Propensity Modelling for Marketing

Abstract: Propensity modelling is a sophisticated technique that leverages advanced data analytics to predict customer behaviour. By examining historical data and identifying patterns, businesses can use propensity models to understand customer tendencies and improve marketing efficiency. This article explores the benefits of propensity modelling, such as targeting the right audience, optimising resources, and enhancing customer engagement, ultimately leading to increased revenue and business growth.

An Introduction to Propensity Modelling

In today's competitive business environment, understanding and predicting customer behaviour have become crucial to successful marketing strategies. Propensity modelling, a powerful technique that leverages data-driven analytics, enables marketers to predict the likelihood of specific customer actions, such as making a purchase, subscribing to a service, or churning (Coussement & De Bock, 2013). By incorporating advanced statistical and machine learning methods, propensity modelling allows marketers to effectively tailor their campaigns and promotions, maximising return on investment (ROI) while minimising customer attrition. This article will delve into various propensity modelling techniques, highlighting their benefits, challenges, and potential applications in the marketing landscape.

Definition and Importance

Propensity modelling is a statistical and data-driven approach used to predict the probability of a particular outcome or behaviour for individual customers (Wu & Chen, 2015). In marketing, propensity models are developed to forecast customer behaviours, such as the likelihood to purchase, upgrade, or churn, based on various demographic, behavioural, and historical data (Kok et al., 2021). Predicting customer behaviour offers several benefits, including improved targeting, increased marketing efficiency, and better resource allocation (Siddiqi, 2017). As a result, propensity modelling has become an indispensable tool for marketers, helping them make more informed decisions, optimise marketing campaigns, and enhance customer relationship management (Verhoef & Leeflang, 2009).

Since the advent of computerisation in the business and industrial sectors during the 1970s and 1980s, companies have leveraged data to gain insights into their operations and clientele. In the past, structured data was the norm, and tools such as SQL and Excel were employed to query and derive valuable information. While this may appear rudimentary by today's standards, Marr (2022) asserts it remains the bedrock for many vital functions, including forecasting, profit and revenue analysis, productivity and workflow tracking, and customer analytics. Organisations can efficiently manage inventory, monitor orders, record client details, and discern sales and revenue sources by utilising spreadsheets, databases, and SQL. All of this supports the understanding of customer behaviour.

Applications in Marketing

Propensity modelling offers valuable insights for marketers in various areas, including customer segmentation, churn prediction, cross-selling, lifetime value prediction, lead scoring, and personalisation. By analysing demographics, past behaviours, and preferences, these models enable targeted marketing campaigns, improve customer retention, and facilitate relevant cross-selling and up-selling opportunities. Additionally, propensity models help estimate customer lifetime value, prioritise high-potential leads, and guide personalised marketing strategies. 

Customer Segmentation: By predicting the probability of specific customer actions, propensity models enable marketers to create targeted segments for personalised campaigns (Van den Poel & Buckinx, 2005). These models consider various factors, such as demographics, past purchase behaviours, and customer preferences, to identify patterns that can be used to group customers with similar characteristics (Leenheer et al., 2007). With this information, marketers can tailor their messaging, offers, and promotions to address each segment's unique needs and preferences, resulting in more effective marketing campaigns and higher customer satisfaction (Buckinx & Van den Poel, 2005).

Churn Prediction and Retention: Understanding the likelihood of customer attrition helps marketers develop strategies to reduce churn and retain high-value customers (Kok et al., 2021). Propensity models can identify early warning signs of churn, such as declining engagement, reduced purchase frequency, or negative customer feedback (Berson et al., 2000). By proactively addressing these issues, marketers can improve customer retention, reduce acquisition costs, and maintain a strong and loyal customer base (Reichheld & Teal, 2001).

Cross-selling and Up-selling: Propensity models can identify customers more likely to be receptive to additional products or services, facilitating cross-selling and up-selling opportunities (Singh et al., 2017). These models can uncover hidden relationships between products or services by analysing purchase history and customer preferences and recommending complementary offerings (Kamakura et al., 1991). This increases revenue and enhances customer satisfaction by providing relevant and personalised suggestions (Ansari et al., 2000).

Lifetime Value Prediction: By modelling customers' propensity to make future purchases, marketers can estimate customer lifetime value (CLV) and allocate resources accordingly (Kumar et al., 2008). CLV is a crucial metric that reflects the net present value of future cash flows generated by customers over their entire relationship with a company (Berger & Nasr, 1998). By identifying high-CLV customers, marketers can focus on nurturing these valuable relationships, providing personalised incentives, and ensuring high satisfaction levels, increasing loyalty and long-term profitability (Reinartz & Kumar, 2000).

