Propensity Modelling: Definition, types and use cases

Propensity Modelling is a powerful tool, led by statistics and machine learning, which can empower brands to confidently predict customer behaviour. Using historical data, propensity models can be trained to forecast a customer’s likelihood to convert or re-purchase. 

By accurately segmenting your audience through the use of Propensity Modelling, you open up the ability to deliver highly customised marketing campaigns. They can be used to identify your highest-value customers so that you can nurture their loyalty and engagement through the personalisation of your service.

These models also aid in the acquisition of new customers, identifying those groups of prospects who are most likely to spend in your business. By identifying users most likely to engage, brands can prioritise their ad spend and boost return on investment:

Ready to get started with Propensity Modelling and other advanced market segmentation techniques? Speak to our team today or explore our service offering to discover more:

• Conversion rate optimisation services
• Email marketing automation services
• Media measurement consultancy
• Paid media services

What is Propensity Modelling?

Propensity Modelling is a tool used by marketers to predict the probability that a customer will take a particular action. Consumer data is used to designate scores to customers which tell us how likely they are to provide future value to the company.

Propensity Models are built using large amounts of customer data. This could include demographic information such as age and gender, as well as behavioural data like browsing habits or engagement with email and ads. More advanced models could also incorporate psychographic data, and interaction with social media channels into their models.

Machine learning tools can be trained, using historical data with known outcomes, to look for patterns in groups of customers which can be used as predictors for future behaviour. For example, perhaps customers who engage regularly with marketing emails are more likely to make a purchase.

Over time, as more data is made available, the algorithm becomes more refined and increases in accuracy. This tool can now be used to form detailed predictions about the buying habits of potential customers. Customers most likely to provide value to the business can be easily identified and targeted with tailored marketing campaigns and offers, maximising conversion rates.

Where does this term come from?

‘Propensity’ refers to the tendency towards a specific behaviour or characteristic. For example, someone with a high propensity for risk-taking may be more likely to partake in an extreme sport, or invest in a volatile stock.

The term ‘Propensity Modelling’ can be attributed to Paul R. Rosenbaum and Donald Rubin in 1983, who recognised that the best predictor for future behaviour is past behaviour. Their findings have been applied in marketing to develop the practice of Propensity Modelling to calculate the probability of a customer performing certain actions.

This knowledge can be used by marketing teams for a range of applications, including tailored product campaigns, sales forecasting and churn predictions.

Introducing propensity scores

Propensity Models provide us with propensity scores. These tell us the probability that a website visitor, lead, or existing customer will carry out a certain action. A score closer to one suggests that the customer is highly likely to take that action, whilst a score closer to zero indicates a far lower likelihood.

This knowledge provides marketers with an opportunity to interact with customers accordingly.

Propensity Modelling techniques

There are a number of different Propensity Modelling techniques which can be used to produce these propensity scores. Decision trees, neural networks and logistic regression all come with unique pros and cons which may influence their uses. All three techniques can be used to classify leads as being likely or unlikely to take a particular action.

Decision trees 

Decision trees work in much the same way as flow charts, splitting data into ever-narrowing sub-groups at each stage of the process. 

The algorithm begins with the entire dataset and splits it into categories. For example, perhaps it divides customers into those who have purchased from the company before, and those who have not. Each of these groups is then divided into smaller sub-groups. Eventually, you are left with a number of distinct categories, each with an assigned propensity score that indicates the likelihood that members of that group will perform a particular action.

Advantages of decision trees

• Highly visual – a decision tree can be clearly laid out in a diagram which is easily interpreted and understood.

• Quick to set up – compared to other methods, decision trees are fast to train and can be started with a relatively small data set with little preparation.

• Non-parametric – they can look at complicated customer data without needing to fit it into a specific pattern. This is useful because not all customer data follows the same trend. For example, the chance that a customer makes a purchase may directly increase with age, but might not directly correlate with the number of visits they make to the website.

Disadvantages of decision trees

• Prone to overfitting – the tree can become highly specific to its training data and may perform poorly on new data. 

• Instability – a small change in the data can lead to changes at multiple levels of the tree’s structure.

• May be biased – for example, a split based on five age categories, will divide customers into smaller groups than a simple yes or no split. This means age would play the greatest role in dividing the group, even if it’s not the strongest indicator of a particular future behaviour.

Neural networks

Neural networks are so-called because they work in the same way as a human brain. Over time, the brain gathers information about the world around it. It begins to form new connections that help it to be able to make predictions about the future. You may know to expect warmer temperatures in the month of August, because your brain has been gathering information and noticing this recurring pattern for a number of years.

