In a recent video blog, I discuss forecast accuracy as a parameter for measuring the ability to forecast and plan demand. I further argue for the use of causal data as a key input to understanding historical demand and forecasting/planning future demand. Forecast accuracy is often claimed NOT to be the only parameter explaining whether your demand planning process boosts the competitiveness of your company. But it is a key indicator for several reasons:

• It is used directly in optimizing inventories. And poor forecast accuracy increases safety stock.
• It is linked to the total uncertainty in your supply chain. Increased uncertainty leads to more firefighting. It also means higher costs for transportation and capacity in order to meet customer expectations.
• It is a key motivator and trust builder in your process. Poor forecast accuracy typically reduces commitment to not only the forecasted figures and subsequent plans but also the process itself.

Data is the key

This is why you should continuously measure and improve forecast accuracy. Surveys indicate that demand-driven companies outperform those that use more traditional ways of forecasting and planning in terms of growth, customer satisfaction and profitability. In an earlier blog series, I introduced the road to become market driven as developing the forecasting/demand planning process along two axes: organizational focus and technology focus.

• Technology: With technology, it is a matter of moving from manually forecasting in a spreadsheet – or even using traditional time-series forecasting software – towards analytical software that can identify and incorporate demand drivers into your forecast.
• Organizational: Organize stakeholders and develop competencies that support a demand-driven forecasting and planning process.

For a more detailed explanation of the development, please read "The Market-Driven Journey."

Hence, data is the key enabler for improving forecast accuracy. It is only through the use of additional data that you can identify, explain and analyze differences between expected demand and actual demand, and subsequently use this information for future planning. Whereas traditional time-series data are great at explaining trends, seasonality and cycles, they are not able to take causal factors into consideration. This will require data explaining why there is a deviation – e.g., a price change. These causal and independent data sets that help explain changes in demand are called demand drivers.

Identifying demand drivers

Identifying demand drivers and their effects requires software that can perform such analysis and forecast models that include them. The forecasting solutions that do not offer such functionality cannot compete in terms of forecast accuracy.

Demand drivers can be found in any company. In B2C focused supply chains (retail, CPG), traditionally price, promotions and events are potential demand drivers. In B2B, these demand drivers may not be as influential. And in my experience in B2B, outside factors are often the most important.

Through the use of relevant demand drivers, you can increase forecast accuracy and reduce those forecasts that are more than 100% off by half. All in all, you will go from uncertainty to expectation – thereby reducing the costs associated with uncertainty. There is, however, a limit to the number of demand drivers you can add to the forecast in time-series forecasting, and thereby a limit to how much of the noise can be explained.

Enter machine learning and AI

Recently, machine learning techniques have been introduced into forecasting. Machine learning looks for patterns in data and draws conclusions. The big advantage is the volume of data that can go into forecasting. Basically, there is no limit. And it is commonly argued that the more data that is used, the better the forecast accuracy.

Machine learning techniques may provide superior results in some forecasting cases.

One such case is for new product forecasting or products where there are no or only a few demand data for a product. By combining the demand patterns experienced in previous product introductions, machine learning algorithms can predict demand more accurately than traditional methods by taking ALL the available causal data into consideration.

Another is for products where there is high volatility due to promotions, price changes, etc. Again it is the application of causal data that better explains how demand will look.

#MachineLearning models can not only improve #forecast accuracy for these products, but also help predict cannibalization effects on other products. Click To Tweet

Needless to say, machine learning requires lots of data to make accurate predictions about the future sales of a product. Lacking data reduces the accuracy of machine learning models. This may favor more traditional time series forecasting models. Further, machine learning models are hard to interpret.

The key is thus to use machine learning forecast models alongside traditional models. You can then choose the best fit to maximize forecast accuracy, thereby improving competitiveness.