14 Statistical forecasting

 

Statistical forecasting models can be roughly divided into two types: Time series forecasting, the most common one, and predictive models. This division can be a bit artificial at times—you could use both approaches in a single model—but let’s keep it for now because it will simplify our understanding of the topic.

14.1 Time series forecasting

Time series models look at historical demand patterns (trend and seasonality) and extrapolate them into the future. To do so, they usually decompose demand into three subcomponents: level, trend, and seasonality. Most time series forecasting models are also called univariate as they only use one variable (historical demand) to predict an outcome (future demand).

Definition: Univariate models

A model is said to be univariate if it only uses a single variable to predict another one. In our case of supply chain demand forecasting, it means that we only use historical demand to predict future demand

14.1.1 Demand components: Level, trend, and seasonality

  • Level: The level is the average value around which the demand varies over time. As you can observe in figure 14.1, the level often looks like a smoothed version of the demand. Example: “On average, we sell ten books per day.” Figure 14.1 illustrates the sales level of Toyota in Norway between 2007 and 2017.
Figure 14.1 Historical sales and level of Toyota cars in Norway42

14.1.2 Setting up time series models

14.2 Predictive analytics and demand drivers

14.2.1 Demand drivers

14.2.2 Challenges

14.3 Times series forecasting vs. predictive analytics

14.4 How to select a model

14.4.1 The 5-step framework

14.4.2 4-step model creation framework

Summary