How Binge-Worthy is your Data?

One of the main tenets of digitalisation initiatives in financial services is to become more familiar with client preferences. Using client behavior data and feedback to be more predictive about their future needs allows banks and large investment managers to develop new products, and cross-market to existing ones, with minimal human analysis and intervention. The trend, therefore, has seen a proliferation of touchpoints in that digital relationship, providing the opportunity for users to express their opinion. At the design stage, the question is how and where to cultivate that data, and what to do with it.

Likewise, data specialists in this area face the reality that they are constantly compared against goings-on beyond the world of financial services. That is particularly true of their efforts at innovation. Whether deploying new technologies or digitalisation strategies, or disruptors’ approach to their respective markets, there is a lot to learn and potentially transfer (or even guard against). Gresham’s approach to innovation takes this into account; our experimental work last year with voice technology and virtual assistants for example, or the recent study that benchmarked Clareti Preon’s data compression performance against Google’s.

Sometimes, the comparison and relevance stare us right in the face. But sometimes, the news is a little more subtle. When major decisions happen quietly for a normally headline-making tech firm, the reasons ‘why’ might be interesting in their own right. One such case popped up late last year, and it involves the question above -- when the content streaming giant Netflix scrapped its well-established user-based review scoring model. Instead, its automated recommendation algorithm for delivering content suggestions to viewers would be more accurate, timely, and insightful.

Recommendation Questions

The Netflix move was immediately interesting because it seems to run counter to what we’ve seen prevailing elsewhere - the wisdom of the crowds. Many in the industry media posited explanations. Maybe Netflix believed the review feature was a poor indicator of audience preferences, and indeed had seen engagement decline. Some argued that the feature could be gamed, i.e. by collective and concerted efforts to tank viewership of specific content. And reviews could prove generally unhelpful for a company that is now actively producing much of its own media, in addition to distributing others’.

Bottom line: user opinions become distracting clutter - confusing the experience and running counter to Netflix’s strategy.

Still, it was surprising. Others pointed out that various forms of user-review tools remain prominently placed, avidly used and well-trusted online, everywhere from Amazon to Trip Advisor. The argument here runs a little differently: the Netflix shift was less about guiding the audience to new content experiences, and more about the insight Netflix could glean from the algo alone - compensating generously for the technology capital it had spent in the cause of doing so. It was a conscious decision.

This is where it gets more interesting. The simpler, manually scored, star-rating review system had already been phased out, and Netflix acknowledged that user review-based data had never been directly coupled up to recommendations to viewers anyway, nor was it even available on some devices. Reviewer data wasn’t ever really prioritised. By contrast, its algorithmic recommendation system represents one of the most famous, debated, and well-funded machine-learning implementations in the industry, if not anywhere.

So, perhaps, this was a case of using the old-fashioned recommendations systems as a temporary facade, spoofing users to believe their experiences were being influenced by ‘people-like-them’ while waiting for the rise of the robots!

Data Lessons Learned?

Philosophically, this asks whether we should care what users directly tell us about their experience, or trust an algo that bases its decisions on aggregate evidence - and this certainly applies to finance. With the vast amounts of data financial services firms can feed the algos – perhaps inevitably, the end is nigh for human input?

We saw that in the rush to develop robo-advisors. Machine learning can improve every stage of the data management process – detecting patterns that indicate malicious internal fraud or reducing the risk of costly settlement delays. By mining the vast lakes and pipes of streaming data, risks can surface triggering faster response times, reducing the chances of a loss, or identifying a new opportunity.

On the other hand, the complexity and consequences are greater in our industry than they are in choosing Russian Doll or Money Heist for a date night. For that reason, many of the most successful digitalisation initiatives on the leading edge of capital markets still incorporate the design around user engagement and feedback, with AI bouncing off of that information but not replacing it wholesale. Institutions want to know as much as possible about what clients are thinking and what they want (and don’t want). There is growing sensitivity to this behavioural information, be it to upvotes, machine-readable comments, the full correlation of what you do when at what time of day, and from which locations. But it’s still hard.

From a data management perspective, the question isn’t if but how?

There is complexity in developing and structuring the kind of data in question. It involves lining up timing and cause and effect, between transactions or other quantifiable outcomes and variables that are more qualitative and interpretable. As the Netflix case illustrates, it requires commitment across channels, obviously needs to be judged for reliability, and in short, needs to be in a usable condition. This places more demand on the data management process and infrastructure. More importantly, errors and trouble lining the data up mean an algo can’t adjust and respond precisely.

Herein lies the data integrity issue. Just because a reaction is fast doesn’t make it useful. On a micro-level, we’ve probably all seen this problem play out in real time: when you’re online and prompted to buy something or take an action that comes completely out of the blue. This is tough enough to get right for Netflix, a young and digitally-native company. For a bank with hundreds of legacy systems and thousands of use cases, it’s a far bigger challenge.

Playing It Out

Given their track record, it should prove interesting in the coming years to see whether Netflix’s decision proves the guide or a noteworthy exception. For now, financial services firms split the difference: take in as much as data as they can for analytics purposes, and do what they can with it with narrowly focussed AI.

Binge on the data while working to make it “binge-worthy”, in other words.

Read Paul's original article on FinExtra.