Go Beyond Predictive Modeling ... Creating the Virtual Prescriptive Tool

So you think that the predictive model that was just rolled out should be good enough to help your company get close to its goal. You are trying to predict how many claims processors you will need to have to run next month’s shift base on how your incoming claims are trending? You want to predict how much of the market you can capture with the new acquisition? You are expanding and want to know how many of the customers will be up for grab?

 

As very good predictive analysts, just give us the problem. Revenue is the problem? We can identify the drivers that are impacting our revenue. We can do the modeling without much sweat. A little of data, structured, unstructured, complete, incomplete, dispersion, no dispersion, normal or non-normal - never mind! A little of brain power, that is normal. Oh, wait a minute, our coefficient of determination, denoted with R2 or r2 which indicates how well our data fit the statistical model is telling us we are coming at adjusted 52 percent. Not good!!! That means we can only account for 52 percent of the drivers influencing our revenue? Not good!!!

 

Can the Coefficient of Determination R2 Be in the Lower 80 Percent?

Let’s see, we want the variance explained by the model to be in the lower 80 percent! Low or mid 70 percent will be like a coasting solution!

We need to add a layer of analysis - How about jackknife? No, we are moving a little bit higher to bootstrapping. Non-parametric or parametric bootstrap? We are even thinking of Bayesian bootstrap. Wait; is there someone around our team who knows whether our software can handle maximum entropy or MaxEnt?  How about simulation outright? Oh – yes, let us try simulation. Depending on what business problem of revenue we are trying to solve for and our data, we are not going to use deterministic, we are using stochastic simulation!

 

Beyond Simulation, Can we Use Optimization?

So after we reran the analysis and allow for complex interactions, we are now in the 68%. We think this is good enough but that is 32 percent variance that we are not accounting for! How about optimization? Yes, let us try optimization - constrained or unconstrained optimization?

Finally, we are at 70 percent of what drivers are influencing our revenue. We have got to move to another deliverable but then there are still 30 percent of those drivers still out there!

Well, here comes the Virtual Prescriptive Analytics!

 

Virtual Prescriptive Analytics

No, doubt, there are need for some of the efforts describe above. Virtual Prescriptive Analytics allows the analysts and the business leaders to actual play with additional variables that may or may not be included in the original model above.

Some of these do not have to be part of the original variables that were used in the original predictive model. The original predictive model will need to include depends on long historical data so as to be stable. The Virtual Prescriptive Analytics still needs the predictive modeling. The predictive model has to be converted into business rule engine. The rule engine is then exposed to different variables. These variables include economic, social, market, consumer behavior etc. These variables can have long, medium, short or one-off  history!

 

Exposing Variables to Rule Engine

As these variables are exposed to the rule engine, the effects become the ingredient for the business leaders to run their businesses.

The Virtual Prescriptive Analytics allows the business leaders to run their business in five year today!!! Just like the pilot can create different weather, crash, speed, aircraft structural integrity scenarios in a flight simulator, so also can the business leaders do the same in terms of growth, acquisition, merger, off-shoring, selling business unit(s), revenue and profit margin in Virtual Prescriptive Analytics.