DEMYSTIFYING MARKETING BUZZ WORDS – LEVERAGE CUSTOMER DATA WITH ANALYTICS TO PREDICT FUTURE BEHAVIOUR

Foreword: BUZZWORDS & ANALYTICS

Like most buzzwords, many of the terms you will hear banded about at marketing and digital conferences relate to processes that have been undertaken by marketers for decades in one form or another.  They are given new names based on changes in related techniques of simply to make them sound more exciting.

In addition to outlining how to leverage your customer data, we aim here to demystify the meaning of a range of current buzzwords, bringing them into a more relatable context.

So who are my customers?

Businesses are collecting more and more information about their customers every day. This brings about challenges in terms of drawing the data together, standardising it, analysing it and establishing ways to leverage it in a measurable way. The vast volumes of data that it can be collected has led to the term “Big Data.” This simply alludes to the fact that there is so much information that it would be impossible to analyse the information with traditional techniques.

The “Big Data” nirvana can be described as the true single customer view, where all known information on all customers are easily available, allowing businesses to make clear decisions based on all data points available.

 If you consider your company, you may be collecting data on your customers in a range of disparate systems (silos). It is not uncommon for businesses to have limited linkage between these systems, which ultimately makes it challenging for marketers to understand all the touch-points they have with their customers.

Data collection points can be grouped into 3 key types.  

Transactional

This related to financial interactions and is often described in terms of RFM (recency, frequency and monetary value of spend) or customer lifetime value.

This includes data collected through:

• ERP (Enterprise Resource Planning) system / finance package - examples include SAP, Oracle, Netsuite, MYOB, XERO, Quickbooks, Microsoft Dynamics
• eCommerce platform (online shop) - You may run this in-house / use 3rd party software like Shopify, Neto, Magento and Zencart

Behavioural

This relates to the buying and research behaviour of your customers and prospects.

This includes data collected through:

• CRM (Customer Relationship Management platform) - Examples include Salesforce, ZOHO, SugarCRM, OracleSalesCloud, MicrosoftDynamics CRM. Data collected include interactions between the customer and your sales & customer service teams, technical support teams
• eMarketing platform (email marketing software) - Examples include Marketo, Salesforce Exact Target, Mailchimp, Drip, Sendinblue
• Website visits
• Recorded behaviour relating to life-stage triggers - Examples include moving home, wedding, having children

Geodemographic / business descriptors 

This relates to data that describes your customer in terms of the type of person or business that they are, in ways that can used to compare then to other people / businesses

This type of data may be self-reported by the customer / procured form third party data aggregators. 

Consumer segmentation variables of this type include:

• Age
• Income
• Household makeup / relationship status
• Presence of children
• Home ownership status
• Household variables including home type (house / duplex / apartment), number of bedrooms, number of car parking spots
• Ethnicity / language spoken
• Religious preference
• Aggregated geo-demographic segmentation variables that cluster consumers in terms of their makeup in terms of combinations of the information above - examples include Mosaic / LandScape / CAMEO. These variables are built using a range of sources, including Census data, expenditure data and media consumption survey results.

Business segmentation variables of this type include:

• Annual sales revenue
• Industry vertical
• Number of employees
• Blue collar / white collar
• Home based / office based
• Private / public company
• Importer / exporter

So what is a persona?

Personas are meaningful customer and/or prospect segments. These might relate to existing or probable behaviour and transactions and/or descriptors that can be used to tailor messaging to the right audience, ideally at the right time.  

The build of a persona may draw upon multiple transactional, behavioural and geodemographic data points, which can be described as “segmentation variables.”

To determine where to start, we can use a range of statistical techniques, including data mining, customer profiling and predictive modelling.

Data mining

Data mining describes the process of statistically analysing a dataset, to understand it better.  

Examples of data profiling might include:

• Establishing the % of female Vs. male customers that bought a product in 2 time periods
• Establishing how many customers that bought product A also bought product B
• Establishing how many marketing touch points are needed on average to close a sale in each city in which you operate
• Running a post-campaign analysis to determine how many people that received your letter later came into a store and bought a product, using their rewards card

These processes enable marketers to discover purchase trends, relationships and correlations between various data points.

While basic data mining is often great place to start your analytical journey, it is often limited by the virtue of the data being analysed in a vacuum.

The best way to explain this is with a simplistic working example:

XBeautyProducts, a fictional online beauty store, wants to understand who they should send an email to, in order to reach people with the highest propensity to buy their new moisturiser. They have limited stock and don’t want to market to the entire database. This is key because they are already suffering from a high number of email unsubscribes, so the message needs to be relevant.

Through data mining, they establish that 40% of their sales of the product have been to females, aged 60-75, whereas only 20% of their sales were to females aged 18-35. Therefore, this seems like the older market is the ideal target for the campaign.

In order to sure up their hypothesis, the decide to profile the data in such a way to remove any bias within the database.

Much of the advertising run by the company to gain new customers is within magazines that appeal to an older demographic. Further analysis highlights that this has created bias within the customer base, with 80% of their female customers aged 60+ and 10% aged 18-59.

The profile highlights, therefore, that older customers: 40% of sales come from 80% of the database. The 40/80 conversion can be described as an index of 50 - this shows that the group are half as likely than an average customer to buy the product.

Younger customers: 40% of sales come from 20% of the database. The 40/20 conversion can be described as an index of 200. This shows that the group are twice as likely than an average customer to buy the product.

