Role of Forecasting in Demand Planning

Statistical forecasting plays a key role in the demand planning process. In a typical consensus-based demand planning process, the role of the forecast becomes much more prominent as the credibility of the demand planning process rests on the accuracy of the forecasts. Until recently, conventional time series forecasting methods have been predominantly used for forecasting in demand planning. A majority of demand forecasting tools in the market leverage these methods in their solutions. With advances in technology and computing power, the sophistication of these time series algorithms has increased thereby increasing forecast accuracy. However, with the advent of machine learning (ML) tools and increased interest in exploring them to improve forecasting approaches and methods, many vendors are increasingly incorporating ML-based forecasting methods in their tools. A key aspect to remember is that most of the frequently used ML algorithms have been around for decades now. Technology and computing power today allow us to leverage them easily in a variety of ways and applications.

Advantages of Using an ML Algorithm for Demand Forecasting

An ML algorithm typically has the following advantages over traditional time series forecasting methods:

• You can develop a solution that can learn by itself to improve recommendations, automatically capture demand signals, recognize patterns, and uncover complex relationships between variables — without having to write any additional programs
• A much wider range of influencing factors and relationships can be handled vs. typical forecasting algorithms
• These algorithms reduce the perceived complexity of demand patterns by selecting features/attributes that truly and significantly influence demand and filter out the “noise” from the data. This noise is typically random and unpredictable demand fluctuations.
• An optimally designed ML algorithm typically processes data much faster
• ML algorithms can handle large quantities and types of data

 What Are Some of the Typical Algorithms Used?

The list below captures some of the frequently used ML algorithms being leveraged currently, across industries:

• Decision Trees -- Decision trees are easily interpretable classification or regression models that essentially split data at decision nodes, where each node is a feature. (ex: color or size). In addition to forecasting, decision trees can also help understand which demand attributes impact demand the most.
• Regression -- Widely used forecasting technique to predict the future value of a dependent variable (like demand) based on the past relationship between the dependent variable and certain independent variables (like color, size, season, etc.). Like decision trees, these also help us determine which factors are more important in predicting future demand.
• k-Means Clustering -- K-means clustering falls into the category of unsupervised algorithms that cluster data points into groups based on certain attributes. This algorithm helps in demand forecasting by clustering the demand behavior of segments of customers and is frequently used in tandem with other machine learning algorithms for demand forecasting.
• Neural Networks (NN) -- NN models typically are made of multiple layers where each layer has multiple calculation logic (called neurons). These can be used for solving most of the demand forecasting scenarios but perform better than others only when certain criteria are met. These models are computationally intensive.

The algorithms mentioned in the list above are high-level categories. Readers who are interested in specific and more advanced algorithms can explore applications of the following algorithms in demand forecasting like Random forest, Support Vector Mechanism (SVM), and Gradient Boosting trees. We will be covering many of these specific algorithms in our future articles.

What Does This Mean For SAPinsiders?

As demand data sets become more complex, the usage of advanced algorithms in demand forecasting will become more common and necessary. Some key aspects that the SAPinsider Community should keep in perspective are:

• SAP is cognizant of the increasing importance of the role ML algorithms are playing in the demand forecasting space and continues to add these algorithms in the forecasting algorithms available in the SAP Integrated Business Planning tool (The most recent being a Gradient boosting decision tree) and will continue to extend the choice of forecasting algorithms in future releases.
• SAP Integrated Business Planning tool also has a “best fit” functionality that allows the planner to select the algorithm with the best accuracy based on the model fit error through training and testing.
• While SAP is actively working on transforming its tools like SAP Integrated Business Planning to infuse the latest ML tools in them, SAPinsiders also have the option to use other third-party applications that can tap into SAP data, leveraging a portfolio of ML algorithms and then use the one that works best for the underlying data. A majority of the off-the-shelf demand planning tools can be integrated with SAP data.
• A key step in selecting a third-party tool is to make sure that the embedded algorithms are compatible and generate a high level of accuracy for your specific demand data set.
• As demand data, patterns, and signals get more and more complex, ML algorithms provide us with tools to manage the intricacies of the data better. Thus, it is imperative for demand planners to develop an understanding of ML algorithms frequently used for demand planning.