your organisation, such as Cognitive Computing and Artificial Intelligence (better called Machine Learning).