At Smart Vision we’re in a pretty strong position to talk authoritatively about the reality of predictive analytics. That’s because we’re comprised of a team of veteran practitioners with decades of experience where we’ve all witnessed plenty of success stories but also one or two ‘data science’ train wrecks.
Moreover, like anyone else, we’re exposed to the seemingly constant torrent of stories about the latest developments in machine learning, data science or AI. But we’re often struck by the fact that there seems to be such a focus on emphasising the power of analytics or on explaining how machine learning works that authors tend to overlook one salient truth: most advanced analytics projects fail because they never even get started.
Very often this is simply because people don’t know where to begin. With that in mind, we’ve always recommended that the smartest thing anyone can do when embarking on a predictive analytics initiative is to make a plan. This is precisely why analytical methodologies like CRISP-DM exist: to provide a framework for planning these projects. However, even methodologies like CRISP-DM can’t tell you what your project should focus on. It turns out that very often the most obvious targets for these initiatives are right under your nose.
Let’s begin with your organisation’s strategic objectives. What actually are they? What yearly goals are stated in the most recent annual report? A random search of annual reports yields strategic objectives such as “deepening customer engagement” (Vodafone 2019), addressing “key areas of risk and compliance” (RNLI 2019), “data to insight to action” (Thames Water 2019), “personalize guest experience” (Hilton 2018) and “bearing down on avoidance and evasion” (HMRC 2018).
Predictive analytics has a role to play in addressing all of these objectives, so any analytical initiative that does so has a clear mandate to state that it is targeting goals at the strategic level. That can be quite a powerful argument when you’re looking for internal sponsorship to get your project off the ground.
What about at the tactical level? Ask yourself, “what is being monitored in my organisation”? If it’s being monitored, it’s because a) it varies and b) it’s important to someone. The most obvious examples of this are shown in BI/MI reports and KPIs. So that means the focus turns to more specific targets such as net promoter score, revenue per available room (RevPAR), quote-to-close ratios, average cost per claim, patient readmission rate, median time to resolution, student retention rate, percentage out of stock items, average waiting times etc.
All of this merely emphasises that the target of the project should address something that is meaningful to the business rather than interesting to the analyst. So a few further considerations should be borne in mind.
- Pick something where you can quantify the potential impact of addressing it effectively. What’s the value associated with this objective? What are the costs associated with getting it wrong?
- Assuming you can use predictive analytics to predict the likelihood/estimate the metric/classify the outcome/quantify the risk, what will your organisation have to do differently to exploit this new insight?
- How will you it worked? How are you going to test the effects of these actions? One simple approach is to compare the effects of paying attention to outcomes that the new analytical approach generates, to those where you adopt a ‘business as usual approach’. Then, critically, quantify the difference.
Predictive analytics represents an analytical approach where the acid test is focussed on better decisions not complex algorithms. Sometimes the greatest insight can be gleaned, not from asking how something works, but why.