Computers can be trained to do the same thing. Data is input to the network and passed through a series of “neurons” which perform mathematical functions at each stage. The data output forms a prediction. As the network is fed more data over time, it notices where its predictions are not matching reality, and adjusts its calculations to refine the predictions further.

In the case of the Propensity Model, our output comes in the form of a propensity score of zero to one. Over time, neural networks can be trained to produce extremely precise predictions for each individual customer based on their previous behaviour.

Advantages of neural networks

• Handle highly complex relationships – neural networks are able to recognise relationships which may be missed by simpler models.

• High predictive power – the accuracy of predictions is high, especially with large data sets.

• Handle very large data sets – large quantities of data can be analysed quickly and precisely.

Disadvantages of neural networks

• Not useful for all data sets – neural networks are parametric, meaning they can only be used to analyse data which follows a specific pattern. Metrics such as customer survey scores may not follow this pattern and could not be used as part of a neural network.

• Hard to interpret – it is not always clear how the scores are produced and it can be hard to identify how predictions have been made.

• Require a large amount of data – they are far less effective with small data sets.

• Computationally expensive – they require huge levels of computational power and time, which may be costly to the business.

Logistic regression

Logistic regression predicts the outcome of a binary decision- one which has only two possible outcomes. When applied to a Propensity Model, this means we are predicting whether an action will or will not be taken. For example, logistic regression could be used to predict whether a customer will or will not make a purchase. 

It has the ability to take multiple data points connected to a customer and use them to formulate a “yes” or “no” outcome based on that data. Once again, as more data is input, the prediction is refined to become increasingly accurate.

Advantages of logistic regression

• Easy to interpret – this method is easy to understand and interpret. It’s useful for making predictions with only two possible outcomes.

• Useful for small data sets – less data is required than would be necessary for a neural network.

• Fast to train – The efficient algorithm means that computational requirements are low and training times are quick, even with large data sets.

Disadvantages of logistic regression

• Not useful for all data sets – logistic regression is another parametric tool which can only be used for data with a linear relationship.

• Limited to simple problems – as logistic regression can only have two outcomes based on linear data, it will struggle to compute more complex problems.

• Sensitive to outliers – data points which fall outside of the normal range can have a large impact on the results produced using this technique.

How do Propensity Models differ from other predictive marketing technology techniques?

Like Propensity Models, uplift and lookalike models can be used to predict future customer behaviour. Each model has a distinct purpose and can be applied to different scenarios to maximise the impact of marketing campaigns.

Propensity Models versus uplift models

Both propensity and uplift models make predictions about the future behaviour of a customer, based on data sets pertaining to how similar customers have acted in the past. They make use of machine learning tools to build predictive models which can be used to identify how an individual is likely to engage with a brand.

However, an uplift model also aims to predict which customers are most likely to change their behaviour based on an interaction. For example, an uplift model could be used to predict which individuals are most likely to change their behaviour in response to a marketing email. These people can then be prioritised in future campaigns.

Propensity Models versus lookalike models

Just as with uplift models, lookalike models take large amounts of historic customer data and use it to identify patterns in those who are most likely to provide value to the company. However, whereas a Propensity Model aims to make predictions about existing leads, lookalike models work to find new customers who match the characteristics of the highest-value existing customers.

A lookalike model will identify the common characteristics of the top subset of a business’ customers. This information can then be used to target ads and marketing budgets at individuals who share those traits, with the intention of drawing new customers to the business.

Types of Propensity Models

Propensity Model use cases include estimating the probability that a customer will either make a purchase, stop using a service or interact with marketing efforts, or estimating the value of a customer to the business over their lifetime. They can therefore be grouped into four categories.

Purchase-based Propensity Models

Purchase-based Propensity Models focus solely on whether an individual is likely to spend money in the business. The model could be based on how often they have bought before, how recently their last purchase was made, or the typical value of their purchase.

For example, if a fashion brand is launching a new seasonal range, it may use a purchase-based Propensity Model to pick out those customers who regularly buy from a new collection. These buyers can be sent targeted promotions to encourage them to return to the brand.

This model is particularly useful at building brand loyalty and improving customer retention.

Churn rate models

A churn Propensity Model figures out which customers are most likely to stop using a business. It may look at buying habits, but key identifiers include decreasing engagement with a product or service, as well as measures of low satisfaction.

A business using a subscription model, for instance, might evaluate the chances that a customer will renew their subscription using this model. If they are identified as a cancellation risk, the company may send them personalised offers or recommendations to encourage them to stay.

This has the benefit of increasing customer retention, but can also highlight areas of the business which may need improvement to increase customer satisfaction.

Engagement models

Engagement models help businesses to understand how customers are interacting with their brand. These models focus heavily on behavioural data and could look at metrics such as email open rates, click-through rates on ads, and likes, comments and shares on social media.