By removing bias in the database, the profiling highlights that XBeautyProducts are more likely to have success by focussing targeting of the promotion to the younger group, which is a reverse of the initial data mining results.

For cold acquisition campaigns, customer data can be compared to the overall market at large, for example all Australian women Vs. all women that bought the product. This relies on the marketer gaining access to reliable statistics relating to the variable from external sources which might include census results (via the ABS) or data compiled by third-party aggregators or data brokers.

Predictive modelling

Predictive modelling can simply be described as the process of combining and leveraging relationships between data points to determine an exact (deterministic models) or a probable (probabilistic models) outcome. In marketing terms, these processes are often used to understand and predict the behaviour and makeup of customers and prospects, based on marketing and sales interactions.

Deterministic modelling

Data mining can sometimes be used for determinist modelling, as it can be used to determine exact relationships. This relies on definitive proof, with no room for variance (deterministic data points).

The data sets used to build these models need to be first-party / self-reported data sets that are not modelled themselves, or one is effectively building a model on a model.  

For example, if your customers must be logged into their account when transacting online with you and they have provided you with details of their age, you could link data from your website and your finance package to build understandings correlations between various product purchase patterns and the age of the related customers.

However, as no marketing truly operates in a vacuum, deterministic modelling has few real-world benefits.

Probabilistic modelling

Probabilistic modelling has more applications for marketing, as it is used to establish statistically viable, but not exact, correlations and related predictions.  

These might relate to likely impacts of existing campaign targeting, who customers are likely to be and/or how they are likely to behave next, according to how you communicate with them, ideally with a sensible degree of variance. 

Some examples of how probabilistic modelling can be applied:

The who:

If customer profiling establishes an extremely high degree of correlation (for example, an index of 200+), between specific purchasing patterns and a younger age group, one might assign the same modelled age group to customers exhibiting the same behaviour, for use in messaging and other targeting.

For example, by simplistically drawing on first name data from births registries, we might be able to predict that people called “Mildred” are likely to be aged 60+.

The behaviour:

By linking past behaviours and/or transactions to profiled segmentation variables the marketer can start to establish which segmentation variables might be predictive of that behaviour in the future.  

For example, in profiling customers that bought Product X, a marketer might highlight indexes over 150 (meaning more than 50% more likely to be the same than random) across variables including:

• Females
• People that also bought Product Y and Product Z
• People that did not buy Product A 
• People aged 30-39
• People within wealthy geodemographic segments, using a solution like Mosaic 
• People within 3KM of postcode 2000

Depending on the modelling techniques the marketer employs, they can now start to determine what this might look like at an individual level. 

One method to do this is CHAID modelling, also known as decision tree modelling, which establishes the relationship between each of the variables.

Decision tree modelling does this by looking at each variable in descending index order and splitting it into nodes. Each node is then scored to determine how likely it is to be predictive of singling out people that are similar. The diagram here shows how the model might start, with additional layers being added below according to lower indexing variables.

Once high scoring nodes are established, the marketer can establish how many nodes they need to select to build a critical mass for campaign selection.  

This might be represented in the form of a gains chart, where each node is shown in the model line. In this example, the Y axis represents the % of responders and the X axis shows the % of people marketed to.  

In this gains chart:

• The random line shows what would happen by selecting people at random (no gain)
• Each point on the model line represents a node and the related gain
• The hindsight line shows how well a perfect model could have been applied in retrospect, for comparison (the model finds everyone that will buy the product)

This model is projected to work better than random and is projecting that it could select 50% of the people likely to exhibit the behaviour by selecting only 20% of the total universe.

Where there is previous purchasing behaviour to compare the model against, it can be retrospectively tested, to determine it’s real world validity.

For example, if selecting the node representing the 20% of people that are projected to include 50% of the people likely to purchase the product from the gains chart above, the node can be compared to people that actually have purchased the product previously.

A simple Venn diagram is useful great graphical representation of this. 

The diagram effectively shows:

A – the number of people that previously bought the product that the model did not select

B - the number of people that bought the product that the model selected

C – the number of people that are projected to buy the product that did not previously buy it

If the value of B represents a significant portion of A, then model has been retrospectively proven to work well.

At this stage, the marketer will need to decide whether they wish to exclude customers in segment B, based on whether they are likely to re-purchase in real terms.

Important note:

It is also important when using modelling to consider that marketing campaigns do not operate in a vacuum. Correlation should not be confused with causation.  

Impacts of changes to the 5 Ps of marketing, for example, might have impacted previous behaviour, which may no longer apply to the current campaign (for example pricing influences of competitors / seasonal consumer sentiment).

An example that is often highlighted in the digital realm relates to digital advertising conversion attribution. A marketer might spend significant advertising budget in search marketing and digital display advertising. In analysing sales conversions, they might inadvertently attribute 90% of sales on a website to re-targeting campaigns, because they built conversion statistics based only on last ad click results. As the results did not take into account how the customers had arrived in the website in the first place, they cannot be taken to be causation. So, moving their ad budget to this medium only would be ill-advised, as it would not likely be sustainable.

 Conclusions:

Marketers have always looked to gain greater understandings of customer behaviour.  

The buzz words used to describe data-driven marketing processes shouldn’t detract from the importance of the outcomes available to marketers that can navigate their true meaning.

In an age of powerful technology and ubiquitous customer data, using the right methods to harness, process and leverage customer data is both challenging and essential, if we are to stay relevant to the customer and successful in our marketing efforts.

If you would like to learn more about how to leverage your customer data, please get in touch.