Subscribe to DSC Newsletter

Prescriptive versus Predictive Analytics - A Distinction without a Difference?

Summary:  Is the addition of “Prescriptive” analytics to our nomenclature really worthwhile or are we just confusing our customers?

I admit to being annoyed when this or that industry wag tries to coin a new term to describe some portion of the discipline we are already practicing.  Some of these folks I think are well intentioned and believe they have seen an area sufficiently unique to deserve separate description. Some I believe are motivated more by the ability to sell more reports if there are now two disciplines to be ranked and rated instead of just one.

It is with this admitted mindset that I’ve come to the conversation about whether ‘prescriptive’ analytics is a real thing or is it just an aspect of what we’ve been doing all along. 

My concern has been that if we make our clients believe that this is some new form of analytics requiring new techniques and new tools or analytic platforms we will have made their perceptions less clear and our own lives that much more difficult.  Believe me, with only about 1/3rd of companies having yet adopted predictive analytics the last thing we need is introducing ‘the next big thing’ into that conversation unless there is some real value to be had.

So with this in mind I decided to focus on this issue over the last few weeks and see if I could convince myself one way or the other.  Could I find real world use cases where the distinction between prescriptive and predictive was important and easy to see?

The Definition

The most common definition for this difference is that predictive analytics describes what will happen while prescriptive analytics describes what should happen.  Perhaps a better way to say this was offered by TDWI senior researcher Fern Halper when she says prescriptive is about improving the outcome.

The graphic that she offers shows the ‘fuzzy transition’.  It should probably be bigger.

The Argument Against

I would argue that most of what has been broken off as ‘prescriptive’ is actually what we’ve been doing all along.  Let me offer two examples.

If I am a bank or insurance company and have say six major service offerings then one of the things I would like to know is if customer John calls my service center, then during that conversation what is the next most likely product my CSR should offer to John after his question is answered.  This is a very common application of predictive modeling and a simple scoring model will reveal that John is most likely to buy X if he already has Y.  I use this score to display on the CSR’s screen that X should be offered at the end of the call.  This described both what will happen (the probability of a cross sell of X) and simply displaying this information to the CSR defined what should be done and improved the outcome.  No additional secret sauce required.

If I am a medical provider trying to decide what portion of say the pre-diabetic population that I serve should receive a specific disease management program and which should not, I can use predictive modeling to determine the likelihood that each member will respond positively to the program. (See my earlier article on ‘Optimizing Disease Management Using Predictive Modeling’).  The disease management program has a cost but it is presumably lower than the cost of treating the member if he converts to fully diabetic.  Knowing that not all members will respond as well as others to the disease management program I can build a predictive model and from that prepare a simple lift analysis that will target my cost breakeven point.  We’ve been using lift analysis from predictive models for more than 20 years.  It describes what will happen.  It also describes what should happen, that is offer the disease management program to the portion of the population that is most likely to respond which best allocates my scarce health care resources.  Anything new here?  No.

The Argument For

I did find several use cases where prescriptive analytics required the use of additional techniques to move from what will happen to suggest or automatically initiate a subsequent action that could produce an optimal result (what should happen).  Curiously, they all borrow techniques from Operations Research, a field far more mature than predictive analytics so not new in that regard.  Specifically these examples all rely on mathematical techniques of optimization in the face of specific constraints (cost, labor, time, or similar).

It makes perfect sense that the learning from the predictive analytics should be tailored to deliver the best possible result in the face of real world situations.  Here are a few of the examples I found.

  • Price optimization.  A particularly good example is airline or cruise ship ticket pricing.  Especially in the case of airline pricing the best solutions are those that can occur in near real time.  The addition of the time constraint however does not differentiate predictive from prescriptive.  It does point to some current challenges in implementing and refreshing models and the realm of big data.
  • Optimizing parcel delivery routes and schedules.
  • Mechanical preventive maintenance.  The optimization here is the tradeoff between leaving the device in service (presumably revenue generating) and proactively removing it from service prior to breakdown (presumably reducing repair cost and time).
  • Staffing optimization.  This is a hot topic for hospitals and other health care providers but could equally apply to the factory floor or service industries.

Are Prescriptive Analytics Different in Big Data?

I did discover one application of what is described as prescriptive analytics specifically relating to big data, in this case the analysis of unstructured text data.

More than one big data analytic platform with a healthcare focus offers a type of text analysis in which they aggregate the text descriptions of different steps in the treatment of specific medical conditions.  Take for example back problems leading up to surgical or non-surgical resolution of the problem.  Through clever analysis relying more on visualization than numerical quantification this helps medical professionals identify the optimum series of procedures leading to the best outcomes for different subsets of this population.