This knowledge allows marketers to increase brand loyalty and focus resources in the areas where they are getting the most engagement, maximising advertising revenue. It can also be used to target the most engaged audience with specific marketing campaigns.

Customer lifetime value models

Customer lifetime value models look at customer buying behaviour, including how often they buy and how much they spend, and use this to predict the total monetary value they are likely to provide to the company over their lifetime. 

There are a number of different versions of this model to choose from. The Pareto/NBD model predicts when a customer is likely to make a purchase. The BG/NBD model can be employed to do the same, but goes further by suggesting how long they are likely to continue to buy from the business. The Gamma-Gamma model, on the other hand, predicts how much they are likely to spend. 

These models can be used individually or in combination to predict the future buying habits of customers over an extended period of time, estimating their customer lifetime value (CLV). This score can be used to contribute to RFM (Recency, Frequency, Monetary Value) segmentation.

RFM segmentation provides insight into which customer profiles provide the most value to the business. Together, Propensity Modelling and segmentation can be used to target these small groups to help reduce unnecessary marketing spend and increase profitability by ensuring the loyalty of the best customers.

Use cases for Propensity Models

The scope for Propensity Models to inform how companies interact with customers is huge. It can help to build loyalty and engagement, increase sales, and attract new customers to the business. Its applications range from informing product recommendations to reducing churn on subscriptions and much more.

Website interactions and customer acquisition

Website interactions provide key insights into a customer’s engagement with a brand. Every action they take builds a picture of their relationship with the brand and can contribute towards generating a lead score for that individual. This allows a business to identify its most engaged prospects and focuses marketing efforts towards those with the highest chance of converting.

Interactions of interest could include which pages a customer has visited, how long they spent on each page and how far they scrolled down the page. All of these feed into our knowledge of which products or services most excite the customer. Identifying returning customers helps to highlight those prospects who are already loyal to the brand.

It also allows the customer to receive personalised recommendations of products and services. This helps the business to show prospective customers more of their offer, and also helps the client to find products which most closely match their needs and solve their problems. This can contribute to improving customer satisfaction, increasing the likelihood that they provide repeat business.

Over time, a highly detailed customer profile can be built which, when used correctly, can streamline marketing efforts and improve budget efficiency by putting adverts and communications in front of the right people at the right time. 

Loyalty schemes

Loyalty programmes such as Nectar and Tesco Clubcard are used to collect customer data. They allow a unique insight into customer behaviour by providing the business with a complete purchase history for each individual member. Propensity Modelling can use this data to highlight the customers most likely to buy on an individual product level. 

By using customer segmentation, a business may present customers with highly personalised offers. This not only increases conversion rates, but allows the customer to feel seen and understood by the company, increasing their loyalty to the brand. Reward schemes can also be used to motivate customers to return to the business.

Subscription businesses

For companies which rely on customers taking out subscriptions to their services, Propensity Models are vital tools for identifying customers likely to churn. Once these customers are identified, they can be targeted with “stay” bonuses and other incentives.

For example, if you skip a few HelloFresh deliveries, you might receive a discount offer for your next box. If you subsequently cancel, you might be enticed back in with incentives like “refer a friend, receive a free box”, which have the added benefit of incorporating a referral into the offer.

Customer journey mapping

Customer journey mapping provides a visual representation of a customer’s touchpoints with a brand. It tracks a customer’s position within the marketing funnel. AI-powered propensity scores allow a company to produce these maps, pinpoint an individual’s place within it and predict their future trajectory with high precision.

This insight has further refined the ability of a business to hit customers with the right offer at the time. For example, a customer in the awareness phase could be shown advertising demonstrating the benefits of a particular product and illustrating how it can provide solutions to their problems. A customer in the consideration phase could receive email marketing to encourage them to purchase.

Product recommendations

Information such as browsing history and past purchases can be used to build a clearer picture of a customer’s needs and interests. Tools like Google Cloud Retail Recommendations use AI to process this data and provide customers with bespoke product recommendations. The AI might assess, for example, that a customer who has previously purchased running gear, might be interested in other running products and show them adverts for running shoes. Publishers can also leverage alternative yet similar offerings which recommend content to users.

Cross- and up-selling based on propensity scores

Propensity Modelling can be used to identify places where there are opportunities to up-sell or cross-sell. This knowledge can be used to tailor marketing to buyers of a particular product, with the intention of increasing the value of the sale, and benefitting customer satisfaction by providing the customer with products aligned to their needs.

Past data might indicate that users of a particular smartphone, for instance, are likely to be persuaded to be up-sold to a newer or more expensive model. Customers buying the basic model can be shown marketing designed to entice them to upgrade. Similarly, if data shows that buyers of that phone commonly purchased insurance to go with it, they could be offered the two products as a bundle.