Acquisition and Lead Scoring: Propensity models can prioritise leads based on their conversion likelihood, allowing marketers to focus their efforts on high-potential prospects (Verhoef & Leeflang, 2009). By incorporating demographic, behavioural, and engagement data, these models assign scores to each lead, indicating the probability of conversion (Pfeifer & Carraway, 2000). This enables marketers to allocate resources more effectively, targeting their sales and marketing efforts toward the most promising leads and improving overall conversion rates (Jap, 2001).

Personalisation and Recommendations: Propensity models can also guide personalised marketing strategies and recommend relevant products or services to individual customers, enhancing their overall experience (Tam & Tummala, 2001). By analysing customer data, these models can identify unique preferences, needs, and interests, allowing marketers to provide tailored content, offers, and recommendations (Xiao & Benbasat, 2007). This level of personalisation strengthens the customer-brand relationship and drives higher engagement, repeat purchases, and customer loyalty (Goldfarb & Tucker, 2011).

By leveraging propensity modelling techniques in these applications, marketers can enhance the effectiveness of their campaigns, improve customer satisfaction, and ultimately drive revenue growth. These models foster stronger customer-brand relationships, higher engagement, and increased loyalty, resulting in more effective marketing efforts and long-term profitability.

Traditional Propensity Modelling Techniques

Logistic regression, decision trees, support vector machines (SVM), and neural networks are traditional techniques for predicting customer behaviour. Logistic regression is a popular, easy-to-implement method for estimating the probability of binary outcomes. Decision trees offer an intuitive, visual approach to classification and regression tasks, handling non-linear relationships between variables. SVMs provide robustness, high-dimensional data management, and generalisation capabilities, particularly for large and complex datasets. Inspired by human brain function, neural networks excel at pattern recognition, classification, and regression tasks, adapting to changing data patterns and handling complex interactions. These traditional techniques enhance marketing applications across various scenarios.

Logistic Regression: Logistic regression is a widely used statistical method for modelling the probability of a binary outcome based on one or more predictor variables (Hosmer et al., 2013). In propensity modelling, logistic regression helps predict the likelihood of customer behaviour by estimating the probability of a specific event, such as making a purchase or churning, given input features (Coussement & De Bock, 2013). Logistic regression is easy to interpret and implement, making it a popular choice for marketing applications (Hair et al., 2010).

Decision Trees: Decision trees represent decisions and their potential outcomes for classification and regression tasks (Quinlan, 1986). Decision trees can be employed in propensity modelling to predict customer behaviour by recursively splitting data based on predictor variables (Breiman et al., 1984). Decision trees are intuitive, easy to visualise, and can handle non-linear relationships between variables, making them suitable for various marketing applications (Venkatesan & Kumar, 2002).

Support Vector Machines: Support vector machines (SVM) are a class of supervised learning algorithms that can be used for classification or regression tasks (Cortes & Vapnik, 1995). In propensity modelling, SVMs can predict customer behaviour by finding the optimal hyperplane that separates data points from different classes (Burges, 1998). SVMs are known for their robustness, ability to handle high-dimensional data, and generalisation capabilities, making them useful for marketing applications, especially when dealing with large and complex datasets (Coussement & Van den Poel, 2008).

Neural Networks: Neural networks are a family of algorithms inspired by the human brain's structure and function, used for pattern recognition, classification, and regression tasks (Haykin, 1998). In propensity modelling, neural networks can predict customer behaviour by learning non-linear relationships between input features and target variables (Zhang et al., 2000). Neural networks are highly flexible, can handle complex interactions, and can adapt to changing data patterns, making them a powerful tool for various marketing applications (Specht, 1990).

For years, Marr (2022) asserts, these analytical techniques have adequately supported businesses and remain potent instruments, particularly when handling structured data. Nevertheless, traditional methods are insufficient with the growing volume of unstructured data available to companies (including video, audio, and voice recordings). Broadly speaking, extracting insights from unstructured data tends to be more challenging, costly, and time-consuming. Consequently, novel analytical approaches have emerged to address these complexities.

