Decoding Your Customers’ Actions With Propensity Modeling

Collecting Relevant Data

To gain a comprehensive understanding of user behavior and drive effective decision-making, it is essential to collect and analyze relevant data. This data serves as the backbone for statistical modeling and empowers businesses to design targeted intervention strategies. Let’s explore the different types of data that play a crucial role in this process.

• Demographic Data: Understanding customer demographics is vital for tailoring products and services to specific target audiences. This data includes information such as age, gender, location, marital status, education level, and other relevant factors that help businesses cater to diverse customer needs effectively.
• Behavioral Data: Analyzing past actions and interactions with the company or website provides insights into customer behavior. This data encompasses purchase history, browsing behavior, click-through rates, frequency of visits, and time spent on specific pages, allowing businesses to personalize experiences and enhance customer satisfaction.
• Transactional Data: Details about a customer’s past purchases, including the types of products or services bought, purchase frequency, average transaction value, and purchase recency, offer valuable insights into consumer preferences and shopping patterns.
• Customer Service Interactions: Data from customer support interactions, including call center logs, chat transcripts, and email communications, offers invaluable information about customer sentiment and satisfaction. Analyzing this data helps in identifying areas for improvement and optimizing customer service experiences.
• Response to Marketing Campaigns: Data related to a customer’s response to previous marketing campaigns, such as email open rates, click-through rates, and conversions, offers insights into the effectiveness of different marketing initiatives.
• Survey Data: Customer survey responses and feedback forms provide direct insights into preferences, opinions, and areas for improvement, guiding businesses in refining their offerings and enhancing customer experiences.

Collecting relevant data from multiple sources enables businesses to build a robust foundation for statistical modeling, gain profound insights into user behavior, and implement data-driven strategies that lead to better customer experiences and overall success. After you identify useful data you need to clean the data. This is discussed in more detail in the next section.

Preparing Data for Modeling

Data preparation is the step that takes 80% of the effort in building and deploying a propensity model. It is crucial to create an accurate data set, because bad data leads to bad models or worse, the wrong conclusions being found. This process involves collecting, cleaning, and transforming raw data into a format suitable for statistical analysis and modeling. Here’s an overview of the steps involved in preparing data for a propensity model:

Data Cleaning and Preprocessing:
Data collected from diverse sources may contain errors, missing values, or inconsistencies, which can significantly impact the model’s accuracy. Cleaning and preprocessing the data involve identifying and handling missing values, removing duplicate entries, and correcting any discrepancies. Additionally, outlier detection and treatment are essential to prevent skewed model outcomes due to extreme data points.

Feature Engineering:
Feature engineering is a crucial step in data preparation, as it involves selecting and creating relevant variables (features) that contribute significantly to the propensity model’s predictive power. This process may involve transforming variables, creating interaction terms, and selecting relevant subsets of data to enhance the model’s ability to capture patterns and relationships.

Creating and Testing a Model

Your extensive efforts in the previous step of preparing the data makes this step possible. In this step we need to select a model type, build it, and test it for accuracy and stability.  Propensity models encompass a variety of statistical approaches, each selected based on the data structure, distribution, and desired output format. We will focus on two types of models in this article, logistic regression and random forests. Below you will find the strengths and weaknesses of each model.

Logistic regression, a fundamental propensity model, has a long-standing history in experimentation. It is easy to interpret, because each feature is assigned a coefficient tied to its effect on the outcome. Moreover, it delivers swift results due to relatively simple math. However, logistic regression performs best with data that are not overly complex.

In contrast, random forests are built on combining the decisions of multiple users and finding the most likely choice. They exhibit greater adaptability and robustness, accommodating various data types without necessitating encoding. Moreover, random forests implicitly perform feature selection, simplifying the modeling process. Yet, these models can be computationally intensive, especially for large datasets, and their “black box” nature offers limited control over their decision-making process, making it challenging to discern why specific trees performed better or were assigned higher weights. Nonetheless, random forests are recognized as catch-all solutions for many classification problems due to their versatility and efficiency across diverse datasets.

After you select a model type you will build the model using a tool like R, Python, or SAS.  This building process is beyond the scope of this article, but a trained Statistician or Data Scientist should be employed to do this properly. 

Deploying the Model

After the model is created the next step is to implement the model into the business. These steps should be outlined in the data strategy to ensure that the business users can benefit from the model scores. A common deployment strategy is to score a customer file using a decile ranking. Deciling scores from a model allow you to break the customer file up into groups representing 10% of the total customers.  This approach helps to more efficiently target the top and bottom performing groups in your list and to track them more efficiently in the next step.

Here is an example of how this works. You have a list of customers that you want to know who is the most likely to buy your product. You built a model that answers this and you score the customers using the model. You then order the customers by score and break them into groupings representing 10% of the total list size. Now You know you can only market to 20% of the total file. So, you take group 1 and group 2 which are the most likely 20% of customers to purchase and you focus your attention on them.

Quadrant scoring can also be an effective method of implementation. Check out our other article that discusses this in more detail.

Performance Tracking

Deploying reporting for propensity models is a pivotal step in leveraging the insights derived from these predictive models to drive data-informed decision-making. Through effective reporting, businesses can communicate the model’s outcomes and predictions in a clear and concise manner, making it accessible to key stakeholders and decision-makers. The reporting should encompass relevant performance metrics, such as accuracy and business impact, to gauge the model’s effectiveness and identify areas for improvement. 

Visualizations, charts, and dashboards can further enhance the understanding of the model’s predictions, enabling stakeholders to identify trends and patterns easily. Overall, deploying robust reporting for propensity models fosters a data-driven culture within the organization and facilitates strategic actions that optimize customer experiences, drive growth, and deliver tangible results.

propensity model case study