The Problem with Marketing & Analytics
Sometimes it just isn’t obvious.
Background
Customer science is a topic I’m deeply passionate about. I started my career in Marketing in the Singapore subsidiary of a Citibank. I was a product manager in the 1990s when data-driven marketing was not the default; campaigns were not properly evaluated, there was no rigour to the P&L simulation. As a product manager, I undertook my own campaign analysis, working with the IT department to figure out how to extract the necessary data. Data Governance was also non-existent at that time, and hence all kinds of “backdoors” could be established to get the necessary data feeds. This early start in “data entrepreneurship” eventually led to my work in shaping the evolution of the datawarehouse and data analytics capability for the Singapore business. I eventually grew into the regional CAO role.
A recent advisory engagement gave me an opportunity to reflect on the perpetual tension between consumer product & marketing (P&M) folks and their data analytics / data science partners. This tension seems to be underpinned by the mis-aligned expectations and understanding on how data can be used to improve P&M activities and outcomes. And so I dedicate my 53rd weekly article to unpacking this tension.
(I write a weekly series of articles where I call out bad thinking and bad practices in data analytics / data science which you can find here.)
Tension in Profiling
Given the availability of data and tools, every P&M person is now expected to be data-driven in their day-to-day work. But this is where a little knowledge can sometimes be unhelpful. Consider a typical question that P&M folks might ask their data analytics partners: “Give me a profile of my most valuable customers, and tell me how I can acquire more of them.” On the surface, this seems like a simple and useful question. But it isn’t.
Now I can go ahead and profile the valuable customers (assuming we have converged on what “valuable” means) by describing their demographics, utilisation of the organisation’s products and services, and their interactions with the said organisation. But what the P&M folks are really asking for is whether the data can help them separate the valuable customers from the non-valuable customers. The reality is that demographics have very little separation strength; products and services utilisation is naturally correlated with value and hence is self-referencing (i.e. valuable customers will use your products more frequently). Furthermore, the profile and behaviour of valuable customers NOW tells me nothing about where to acquire more of these customers, as the current information signals are out of sync with when these customers were en route to becoming valuable. I’ve learnt that such profiling exercises are mostly not actionable as they don’t help to improve the decision-making on acquisition, cross-sell and relationship deepening.
The concept of profiling resonates with many P&M folks who grew up on the use of personas for their work. They use it to shape their acquisition and marketing strategies. But there is a ton of research that indicates that using personas to develop value propositions or marketing actions simply don’t work as they are not rooted in evidence (see this article). But persona-thinking is simply hard to kill. And P&M folks become irate when they realise the data isn’t doing what they thought it should.
Tension in Modelling
The other challenge in using data in P&M is predictive modelling. There’s a bunch of mis-conceptions in practice here. Predictive modelling in P&M is closely associated with campaign activities. But your campaigns are constantly evolving (and they should) in response to competitive pressures and changing customer preferences. Organisations don’t have the luxury, nor is it practical, to build campaign response models, i.e. predicting customer response to a specific campaign offer (this is the right thing to do technically speaking). Instead, organisations build generalised models on the likelihood to have a need, and then figure out the right campaign configuration that might appeal and fulfil that need. This means that campaign responses are experience-based or intuitive estimates, and are not derived from the predictive models. So when the campaign doesn’t work, was it because of the model or was it because of the campaign design? This makes model fine-tuning an impossible exercise. This is very different from predictive modelling in Risk or Fraud where the problem statement remains constant. In such situations, you can constantly refine the model to improve its efficacy.
Building model to predict customer needs, instead of campaign responses, requires a lot of thoughtfulness and ingenuity. Internal data is often insufficient to detect the appropriate information signals. Furthermore, there is the confounding problem with brand affinity — you can get it right in predicting the need, but it doesn’t mean that the customer wants the need to be fulfilled by your brand. This reminds me of an interesting project I did with a bank a couple of years ago where we wanted to target corporate clients who might require remittance services from the bank. The consult was to create a model to predict likelihood for remittances services. Together with my team, we ended up creating a model that predicted changes in a customer’s remittance take-up (because that’s the data we had) while optimising for false positives; the argument we made was that the false positives were actually customers who had remittance needs (and were already remitting with another bank) but may not remit with said bank. We got the bank’s relationship managers to call a sample of the false positive to validate our assumption, which turned out to be generally true. Next, the campaign had to be designed specifically to encourage the customer to switch their remittance relationship from their current bank to the said bank, since a corporate customer with a remittance need would naturally seek to have that need fulfilled immediately. A P&M manager may not necessarily see these intricacies or have the appetite for competitive promotional pricing, believing that data-mining their existing customer base for potential would be sufficient for them to generate value. (I’ve written here about harvesting knowledge from your campaign activities.)
Conclusion
The tension between P&M folks and their data analytics partners often comes down to operating on different wavelengths with regards to what it means to leverage data for problem solving. Despite decades of exposure to data-driven solutioning, P&M folks may still not fully understand the subtleties of getting the right information signals that will directly be useful in their decision-making process. The data analytics community can go some way to mitigate this disconnect by anticipating it, and explaining what the requested data or solution approach represents and their associated limitations.