This meets the definition of prescriptive analytics (improves the outcome) but to my eye looks more like visualization in big data than something uniquely different from highly quantified predictive analytics.

Wrapping It Up

OK so my little adventure in self-education did indeed find some examples of prescriptive analytics identifying improved outcomes.  And these examples did involve analytic steps beyond what some of predictive analytics usually encompasses.  However, the use of optimization techniques on predictive models is not new.  Those of us who come from operations research or a statistics background have been applying these techniques all along.

On the customer side I’d say the impact is mixed.  By adding complexity through the ‘invention’ of another type or dimension of analytics we’re certainly making it more difficult, not easier for them to understand.  And if you’re dealing with customers who are low on the ladder of adoption I’d say, watch out, too many choices is the mother of indecision.

However, I must admit that the definition of prescriptive analytics with its immediate call to action – tell me specifically what to do to make it better – can be less theoretical than explaining the whys of predictive analytics and even may be the compelling message that we need.

 

October 23, 2014

Bill Vorhies, President & Chief Data Scientist – Data-Magnum - © 2014, all rights reserved.

 

About the author:  Bill Vorhies is President & Chief Data Scientist of Data-Magnum and has practiced as a data scientist and commercial predictive modeler since 2001.  He can be reached at:

[email protected]

The original blog can be viewed at:

http://data-magnum.com/prescriptive-versus-predictive-analytics-a-d...

Views: 12679

Comment

You need to be a member of Data Science Central to add comments!

Join Data Science Central

Comment by Jack Lampka on September 22, 2017 at 3:40am

It's interesting that William's article from October 2014 somehow made it to the "18 Great Articles About Predictive Analytics" newsletter from September 2017. I wonder why, but at the end the same problem still exists three years later: new terms being created mainly to sell more reporting and consulting (and to overwhelm business folks). Although over the past three years "prescriptive analytics" has been defined more clearly, I avoid bringing it up if not needed. If asked by clients or in workshops, however, I describe prescriptive analytics as an extension of predictive analytics where a decision is made without human intervention. I use as an example a manufacturing line where the input mix, temperature, pressure, etc. are automatically adjusted by monitoring output and predicting deteriorating output quality.

The graphic posted below by Pat nails it. I personally believe that predictive analytics outcomes should not be called prescriptive analytics if they involve a human to implement. If a human is needed for implementation, it's nothing else than a business decision that usually follows any analysis. In that case any analysis could be called prescriptive analytics, adding more to the confusion ... add allowing consultants to sell more reports and ranking scores as William pointed out. 

Comment by Irv Lustig on September 19, 2017 at 12:47pm

The term "prescriptive analytics" is something myself and some other IBM colleagues wrote about back in 2011. See the article http://analytics-magazine.org/the-analytics-journey/ .  In that article, we put optimization in that category, and nothing else.  One could argue today (late 2017)  that there are other techniques that can be called prescriptive, but there is a real distinction.

Comment by Mark Meloon on July 28, 2015 at 5:50am

I appreciate your efforts to keep things simple and for being vigilant against hype, but I think there's a very important added complexity between predictive and prescriptive approaches that I haven't seen any thought leader talk about: causation.

Your examples in the "against" section are all stationary processes (e.g., time does not appear anywhere). So a predictive approach is certain sufficient. In your "for" section, you correctly point out that simply adding time into a process does not automatically make it a problem for the prescriptive approach. You can still use predictive modeling with optimization.

The cases where you need prescriptive approaches are those where you need to model the changes your actions cause to the world state in order to solve decision problems over multiple time steps. In mathematical notation, if you need to model how your action at t=t_0 causes changes to the world state from t=t_0 to t=t_1 in order to decide what action to take at t=t_1, then you should use prescriptive analytics. If you want to neglect that effect, the predictive approaches are fine.

Said more simply, predictive analytics is about correlation. Prescriptive analytics is about causation.

In your bank customer example above, you model the sales pitch as an instantaneous event so predictive analytics are fine. If you wanted finer granularity and to model the sales pitch as being of finite length (e.g., a sequence of time steps, each with their own opportunity for a unique action) then you'd want to model how one statement you make causes the internal state of the prospect to be more or less receptive to the next statement you make.

Is prescriptive analytics necessary? Like all things, it depends on the problem being solved. If you are getting information about the effects your action have on the system fast enough, then you can break the multi-time-step process into a bunch of one-step problems and use predictive analytics. After all, receiving real-world data on the state of the world is better than settling for your estimated changes. But if the action is happening faster than you get new info and/or your actions have significant effects on the state of the world, then using prescriptive approaches -- which explicitly have a causative component in the model -- is the way to go.