Advanced Propensity Modelling Techniques

Advanced techniques like ensemble methods, deep learning, and reinforcement learning offer improved predictive performance in marketing applications. Ensemble methods combine multiple learning algorithms, enhancing individual model performance and reducing overfitting risks. Deep learning, a subset of machine learning, uses artificial neural networks to model complex data patterns, effectively capturing non-linear relationships and hierarchies. Reinforcement learning involves agents learning from their environment and optimising marketing strategies by iteratively adjusting actions based on outcomes. These advanced techniques contribute to more accurate propensity modelling, enabling marketers to maximise customer engagement, conversion, and retention, ultimately improving marketing performance and long-term customer value.

Ensemble Methods: Ensemble methods combine multiple learning algorithms with improving predictive performance (Dietterich, 2000). According to Opitz and Maclin (1999), ensemble techniques such as bagging, boosting, and stacking can enhance the performance of individual models, leading to more accurate predictions in propensity modelling. By leveraging the strengths of various models and reducing the risk of overfitting, ensemble methods have become famous for predicting customer behaviour in marketing applications (Polikar, 2006).

Deep Learning: Deep learning is a subset of machine learning that uses artificial neural networks with multiple hidden layers to model complex patterns in data (Goodfellow et al., 2016). Hinton et al. (2012) suggest that deep learning techniques can effectively capture non-linear relationships and hierarchies in data, making them well-suited for propensity modelling tasks. In marketing, deep learning has shown promise in various applications, such as customer segmentation, recommendation systems, and sentiment analysis (Nguyen et al., 2018).

Reinforcement Learning: Reinforcement learning is a type of machine learning where an agent learns to make decisions by interacting with an environment and receiving feedback in the form of rewards or penalties (Sutton & Barto, 2018). Li (2017) has argued that reinforcement learning can optimise marketing strategies, such as bidding in real-time advertising auctions, by iteratively adjusting actions based on observed outcomes. In propensity modelling, reinforcement learning can help marketers identify the most effective actions to maximise customer engagement, conversion, and retention, improving marketing performance and long-term customer value (Zhao et al., 2019).

Using Python For Propensity Modelling

Python is a versatile programming language widely used for propensity modelling due to its simplicity, readability, and extensive library support. With Python, marketers can quickly implement various machine learning algorithms and handle data preparation, feature engineering, and model evaluation tasks. Python's popular libraries, such as Scikit-learn, TensorFlow, and Keras, provide powerful tools for developing, training, and deploying propensity models, ultimately enhancing marketing strategies and decision-making. The dataset we use for the example below will consist of 1000 samples with 3 features: age, income, and a website engagement score. The target variable represents whether a customer purchased (1) or not (0).

# Create a target variable based on some simple rules
data['purchase'] = ((data['income'] > 70_000) & (data['engagement_score'] > 50)) | (data['age'] < 30)
data['purchase'] = data['purchase'].astype(int)
X = data.drop('purchase', axis=1)
y = data['purchase']

# Split the dataset into training and testing sets
X_train, X_test, y_train, y_test = train_test_split(X, y, test_size=0.3, random_state=42)

# Logistic Regression
logreg = LogisticRegression()
logreg.fit(X_train, y_train)
y_pred = logreg.predict(X_test)
accuracy = accuracy_score(y_test, y_pred)
print('Logistic Regression Accuracy:', accuracy)

# Decision Trees
dtree = DecisionTreeClassifier()
dtree.fit(X_train, y_train)
y_pred = dtree.predict(X_test)
accuracy = accuracy_score(y_test, y_pred)
print('Decision Tree Accuracy:', accuracy)

# Support Vector Machines
svm = SVC(kernel='linear', C=1)
svm.fit(X_train, y_train)
y_pred = svm.predict(X_test)
accuracy = accuracy_score(y_test, y_pred)
print('SVM Accuracy:', accuracy)

# Neural Networks
scaler = StandardScaler()
X_train_scaled = scaler.fit_transform(X_train)
X_test_scaled = scaler.transform(X_test)

model = Sequential([
    Dense(32, activation='relu', input_shape=(X_train_scaled.shape[1],)),
    Dense(16, activation='relu'),
    Dense(1, activation='sigmoid')
])

model.compile(optimizer='adam', loss='binary_crossentropy', metrics=['accuracy'])
model.fit(X_train_scaled, y_train, epochs=50, batch_size=32, verbose=0)
_, accuracy = model.evaluate(X_test_scaled, y_test, verbose=0)
print('Neural Network Accuracy:', accuracy)

