1
Lecture 1
Introduction
Why analyse customer behaviour?
Analyzing customer behavior serves three main purposes:
• Prediction: By studying customer behavior, businesses can forecast future trends, preferences,
and purchasing patterns.
• Understanding customer behavior: Examining how customers behave provides valuable
insights into their motivations, decision-making processes, and preferences.
• Informing company policy (prescription):
- Delivering/obtaining value: Analyzing customer behavior helps businesses identify ways to
enhance value delivery to customers through personalized offerings, improved service, and
targeted marketing efforts.
- Improving customer experience: Understanding customer behavior enables businesses to
optimize touchpoints, streamline processes, and address pain points to enhance the overall
customer experience.
- Nudging towards sustainability: By analyzing customer behavior, businesses can promote
sustainable practices by encouraging eco-friendly choices, reducing waste, and supporting
ethical consumption.
Why models?
Models play a crucial role in analyzing customer behavior by abstracting away from details and
focusing on key elements. They consider more than one factor and facilitate a spectrum of analyses,
from simple to complex, to describe, predict, and prescribe actions based on insights. Importantly,
models are not limited to observed cases, allowing for extrapolation and scenario analysis to
provide a deeper understanding of customer behavior.
Why not always models?
Not everyone relies on models due to skepticism towards their effectiveness. Additionally, concerns
about the time-consuming nature of developing and utilizing models, as well as issues related to
data availability, can deter their widespread adoption.
Different types of DVs
Euro Sales = Continuous DV = Regression (OLS)
Units Sold = Continuous DV = Regression
Purchase Intentions = Continuous DV = Regression
Customer Retention (stay or go?) = Choices (2 options)
Email Response (yes or not) = Choices (2 options)
Type of milk purchased = Choices (>2 options, unordered)
Type of subscription (basic, regular, deluxe) = Choices (>2 options, ordered)
How many products does a costumer buy? = Limited number (discrete and ordered)
, 2
Model Building Basics
What is a model?
A model is a stylized representation of reality.
Its goal is to understand this reality.
What are the basic elements of every model? Inputs ➡ Process ➡ Outputs
It should contain the most important elements but it is never complete
The four steps of model building
1. Specification (S)
2. Estimation (E)
3. Validation (V)
4. Use (U)
Regression example
, 3
A good model is:
- Simple
- Robust
- Easy to control
- Adaptive
- Complete on important issues
- Easy to communicate with
- Evolutionary
Simple
The simplicity of a model is characterized by a small number of variables that include only
important phenomena. This can be achieved by combining variables, analyzing sub-problems and
constraining parameter values. The goal is parsimony, which means explaining much by means of
little.
Robust
Robust models are hard to producing illogical results and are based on a deep understanding of the
market and customers. They generate meaningful outcomes values.
The requirements are:
- Correct marginal effects
- Meaningful interaction effects
- Taking into accounts endogeneity effects.
Adaptive
Adaptive models acknowledge the dynamic nature of the environment. They are designed to evolve
with changing circumstances as yesterday, today, and tomorrow may differ. Parameters may need
adjustment as the world changes over time and new significant factors emerge. It is essential to
critically review existing models to ensure adaptability and relevance.
Complete
Complete models include all essential factors such as competition and dynamics. However,
achieving completeness may conflict with simplicity. The completeness of a model is relative to the
specific problem, organization, or user at hand. For instance, decentralized firms may face different
complexities compared to centralized firms.
Evolutionary
In the process of model development, it is advisable to start with a simple model and allow it to
evolve over time. This evolution involves meeting with stakeholders to build a basic model, using
and gaining experience with it, and gradually expanding its complexity.
The approach involves two steps: developing a conceptual model to understand the nature of the
phenomenon and then creating a statistical model to quantify and capture the phenomenon's
characteristics.
( ➡ see slides 56-75)
, 4
Lecture 2
Model Estimation
Why not OLS?
So, you might be asking yourself why we can't just use standard OLS regression (Yi = β0 + β1Xi +
β2Xi +…+εi), which is common in courses like market models. Remember, in ordinary least
squares, or OLS, we usually deal with continuous dependent marketing
faculty of economics variables, not binary ones, along with
and business
predictors (x) that can be either continuous or dichotomous (0,1). In this scenario, our dependent
variable is influenced by various betas, x variables, and an error term, creating a simplified
| 16
model
of our dataset, a conceptual depiction of reality, the points.
Predicted
sales
Observed
sales
The line is the model: a stylized representation of reality (the dots). Essentially, we align these
values to a trend line. For instance, if we consider the relationship between advertising and sales,
you can observe the sales outcomes at a specific advertising intensity.
You can see at a certain level of advertising the observed sales by looking at the y axis. But you can
also see the predicted level of sales by our model. And both, of course, don't have to be equal to
each other. Perhaps you recall, indeed, the graph from last week showing our data, which we
suggested was normally distributed along the regression line. However, that's not possible here
(binary y variable) because, in reality, at most points along the regression line, there are no data
points. Instead, all data points are located solely at one and zero.
Now we use as an example a firm that wants to know if their current mail was getting responses
from customers.
But the data they have is tricky because it only shows responses as either a yes (1) or no (0). This
made it hard to use the usual way of drawing a line on a graph to predict responses (data points only
on 1 or 0).
To deal with this, we can treat the responses (dependent variable) as probabilities. Instead of just
yes or no, they could be seen as how likely it is for someone to respond. This way, the line on the
graph could predict responses between 0 and 1, which matches the data better.
