WE ARE EXPERTS IN TECHNOLOGY

# StatusNeo

## How to Determine Forecastability of Demand?

In demand forecasting, there are a lot of cases where data scientists struggle to reach a good forecast error. On these occasions, they find it difficult to provide clear explanation to why some threshold of error is un-attainable no matter what technique is used. Forecast error of a product / event highly depends on its forecastability. If some event is unforecastable, it doesn’t even make sense to create a forecast model for it. We can rely on two parameters in order to determine whether the event is forecastable or not:

• ADI (Average Demand Interval) – Represents the measure of intermittency.
``ADI = Total Number of Period / Number of non-zero demand occasions``
• CV2 (Squared Coefficient of Variation) – Represents the measure of variation excluding zero demand events.
``CV2 = Standard Deviation of Population /  Average of Population ``

Based on these two parameters, we can classify the demand profile into one of these 4 different classes:

• `Smooth Demand`
• `Erratic Demand`
• `Intermitted Demand`
• `Lumpy Demand`

### Smooth Demand

ADI < 1.32 & CV2 < 0.49 is classified as Smooth Demand because the demand is very regular in magnitude and occurrence. This class is very easy to forecast and low forecast error is attainable.

### Erratic Demand

ADI < 1.32 & CV² >= 0.49 is classified as Erratic Demand because the demand is irregular in magnitude even-though it is in regular occurrence. Low forecast error will be hard to attain.

### Intermittent Demand

ADI >= 1.32 & CV² < 0.49 is classified as Intermittent Demand because the demand is very regular in magnitude but it is very irregular in occurrence. Low forecast error will be hard to attain.

### Lumpy Demand

ADI >= 1.32 & CV² >= 0.49 is classified as Lumpy demand because the demand is very irregular in magnitude and occurrence. It is improbable to create a forecast model that can have some predictability. Lumpy demand is unforecastable.

Determining the classification of historical demand makes it easy to plan out the research and development of a demand forecast model. This includes but no restricted to:

• Prioritization
• Algorithm selection
• Setting baseline for acceptable forecast error

A Data Scientist with proven track record of working closely with product owners, managers, front-end developers, data engineers, and data scientists to execute research/analysis with rapid prototyping.