In the context of the above examples, the accuracy score is a metric that quantifies the performance of the classification models. It represents the proportion of correct predictions made by the model out of the total number of predictions. In simple terms, the accuracy score tells us how often the model makes correct predictions. The score ranges from 0 to 1, with 1 indicating that the model has made correct predictions for all instances and 0 meaning that the model didn't make any correct predictions. To provide more context, let's say we have the following results:

• Logistic Regression Accuracy: 0.90
• Decision Tree Accuracy: 0.85
• SVM Accuracy: 0.91
• Neural Network Accuracy: 0.93

These scores suggest that the Neural Network model performs the best in predicting customer purchase behaviour on the given dataset, followed by the Support Vector Machines, Logistic Regression, and Decision Tree models. However, it's important to note that accuracy may not always be the best metric to evaluate a classification model, especially if the dataset is imbalanced (i.e., one class is significantly more frequent than the other). In such cases, other metrics like precision, recall, and F1-score can provide more insights into the model's performance.

Data Preparation and Feature Engineering

Data preparation and feature engineering are vital for propensity modelling, ensuring the quality, reliability, and relevance of input data. These processes enhance model performance by reducing complexity, mitigating overfitting, and improving interpretability. By carefully selecting features and addressing imbalances, propensity models can accurately predict customer behaviour, leading to more effective marketing strategies, increased customer engagement, and long-term value.

Data Collection and Cleaning: Effective propensity modelling begins with data collection and cleaning, which are critical steps to ensure the quality and reliability of the input data (Chen et al., 2012). According to Hand (1998), data cleaning involves identifying and correcting errors, inconsistencies, and missing values, which may arise for various reasons, such as data entry errors or system glitches. Thorough data cleaning helps reduce noise and improve the performance of propensity models, leading to more accurate predictions of customer behaviour (Wang, 1998).

Feature Selection: Feature selection identifies the most relevant variables from a more extensive set of input features for use in propensity models (Guyon & Elisseeff, 2003). As Guyon and Elisseeff (2003) suggest, appropriate feature selection techniques can help reduce model complexity, improve interpretability, and mitigate overfitting, ultimately enhancing the predictive performance of the models. Standard feature selection techniques include filter, wrapper, and embedded methods, each with advantages and limitations (Saeys et al., 2007).

Handling Imbalanced Data: Imbalanced data is a common issue in propensity modelling, where one class of customer behaviour, such as churn, is underrepresented compared to other classes (He & Garcia, 2009). He and Garcia (2009) state that handling imbalanced data is crucial. Traditional learning algorithms tend to be biased towards the majority class, leading to poor predictive performance for the minority class. Techniques for addressing imbalanced data include resampling methods, such as oversampling the minority class or undersampling the majority class, and cost-sensitive learning, which assigns different misclassification costs to the classes (Chawla et al., 2004).

Model Evaluation and Validation

Model evaluation and validation are essential when utilising propensity modelling, as they assess the models' performance, accuracy, and generalisability. These processes involve comparing predictions against actual outcomes and testing the models on unseen data to ensure their robustness. Through thorough evaluation and validation, marketers can identify the best-performing models, fine-tune them, and make well-informed decisions. Ultimately, this leads to more effective marketing strategies, improved customer targeting, and higher return on investment.

Performance Metrics: Evaluating the performance of propensity models is essential for determining their effectiveness in predicting customer behaviour (Witten et al., 2011). According to Kelleher et al. (2015), selecting appropriate performance metrics is crucial for assessing the quality of model predictions. Standard metrics used in propensity modelling include accuracy, precision, recall, F1-score, and area under the receiver operating characteristic (ROC) curve. Each metric has its own strengths and limitations, and the choice of metric should be guided by the specific goals and requirements of the marketing application (Provost & Fawcett, 2013).

Cross-validation: Cross-validation is a widely used technique for assessing the generalisation performance of propensity models (Kohavi, 1995). According to Hastie et al. (2009), cross-validation involves dividing the dataset into multiple subsets, training the model on a subset, and validating its performance on the remaining data. This process is repeated for each subset, and the average performance across all subsets estimates the model's ability to generalise to new data. Cross-validation helps mitigate overfitting and ensures the model performs well on unseen data (Varma & Simon, 2006).

