Creating customer win-back marketing campaigns with machine learning

In the world of retailing, the Win-Back campaign is a staple for almost every marketing team.

There are a multitude of studies that examine its value and the success of related customer retention strategies: Seminal research published in the Harvard Business Review is a good place to start if you want to examine the evidence, along with the book, Marketing Metrics - The Manager's Guide to Measuring Marketing Performance , by Paul Farris and Neil Bendle.

However, Win-Back campaigns are not as common in travel retailing businesses such as airlines, online travel agencies and loyalty brands. This is puzzling because the retention effect is amplified for e-commerce and digital businesses.

Why? Because referrals and recommendations from loyal customers through digital channels such as social media can be distributed and amplified much more quickly than traditional word-of-mouth.

This reduces the upfront cost of customer acquisition enabling new customers to become profitable and loyal earlier.

Retention, loyalty and data capability are now on the CEO’s agenda

Some airlines have recognised this opportunity. Newly appointed EasyJet CEO, Johan Lundgren talked about the importance of retention, loyalty and data on a recent investor call:

"Loyal regular customers create the most value and at the moment only 46% of our customers fly with us more than once a year, which creates a great opportunity. We’ve identified three significant areas of opportunity: holidays, business customers, and loyalty. All three will be unlocked by supercharging our data capability."

So, retention initiatives such as win-back campaigns speak for themselves in terms of their potential to generate a profitable return for any travel retailing business.

And larger airlines such as British Airways are clearly investing heavily in both manpower and enabling Big Data technologies.

But what is the best way to approach a win-back campaign in travel?

There are three typical requirements for a successful Win-Back campaign:

1. Collect high quality customer data on purchasing behaviours.
2. Use machine learning to separate ‘Win-Back’ customers from the broader base.
3. Deliver micro-segmentation to identify sub-segments for personalised messaging.

Each of these must be delivered to a high standard for a win-back campaign to be successful.

Three requirements for a successful Win-Back campaign

#1 Customer data

The success of every win-back campaign is dependent on the depth and breadth of customer data available to the marketing team.

To create a successful win-back campaign, customers that are good candidates for win-back campaigns must be clearly identified and separated from the rest of the customer base.

This is non-trivial because we need to measure this across multiple dimensions simultaneously, for example:

1. The frequency of purchases.
2. The recency of purchases.
3. The average spend (order value) of every customer.

Each of these dimensions will vary by industry.

For example, a customer may have lapsed from a grocery chain if they haven’t purchased for three months, but in travel, this may be as long as 9-12 months, especially if we are selling holidays rather than more frequent business travel.

Once a group of candidate win-back customers are identified, they need to be further segmented into smaller groups that follow a demographic or related theme.

We need this so that the win-back marketing communications (typically distributed via e-mail, but can also use social media and other online channels) can be personalised and crafted in a way that offers incentives and propositions that are relevant and attractive to specific segments of customers.

# 2 Machine Learning

Identifying the customers that are suitable for win-back campaigns across multiple dimensions simultaneously is very difficult to implement manually for even small customer bases.

For larger, real-world retail customer bases, it is practically impossible to do this manually, no matter how good your Excel skills are!

Here’s where machine learning comes in. With algorithms like k-means clustering and self-organising maps, a customer base can be segmented into groups with each group representing some common pattern across multiple dimensions.

For Win-Back campaigns, we are primarily interested in the purchasing behaviour of customers, so we cluster customers based on their Recency, Frequency, Monetary profile – called RFM ("Monetary" in this context is simply average spend / order value).

Let’s take a look at this in action. The screenshot below shows how a marketing analyst can quickly identify a group of customers for a win-back campaign.

The system we’re using here is t-Data, OpenJaw’s new Big Data platform for travel retailers.

At the top of the screen are some Key Performance Indicators (KPI’s) that summarise the total number of customers in the customer base, their total spend in Euros for the past year, plus their average recency score, frequency score and monetary (we call it "spend") score.

The bubble chart identifies five separate clusters of customers. To find these clusters of customers, t-Data’s clustering algorithm was run across the full base of 43,142 customers.

The most interesting cluster amongst these five clusters for a Win-Back campaign is the group labelled Risk of Churn/Big Spenders.

This group is important because although it represents a group of "churned" or "about to churn" customers, they were good customers in the past if we look at their spend profile (denoted by the size of the bubble).

Next step is to click on this cluster in the bubble chart and a filter will be applied to create your segment of win-back customers – in this case 11,053 customers per screenshot below.

