[This post is a nod to this famous editorial. I don’t want you to think I’ve gone crazy. The subject of this post was inspired by the Lustig/Dietrich/Johnson/Dziekan article “The Analytics Journey” and the last bit (regarding skills) by a conversation with my boss.]
Eight-year old Virginia O’Dual writes:
Dear Blogger: I am eight years old. Some of my little friends tell me that prescriptive analytics is just a fancy way to say optimization. My advisor tells me that if it is on your blog it must be so. Tell me, is it so?
Virginia, your little friends are wrong. Analytics is not the same as optimization. I realize it may sound strange to be so blunt – after all, this is merely semantics, right? The internet is filled with people ceaselessly arguing about terminology. Millions (billions?) are afflicted with the disease of thinking that knowing the name of a thing is the same as knowing it. Virginia, you may be asking why this is any different. Am I not falling into the same trap by insisting there’s a difference? On top of that, why pick a fight about “analytics”? The word “analytics” is on everyone’s lips these days: on twitter, in magazines and newspapers, during coffee breaks at conferences. More often than not, rushes of hot air follow. So why on earth am I telling you there’s a difference? Virginia, the reason is that there is a connotation to “prescriptive analytics” that is missing from “optimization”:
Optimization speaks to process and prescriptive analytics to purpose.
[Numerical] optimization is finding the best option from a set of alternatives for a mathematical model. In more staid company it is called “operations research”. In less refined circles some simply speak of “search problems”, but this detracts from the majesty and beauty of the techniques that are so well-known and well-loved by practitioners. The theories and techniques of optimization date back centuries, though they were not labeled as such until much more recently. Optimization usually implies a formal mathematical model and a clearly articulated method that provides a guarantee about the quality of the alternatives are discovered.
Virginia, you read this blog, so you already know all of this. You’re an optimizer. You’re proud. Proud of your heritage, proud of the sound reasoning behind your methods, and proud of the profligate applications of your craft in matters of war, peace, and profit. More or less. You may consider those who use highfalutin terms such as “prescriptive analytics” to be late to the party, particularly when they speak in the language of business rather than in the more precise language of mathematics. Skepticism abounds; in the past I have even called analytics “old wine in a new bottle“. The fact that the terminology comes with a sales pitch doesn’t help: a cluster of computers, consulting services, a web service. If you listen to the hucksters, everything under the sun seems to be an essential part of an analytics “solution”: a filesystem, a database, a charting library. Even worse, the aspirations the “analytics community” speak of seem so depressingly modest. A pivot table is not analytics, we optimizers snort. We can factorize matrices already, thank you very much. Yes, my data fits on my machine. Go away so I can continue to bear the world on my shoulders like Atlas (or Dantzig). We know what you are talking about, we say to them, and each other. You are talking about optimization. (I notice a similar relationship between statisticians and the term “predictive analytics”.)
But prescriptive analytics really is something new. I have come to understand that though optimization is the primary tool of prescriptive analytics, it would be very limiting to conflate the two. Prescriptive analytics refers to turning data into decisions – but not just any decision. Decisions that are to be made by people, answering questions posed by people. This is not always the case for optimization. Sometimes decisions produced by optimization are in service of another software process. Other times the decisions aren’t really decisions in the everyday sense. Root finding and regression are properly speaking optimization, even if we don’t always label them such. Other times, the answers produced by optimization are far removed from the questions that were originally posed. “Consider a spherical cow“, a professor once said. In many cases, such as scheduling, supply chain optimization, or routing there are so many transformations, simplifications, and layers of rewriting from prose to math to AMPL (or GAMS, or AIMMS) to C and back have been undertaken that it is hard to connect solution to problem. The plot is lost. I wrote my Masters thesis on semidefinite programming formulations for truss topology design – bridge building. During my defense I was asked by a committee member if I would drive over a bridge I had built using the results of my formulation. The correct answer was “hell, no” – because I was trying to solve an optimization problem and not a prescriptive analytics problem. (I wasn’t brave enough to say so at the time, of course.)
Prescriptive analytics implies a specific purpose – building a system that automatically produces decisions to choices that humans face. As a wise man once said: “begin with the decision in mind and see where it leads you“. In the practice of prescriptive analytics this may lead you down unfamiliar roads, but brainpower and software are always provisions for the journey. Software that lies at the center of data, models, algorithms, and computational platforms. Sometimes the job simply boils down to assembling data. Real time data, historical data, disjointed and incomplete data that must be joined, refined, summarized, synthesized and completed. Once this is done we are left with nothing but a measly linear program, or even more pathetically, a sort. There is no shame in this – there are no bonus points awarded for complicated formulations in the world of analytics.
Prescriptive analytics concerns an entire process, and this process typically involves assembling data, building models, evaluating them, and presenting the results. Optimization comes into the play for the middle two steps, but the first and last are every bit as important. In fact, many times the first and last steps – assembling the data and presenting results in a form that people understand – are the most difficult and time consuming ones. And you do need tools to help you to pull them off. The problem with the analytics crowd, and the reason why optimization people sometimes roll their eyes, is that analytics people often focus entirely on the edges and ignore the middle: formulating and solving a real model. In prescriptive analytics we give all steps their proper attention, because without them the goal of delivering actionable plans for success cannot be attained.
So, Virginia, this offers fantastic opportunities to those who choose to take up the challenge of mastering prescriptive analytics – young and old alike. A practitioner of prescriptive analytics needs to understand not only the formal methods of optimization, but much more besides. They need to understand how data is obtained, manipulated, and joined together. They need to understand the art of model building, which is experimental in nature and requires knowledge of the original problem domain. They need to be able to communicate the insights produced by their models in words and in pictures. Finally, they need to be able to understand the business objectives and motivations behind their projects, so that they can not only deliver what is being asked for, but anticipate (and answer) the questions that come next.
2 thoughts on “Yes, Virginia, there is a difference between optimization and prescriptive analytics”
Thanks for the link and the kind words (never been called wise before). The only thing I would add to your post is that increasingly predictive analytics are being embedded in software – what I call Decision Management Systems. Often combined with explicit decision logic and sometimes even with optimization predictive analytic models are being used to score transactions or customers so that these predictions (probabilities if you like) can be used as part of an automated decision.
These three technologies – predictive analytics, business rules and optimization – can and should be used together. Business rules can specify the regulations or policies that need to be applied to the decision. Predictive analytic models can turn historical data into usable assessments of risk, fraud or opportunity. Optimization can be used to manage the trade offs between decisions or even within a complex decision.
And, as you say, if you begin with the decision in mind you can focus on the right techniques and technologies for the decision you care about it, whether you plan to have a person make it or a computer.
Thanks James! I have been a long-time follower of your blog and I appreciate your insights. I certainly agree that the trend of embedding analytics into decision management systems to automatically produce decisions will continue. Certainly that is the case here at Nielsen.