Model Interpretability and Explainability: Model interpretability and explainability are increasingly essential considerations in propensity modelling, as marketing professionals need to understand and trust the models' predictions to make informed decisions (Guidotti et al., 2018). According to Ribeiro et al. (2016), explaining model predictions can help practitioners gain insights into customer behaviour, improve decision-making, and maintain compliance with regulatory requirements. Techniques such as LIME (Local Interpretable Model-agnostic Explanations) and SHAP (SHapley Additive exPlanations) can be used to enhance the interpretability and explainability of complex models, enabling marketers to derive actionable insights from their propensity models (Lundberg & Lee, 2017).

Challenges and Limitations

Marketers should consider the challenges and limitations of propensity modelling to ensure realistic expectations and effective decision-making. Acknowledging these limitations helps choose appropriate techniques, refine models, and mitigate potential biases or inaccuracies, ultimately leading to more successful marketing strategies and better resource allocation.

Data Privacy and Ethics: Data privacy and ethics are significant concerns in propensity modelling. Marketers must balance the need for personalised marketing strategies with the obligation to respect customer privacy (Martin, 2015). According to Dwork and Roth (2014), marketers must adhere to data protection regulations, such as the General Data Protection Regulation (GDPR), and consider ethical implications when collecting, processing, and using customer data for predictive modelling purposes.

Overfitting and Bias: Overfitting occurs when a propensity model captures noise in the training data, resulting in poor generalisation to new data (Hawkins, 2004). Bias, on the other hand, refers to systematic errors in model predictions due to underlying assumptions or prejudices in the data (Zliobaite, 2015). To mitigate these issues, marketers must employ techniques like regularisation, cross-validation, and careful feature selection and remain vigilant about potential sources of bias in their data and models (Hawkins, 2004; Zliobaite, 2015).

Scalability and Deployment: Scalability and deployment are crucial when implementing propensity models in real-world marketing applications (Witten et al., 2016). As marketing data grows in volume and complexity, marketers must ensure that their models can handle large-scale datasets and be efficiently deployed in production environments (Chen et al., 2012). This may involve optimising model architecture, leveraging distributed computing resources, or adopting cloud-based solutions for model training and deployment (Witten et al., 2016).

Conclusion and Future Trends

In conclusion, propensity modelling techniques have emerged as powerful tools for predicting customer behaviour and informing data-driven marketing strategies. These techniques, ranging from traditional methods like logistic regression and decision trees to advanced approaches like ensemble methods and deep learning, enable marketers to target better and engage customers, optimise resources, and ultimately improve their return on investment.

However, we have discussed several challenges and limitations associated with propensity modelling, such as data privacy and ethics, overfitting and bias, and scalability and deployment. Addressing these concerns requires robust model evaluation and validation, adherence to data protection regulations, and continuous vigilance for potential sources of bias in the data and models.

Looking forward, several trends will likely shape the future of propensity modelling in marketing. First, the increasing availability of granular and real-time data, such as social media activity and clickstream data, will enable marketers to build more accurate and timely propensity models (Russom, 2011). This will allow for more dynamic targeting and personalisation of marketing campaigns, further enhancing customer engagement and satisfaction.

Second, the growing importance of model interpretability and explainability will drive the development of new techniques and tools for understanding the inner workings of complex models, such as deep learning (Lipton, 2018). As marketers and regulators demand greater transparency in algorithmic decision-making, research on interpretable and explainable machine learning will flourish.

Finally, integrating artificial intelligence (AI) and propensity modelling will open up new opportunities for automating and enhancing various aspects of marketing, such as campaign management, customer service, and content creation (Gupta & George, 2016). For instance, reinforcement learning, a subfield of AI, holds promise for optimising marketing actions over time by adapting to changing customer behaviour and market conditions (Li et al., 2010).

In summary, propensity modelling techniques have become integral to modern marketing practices, potentially delivering significant value to businesses and customers. As marketing professionals continue to navigate the complex landscape of data-driven decision-making, staying informed about the latest developments in this field and being prepared to adapt to emerging trends and challenges are essential.

References:

Posted by Steve King

I am a marketing and analytics professional with over 15 years experience in strategic marketing development. I am passionate about working with organisations that want to improve their marketing effectiveness and get more from their data; who wish to use its potential to describe what has happened, prescript operational activity and predict business outcomes.

Related Theory

Big Data and Market Research Ethics

• Oct 15, 2021
• Big Data and Market Research: A Primer on Ethics

A Guide to Customer Journey Analytics

• Jan 3, 2024
• Understanding Customer Journey Analytics

The Influence of Electronic Word of Mouth

• Oct 5, 2022
• The Influence of Electronic Word of Mouth