#3 Micro-segmentation for personalised messaging

This segment can be further refined by choosing only those customers that have a frequency score greater than one i.e. are not one-off customers.

The demographic profile of this segment can also be examined, and, if appropriate, used to split the win-back segment into separate sub-segments e.g. families / senior couples/millennials so that that different, tailored marketing messages can be delivered to each sub-segment.

The screenshot below shows how this can be done inOpenJaw t-Data by clicking on the Family segment in the Passenger Mix chart in the dashboard below to create a micro-segment of 2,653 customers.

The last step is to take the segment and export it using Openjaw t-Data’s DaaS (Data as a Service) tool to the system used to execute the win-back campaign. This will typically be an email marketing campaign tool, but could also be delivered to online channels such as social media or search platforms.

Understanding analytics is the platform to understand big data

The clustering algorithm used by OpenJaw t-Data described above created the bubble chart that identifies five separate clusters of customers. This algorithm is really a sophisticated form of descriptive analytics.

Descriptive analytics is focused on past behaviours,so works on the principle of hindsight.

As a result, it is a very useful (and valuable!) tool for marketing campaigns like Win-Back campaigns, because it can quickly and clearly separate churned, but valuable customers from new or loyal customers that require a different marketing campaign.

Descriptive, Predictive and Prescriptive: the three different types of analytics

The next step after Descriptive Analytics is Predictive Analytics. Predictive Analytics, as the name indicates, can be used to target individual customers based on predictions of their future behaviour.

Fewer travel brands use this approach, mainly because it is more difficult, or large histories of good quality data are not available.

In the world of ecommerce, it is typically used to segment customers based on predictions of their future behaviour e.g. their propensity to purchase a high-margin product.

It is based on predicted future behaviours, so works on the principle of foresight. These propensity forecasts can be used in many different selling contexts.

For example, to refine customer segments to smaller, more precise groups that have a high propensity to purchase the product in question, or to embed in business rules for dynamic pricing.

The final type of analytics is Prescriptive Analytics: only a small number of sophisticated organisations have deployed these solutions.

In this scenario, very large quantities of high quality data are required, along with powerful computing resources and machine learning expertise.

The power of prescriptive analytics is that the ideal action for each customer can be automatically taken. In this way prescriptive analytics can either replace or augment a task that previously required a human user to manage.

Picture a scenario where A.I. assistants like Alexa or Siri pro-actively engage with a customer to manage a flight over-booking situation, using propensity predictions to determine the best alternative flight for that person and any relevant cross-sell or up-sell opportunities.

In this way, revenue can be maximised, communications personalised and overall customer satisfaction improved. Prescriptive analytics work on the basis of using predictions of future behaviour together with other data to automate a task that requires some judgement, working on the principle of insight.

Where to next with supercharging your marketing campaigns using data?

The next logical step is to identify which products different customers segments are likely to purchase. This is where predictive analytics and the power of propensity modelling come into play.

So, what does this word "propensity" actually mean?

Let's take a simple example to illustrate. A person may have a high propensity to purchase travel insurance if they are naturally risk averse, purchased travel insurance in the past and fit the demographic profile of people who usually purchase insurance.

In many cases, a person may not even have consciously registered that they have a propensity to purchase something, but the propensity model can predict it.

Propensity modelling uses sophisticated machine learning algorithms to predict what a customer is likely to do next, and deliver customised offers and personalised travel experiences that optimise conversion rates, media spend and revenue per customer.

Propensity modelling is not a new idea, but it is only with the availability of large amounts of high quality data, cloud infrastructure and machine learning that these models are starting to generate predictive signals with reliable accuracy.

Propensity in travel is revealed by a combination of a consumer’s demographic profile, their transactional history, loyalty profile and psychographic profile.

The data used varies depending on what propensity we want to predict. But, generally, transactional data serves as the foundation to everything else in a propensity model as this is hard evidential data that tells us what buying patterns and preferences a consumer has had over time and therefore serves to validate everything else.

This is the next stage in sophistication with Marketing Campaigns and Machine Learning: using accurate propensity predictions to support very sophisticated digital marketing campaigns that target a segment of one with highly customised offers and a super-personalised digital experience right across the customer journey.

For an airline operating as a travel retailer using Machine Learning to power your marketing campaigns is very valuable and has many applications.

Now is the time for airlines to adopt Machine Learning to substantially increase conversion rates, win-back old customers, optimise media spend and increase the revenue per visit to through direct channels.