Propensity Modelling tools

There are a vast and growing number of Propensity Modelling tools now available. Each offers a slightly different range of services. 

Google Analytics 4 (GA4) offers predictive metrics such as purchase probability, churn probability and predicted revenue, whilst Google Cloud provides tools to present customers with relevant advertising. 

Customer Data Platforms (CDPs) like Bloomreach and Klaviyo also collect data to make informed product recommendations. They can be used for audience segmentation, automated email marketing and search recommendations among other things.

Python packages can provide data used for CLV modelling. Their “Lifetimes” package can predict how long a customer is likely to use the business, whereas “PyMC” is used for Bayesian modelling. This is a form of modelling which evolves over time as more data is collected and the algorithm is refined. It could be used to predict the performance of a new product or recommend existing products to customers.

How to build a customer Propensity Model

The process of building a customer Propensity Model requires the use of historical customer data. Statistical analysis can use this to identify relationships between features of customers, and the subsequent behaviours they were likely to exhibit. Once we know this mathematical relationship, it can be used to make predictions about new customers.

1. Choose a Propensity Modelling technique 

Selecting a suitable Propensity Modelling technique will require an evaluation of the unique features offered by each option. Logistic regression may appeal because of its simplicity, but a decision tree or neural network may allow for more robust and sophisticated analysis.

One consideration may be the nature of the data set. Neural networks demand huge amounts of data to operate effectively, and may not be suitable for a company with limited data sets. Logistic regression can function using small amounts of data, but is limited to analysing linear data. If your data comes in a range of distribution patterns, a decision tree might be more suitable.

The desired outcomes should also be taken into account. Logistic regression outputs a simple “yes” or “no” response to a query, whereas neural networks can deliver a precise propensity score. Neural networks are more likely to uncover subtle and complex patterns in data which may not be picked up in other models.

2. Define the features of your model

Once a model is selected, the next step is to decide which data are to be input into the model. These are known as the “features” of the model. The features you choose will be dependent on the output you are trying to achieve.

Channel grouping identifies how customers reach your site. This may be a suitable feature to put into your Propensity Model if you are interested in finding out which channels are most likely to provide sales. Is someone arriving by organic search more likely to convert than someone who comes through social media, for example?

Other features could include the time a user spends on your website, the pages they visited, the actions they took whilst they were there, their purchase history, demographic data like age and gender, or their net promoter score (a measure of customer loyalty and satisfaction). 

3. Construct your model

Your model is built using historical data. This data is useful because we already know the outcome. Logistic regression analysis can be performed to look for connections between the selected features, which act as independent variables, and the outcomes, which act as dependent variables.

For example, you might look for how the time spent on your website impacts customer lifetime value. Historical data, showing how long previous customers have spent on the website and their subsequent customer lifetime values, can be used to perform a logistic regression analysis. This will reveal the mathematical relationship between the two variables.

The most basic outcome would be a perfect correlation between the two variables. Imagine that it showed that as time spent on the website increased, the customer lifetime value also increased. This would appear as a straight line on a graph, which could then be used to predict the lifetime value of a future customer spending any length of time on the website. 

In reality, the relationship is likely to be far more complex, especially when more features are incorporated, but the same principle can be applied- relationships identified in old data can be applied to new data to make highly informed predictions.

4. Calculate propensity scores

The logistic regression analysis has been performed on real-world historical data to achieve accurate results. Once the model is created, you might test it using experiments to ensure its accuracy and efficacy, and to see how the model changes as the parameters are changed. For instance, does increasing the money spent on advertising affect the outcome? Adjustments may be made to the model as you include new features.

Once you are confident that the model accurately reflects the behaviour of your customer base, it can be used to calculate propensity scores for future customers. In our example, the length of time a new customer is spending on the site could be input, and the model can now provide a propensity score to indicate how likely they are to become a high-value customer.

How Propensity Modelling and experimentation work together

Propensity Modelling techniques can be used to improve conversion rates and enhance customer experience. By personalising their experience to show them the most relevant products and information at the point when it is needed most, consumers can feel more connected to the brand. Using Propensity Modelling to refine audience segmentation can ensure recommendations are further tailored to the individual.

This can be further improved through experimentation. Split testing can be used to trial different techniques on different customers. The outcomes can be input into Propensity Models to evaluate the impact on a propensity score. As an example, does a particular website design impact the propensity of customers to make a purchase?

 Propensity Modelling by Impression

Impression is an independent, operator-owned agency, offering a range of digital experience and media solutions services. Whether you are just starting out with Propensity Modelling – or you need support with refining your approach – we are well-positioned to help.

Interested in learning what we can do for you? Speak to our team for a free consultation.