Prophecy Newsletter

The Theory

Forecastability theory* scores a product's demand characteristic on two measures:

The regularity of demand (i.e. the number of periods between demands).
The variability of quantity per demand observation.

Therefore an item can be categorised into one of four quadrants:

Each quadrant has implications on how easily an item can be forecasted and also how it can be most effectively forecasted. Specifically:

Intermittent

Consistent sales level but large time interval between sales.
Forecastability depends on predictability of intervals between sales.
If it isn't possible to collaborate with customer to make intervals more predictable then consider safety stock strategy.
Alternatively, consider forecasting the item at a higher level of customer hierarchy (so that different customers' lumps are smoothed out through aggregation).

Lumpy

Highly variable sales level combined with large demand intervals makes lumpy demand items virtually impossible to forecast reliably.
Can you negotiate a more predictable demand pattern with the customer in terms of managing the interval between orders and pre-communicating their likely levels?
If not, is this item an economic sku to sell to this customer when the true costs are fully reflected?
Alternatively, consider forecasting this item at a higher level of customer hierarchy (so that different customers' lumps even themselves out).

Smooth

Consistent sales level AND regular demand interval.
These items should be relatively easy to forecast.
Consider using Prophecy's R Statistical Forecasting for these items.

Erratic

Regular but highly variable demand - possibly highly promotional or seasonal.
Explore Prophecy's R Statistical Forecasting for seasonal items.
Use Promotional Volume fact in Prophecy and consider implementing Prophecy's Trade Promotions module for itemised promotional forecasting.

*Forecasting References:

Article: Demand classification: why forecastability matters by Johan De Taeye
Academic paper by Adel A. Ghobbar and Chris H. Friend

The Practice

Forecasting literature recommends measuring demand level variability using the Coefficient of Variation (a measure of dispersion from the mean). Demand intervals are measured as an average. For example, if there are 24 periods in the history but only 6 have sales then the Average Demand Interval (ADI) is 24 / 6 = 4.0. So, if there are sales in every month then the ADI is 1.0.

There are recommended thresholds for using these measures to determine forecastability quadrants, as shown in this screenshot, although you can change them as required.

Prophecy screenshot from the new Forecastability Analyser

The red lines dileneate the quadrants corresponding to each category. A pop-up tooltip appears when you hover your mouse over white space in a quadrant, identifying the quadrant. Another tooltip appears if you hover the mouse over any plot point, showing that plot point's product/customer.

The Forecastability Analyser can present the same data as a 'bubble chart', where the size of each bubble corresponds to the average sale per period.

The button group labelled 'Push into Prophecy Report' on the screenshot lets you push the items from a chosen forecastability category into the current Prophecy report (replacing the existing selection) so you can start working with them as a group. You don't even need to close the 'Forecastability Analyser' window, as it's actually a floating window that runs in parallel with Prophecy's main window.

A potential workflow is (for example) to push all 'Smooth' items into your Prophecy report and then use Prophecy' s R Open automatic statistical forecasting to generate forecasts for these items.

You could push the 'Erratic' ones into the Prophecy report to analyse for seasonality. The MAPEs generated from running these items through Prophecy' s R Open automatic statistical forecasting can be analysed and, where the MAPE infers a reasonable forecast can be obtained, these items can be statistically forecasted. (A low MAPE implies that the statistics algorithms successfully identified the seasonality in the history and applied it to the resulting forecast.)

The remaining 'Erratic' items are likely to be heavily promotional and will need to be forecasted as such.

Finally, where items appear 'unforecastable' (i.e. 'lumpy' - high variability in sales and high interval between sales) then consider forecasting at a more aggregated level (of customer) to even out the spikes if smoothing the customer ordering pattern is not possible.

August 2019Sales Forecasting Software

'Forecastability Analyser' added to Prophecy!