Comment by Matthew O'Kane on May 16, 2015 at 4:04am

I agree that people are using the term rather loosely but I believe there are techniques,both old and new, that are specifically prescriptive rather than predictive analytics.

The simplest example is A/B or mult-variant testing. By understanding the average treatment effect for alternative decisions we can 'prescribe' the one with the most impact. The only issue with this approach is that you are only observing an average for one population, segment etc.  but there are many techniques based on machine learning such as uplift modelling, net lift modelling, heterogeneous treatment effect models and such which move towards an individual treatment effect so that we can 'prescribe' the best treatment on an individual basis.

To apply this to an example, think about churn management.  A predictive analytics approach would be to predict a customer's probability of churning but what should I do with a customer with a 99% chance of churning?  Should I definitely call him and offer him my best offer?  Or is he a lost cause and I should not waste my time.  Now a prescriptive analytics approach, say using uplift modelling based on uplift random forests, would tell you the change in churn probability for each treatment you test.  We can use this information to 'prescribe' the right action - in some cases this will be 'do nothing' if every other treatment has a negative effect on churn.

Comment by Paul Capicik on October 30, 2014 at 8:55am

I view this as just using terminology to help in better analytics process understanding. To me the difference between predictive analytics and prescriptive analytics is similar to what the difference between analytics and predictive analytics would have been before the invention of that term. It can help in understanding the full details of an involved process as long as everyone agrees to the definitions. As hardware, software, and data capabilities have evolved, the addition of predictive and prescriptive to the word analytics seems to indicate a higher evolution of steps to take to result in a better solution or decision. But if you don't understand the new words, or you are a diehard that says analytics alone tells me to do it all, then this discussion is moot.

Another thought on prescriptive analytics is that it is a two step process, once you do predictive analytics you will generally 1) do plain analytics to determine what options exists based on the prediction, and then do predictive analytics on each option to see which path is the best option.

The chart that Pat added makes things very clear and so the use of all terms are very useful to me.

Comment by Pat Lapomarda on October 27, 2014 at 10:00am

The difference between predictive & prescriptive is how much closer to "action" the analytic gets us.  Prescriptive implies taking an action based upon the result, whereas predictive just lets us know what's likely to happen next.  It's the same difference as between descriptive & diagnostic analytics.  Both tell you "what" happened, but diagnostic gets you closer to action by including the "why" context.  In this case, both predictive & prescriptive will tell you "what next", but prescriptive gets you closer to action.  This image from Gartner does a good job illustrating why they make a distinction:

Comment by Wayne G. Fischer, PhD on October 27, 2014 at 9:45am

I think the latter half of this blog's title captures it.  Whether we waste time taking about "predictive" versus "prescriptive," perhaps we should all keep in mind that all these methods have been around - and used - for decades (probably longer).  It's called modeling / simulation / optimization (MSO).

The only thing relatively recent is the ubiquity of low-cost, high-power computers and the software, that together allow many more people to apply MSO to many more situations (some of them even appropriately).  :-)

Comment by Ralph Winters on October 24, 2014 at 10:21am

OK.  I get it.  Prescriptive Analytics is sort of predictive analytics with optimization and scenario 'what-ifing' built in. True.  We've been doing this for years.  However, this definition does seem to make the whole thing a little more 'Data Sciency' and probably would appeal to companies who like to have action plans, all in a format that is more business rules oriented,  and less dependent/independent variable R-square oriented.

Comment by William Vorhies on October 24, 2014 at 8:22am

Ralph:

I've seen this too and I've wondered if Gartner isn't driving this distinction  just to increase profits by creating two different ratings.  While I believe that Gartner is correct in identifying predictive analytics as mature (and therefore presumably safe for late adopters to embrace) that doesn't address what to me is the illogic of a market in which only about 1/3rd of companies have adopted it.

Comment by Ralph Winters on October 24, 2014 at 7:52am

If you a believer in the Gartner Hype Cycle "Predictive Analytics" has reached the Plateau of Productivity, and thus fallen off the chart, while "Prescriptive Analytics" has appeared and is making its climb towards the "Peak of inflated expections".  Whether or not it is an improvement upon Predictive Analytics, or just a rebranding remains to be seen.

http://www.pewresearch.org/fact-tank/2014/08/15/chart-of-the-week-t...

Follow Us

Videos

  • Add Videos
  • View All

Resources

© 2017   Data Science Central   Powered by

Badges  |  Report an Issue  |  Privacy Policy  |  Terms of Service