PredictiveAnalyticsMethodsPredictive analytics allows marketing experts to forecast consumer behavior and customer wishes so that they know how to better exploit sales and profit potential. Apteco, a pioneer in this field, is for more than 30 years committed to building industry-leading software to speedily convert customer data into actionable information and makes use of the following common methods:

1. Profiling is used to describe people based on their characteristics or behavior. For example, it can be used to characterize the typical behavior of a top customer by analyzing the target group compared to a population. To do this, you define several variables that are to be considered in the model. The profile shows you the penetration of these variables for your analysis set. Using the histogram, you can see which characteristics are over- or under-represented.

2. Association analysis (shopping basket analysis) means looking for patterns where one event is related to another. A frequently used example is the shopping cart as it includes the mix of products/brands purchased in a defined period. The shopping cart analysis determines the purchase probability for each of the products contained in the shopping cart with the aim of uncovering patterns and rules in purchasing behavior. A typical field of application is the identification of contexts when making a purchase, to react to this with targeted advertising measures, e.g., through cross-selling.

3. Logistic regression can be used to determine the probability of success of an event depending on certain variables. For example, conversion forecasts can be created by asking questions like “What is the probability of a customer buying a certain product?”. To create the statistical model, training data must be prepared, and prediction variables must be specified. The result is a series of coefficients. The more significant these are, the more meaningful the corresponding property is for the occurrence of the event.

4. A decision tree presents decision rules that are used to classify and graphically display data. In the decision tree analysis, the statistical classification takes place through the successive splitting of an analysis set, so that more homogeneous groups regarding the classification variables are found in the subsets that arise from this. Based on this, a statistical model is then developed that helps classify new data. The more intense the red in the tree structure, the better this node is suitable for describing the target group to be analyzed.

5. Best-Next-Offer assigns each customer the products that they are most likely to buy. It is precisely these products that are then displayed to the customer, for example, the next time they visit the online shop. It has now been proven that significantly more sales are achieved in this way than with a randomly selected product advertisement. The customer's purchase history is required as a prerequisite. Two parameters are added to the calculation, which describe the general popularity of the product combination (popularity) and the statistical significance of the product combination (propensity). This ratio can be weighted differently depending on the requirements.

6. The cluster analysis is a group formation process that searches for patterns in data with the aim of identifying homogeneous subsets from a heterogeneous ensemble. The differences between the individual groups (= clusters) should be as clear as possible and the differences within a group should be as small as possible. This is achieved by calculating proximity measures (e.g., Euclidean distance) and then combining objects into groups. Depending on the characteristics chosen, marketers can tackle very different tasks with the cluster analysis. A product for a specific target group can be developed or personality types can be found. For this it is necessary to know the existence, the size, and a detailed characterization of these target groups.

7. Linear regression tries to predict or estimate generally the numerical value of an observed dependent variable, e.g., annual sales per customer, by establishing a functional relationship between several variables. Future forecasts or trends can then be derived from the knowledge gained. However, the determination of the regression function does not yet indicate whether it is a significant determined relationship as it must still be verified by a so-called F-test.

By Daniela La Marca