by My interest in predictive analytics was first piqued by a presentation by Garry Golden at the 2008 APF Gathering on Big Questions. As too often happens these days, I got distracted and only recently committed to digging in. I asked around for a good “intro” book on predictive analytics and was pointed to Eric Siegel, Predictive Analytics: The Power to Predict Who Will, Click, Buy, Lie, or Die, Wiley, 2013.
If I could come up with an excuse, it’s that I’d sat through dozens of data/text mining demos over the years that looked really impressive – until I tried to do one myself. The topic is not one of great intellectual interest to me compared to others, so I didn’t try as hard as I might have.
But let’s get to heart of the matter. Predict analytics (PA) is defined as “technology that learns from experience (data) to predict future behavior of individuals in order to drive better decisions.” The author, Eric Siegel, is a big name in the field and he set out to write an accessible bestseller-type book, and he succeeds in that. It is currently the #1 best seller in Amazon’s Business Forecasting and Planning category. [For what it’s worth, Thinking about the Future, occasionally sneaks into the Top 100 in this category]
A key distinction is the “scale” of concern. PA focuses on the micro level. It is aimed at predicting at what individuals might do, what will they click or buy? Siegel contrasts PA with forecasting by saying that forecasting makes aggregate predictions on a macroscopic level. That’s helpful context, but a question that’s left unanswered and nags me is whether it will someday graduate to the macro-level.
There are lots of stories in the book and lots of example – a section in the middle has 147 examples. While some might find the examples a bit mundane — like more effectively targeted mass mailings – there’s big dollars involved and well, 147 examples is nothing to sneeze at.
A lot of the examples seem to mirror common sense, but common sense with data behind it is more effective than without it. My quick-and-dirty “current assessment” of what’s new:
- We have a lot more data that we used to – a really, really lot.
- The quality of the data is higher – better tagged and organized
- And here’s the punch line for me – improved machine learning, which is basically crunching the data to build the predictive model, which predicts the behavior of an individuals by taking their characteristics as input and providing a predictive score at the output, e.g., credit rating is a score, the higher the score, the better your rating and the higher likelihood you will pay back what you borrow
- The field is excited about the “ensemble approach,” which is basically aggregating multiple models around the problem under investigation
The machine learning part seems to comprise chiefly of putting together multi-variate decision trees based on historical data. Siegel notes there is still some art in there, in the sense that one has to know when to stop branching, because after a certain point, over-learning occurs and one can mistake noise for information. There are testing protocols to protect against this.
It seems pretty obvious to me that predictive analytics is going to continue to get better, as better data, improved machine learning and better modeling in the future seems a pretty safe bet.
Siegel raised a really interesting comparison of differing schools of thought. So, predictive analytics is part of a larger suite of capabilities going doing the prediction path, while Taleb and the Black Swan folks suggest the futility of such endeavors – again, keep in mind that PA in particular is aiming at the microscale and Taleb focused on the macroscale. In support of the predictive tools school of thought, the Houston Foresight Program’s “Student Needs 2025+” study for Lumina Foundation identified the growing use of IT/AI/PA etc. tools as a huge driver of change across all of student life (which we broke into living, learning, working, playing, connecting, and participating in student life). Teams in each of these six areas came to the same conclusion, and when we tried to develop plausible arguments for this not happening, it was, well, tricky, beyond some sort of system collapse.
So, where does that leave me? I think our Houston Foresight curriculum needs to incorporate some PA, and the hard thinking now is “how much?” It’s still a bit unclear to me where PA fits in a futurist’s tool kit, but it seems to fit. I had hoped for a bit more in the book on the future of PA, but got very little. Admittedly, the book didn’t promise it. And I suppose, thinking about the future of PA seems like a task suitable for futurists! Andy Hines
Thank you for sharing your thoughts. I have also wondered about how mature and wide-spread PA approaches are today. The most inspiring examples I have come across have been described by Jay Galbraith in his article on “Organizational Design: Challenges Resulting from Big Data” (available at http://dx.doi.org/10.7146/jod.8856). In the article he describes how consumer goods companies, such as Nike use big data mining to predict consumer patterns and shape them than in real-time.
I have to agree that we better insert some PA in our Strategic Foresight/Futures curricula.
Andy re: your “question that’s left unanswered and nags me” (whether PA will someday graduate to the macro-level) I’m willing to confidently predict that NO is won’t graduate to the macro-level, with one important proviso. I’m assuming that human societies remain highly dynamic, and don’t move towards the Soviet-style tightly state-controlled models. Your question also reminds me of Isaac Asimov’s famous Foundation book series, in which folk tried to develop a field called “psychohistory”. Have you read this series? I confidently predict that it will remain science fiction. Cheers, Stephen
Good point. I do remember Hari Seldon and psychohistory!
Regarding macro vs. micro..PA works at the macro level just fine; I’ve done it lots of times. After all, numbers are numbers, data is data…For example, I’ve built circulation models for two different major newspapers that you know but I am not allowed to divulge. For one of them, we predicted stops each week in aggregate, not the likelihood an individual subscriber would stop. Similar to a time series approach but including more “exogenous” variables than one would typically include with ARMAX models. This was used operationally.
A second example was predicting acquisition likelihood by zip code. Interestingly, this turned out to be the reverse of Eric’s “model at the micro, describe at the macro”. In this case, I modeled at the macro level and applied the model to the micro. Since the demographic data we had was at the ZIP level not the rooftop level, I built models at that level. Then if a new potential subscriber (from a list) lived in that ZIP, he/she would get the ZIP score.
I hope I’m understanding the distinction you are placing on micro/macro here. But to me, the level of analysis is irrelevant from a technology perspective. PA can be applied to either as long as the data supports the business question being addressed by the models. I cover this to some degree in my Applied Predictive Analytics book. How to set up problems for PA is a great topic and deserves a book unto itself (I only have part of a chapter devoted to the topic).
Dean
Hi Dean…thanks for adding the comment and perspective. I’m using the micro/macro distinction that was in Eric’s book. I’m not knowledgeable enough at this point to declare a strong position. I’m more in the “gathering” phase. Your comment is very helpful. I guess, intuitively, I don’t see why PA couldn’t work to some degree at the macro scale — using Eric’s notion that some predictability is better than pure guesswork. But to refer to Stephen’s earlier comment, I’m not a hari seldon pyschohistory guy either. Appreciate your sharing!
nhkxqt