Analytics is a very broad and evolving field that helps companies answer business questions about what has happened in the past, what is likely to happen in the future, and what decisions they can make to take advantage of these likely future developments. Reflecting the importance and growth of analytics, more and more universities are adding analytics degree programs, and many companies around the world are creating analytics departments.
Somewhat surprisingly though, even though prescriptive analytics, the decision-making part of analytics, provides the largest potential business value, it is just now gaining widespread adoption. In fact, Gartner expects the number of companies using prescriptive analytics tools, such as math programming solvers like the Gurobi Optimizer, to increase from ~10% of companies in 2016 to over 35% by the end of 2020.
While prescriptive analytics is growing, analytics overall is still dominated by descriptive (what happened in the past) and predictive (what is likely to happen in the future) tools. As a result, it’s important to take a step back to see where prescriptive analytics fits into the broader analytics landscape, and how all three types of tools can work together to help companies make better decisions.
One analytics framework, already alluded to above and discussed by Tom Davenport (external link), that seems to offer useful perspective on the main components of analytics is as follows:
Descriptive Analytics gives you insight into the past and current state of your business through the use of business intelligence tools. These tools can help you obtain a range of insights into your business, such as how much of a given product you sold over a certain time-period or your current inventory levels for various products in your distribution centers. Most every business function in your company (Production, Sales, Finance, etc.) likely is already using Descriptive Analytics in the form of recurring or custom reports.
Note that some paradigms further divide descriptive analytics into two categories, with a new category, diagnostic analytics, focusing on understanding why something happened. However, the general trend seems to be to simplify things and just include both "what" happened and "why" within the scope of descriptive analytics.
You can learn more about Descriptive Analytics (and the related field of descriptive statistics) here [wikipedia].
Predictive Analytics seeks to give you a glimpse into the future. It takes existing data and applies statistical techniques (often using machine learning) to help provide insight into what the future may hold for your business. The resulting predictions may be very coarse-grained (e.g., expected industry growth or raw material pricing), company-centric (e.g., revenue or profit growth), or operational (e.g., expected changes in demand by product line).
You can learn more about Predictive Analytics here [wikipedia].
Prescriptive Analytics, the application of computational sciences to optimize the set of decisions one should make in a given situation, often through the use of math programming models, is considered by many to be the most valuable part of analytics because it can be used to recommend specific decisions you should make in order to achieve a desired business outcome.
By building models of your decisions and using, for example, math programming solvers, you can quickly evaluate trillions or more possible combinations of choices (for example: in which order should you produce what products in which manufacturing facilities on what product lines and in what quantities), subject to a range of constraints (for example: minimum production of a given product, required manufacturing time and cost on a particular machine, raw material inventory, finished goods inventory capacity, etc.) to maximize or minimize your objectives (for example: total product costs).
Prescriptive analytics tools provide analytics professionals a number of additional benefits, enabling them to not only reduce decision-making risk, but also freeing up their time to focus on higher-value efforts such as performing scenario analysis or considering larger strategic questions. This increases the importance of the role of the analyst in the organization and allows them to take part in a wider range of important work. You can learn more about getting started with Gurobi and Math Programming by getting your own copy of Gurobi and visiting our Getting Started page.
You can learn more about Prescriptive Analytics here [wikipedia].
Math Programming Solvers are the primary tool used in prescriptive analytics. They help to turn data and models into smarter decisions. Using any of a broad range of programming languages, users can state even their toughest business problems as mathematical models, then call a solver to automatically consider trillions or more possible decisions to find the best one. Solvers such as the Gurobi Optimizer can be used as both a decision-making assistant, to help guide the choices of a skilled expert, or as a fully automated tool to make decisions with no human intervention.
The ability of solvers to quickly consider large numbers of business constraints and decision variables allows them to quickly consider in minutes or days many more choices than a human brain could consider over the course of many years. This capability allows companies to refine the way they currently make decisions, enabling them to efficiently and effectively take many more factors and decision options into consideration than before. The result is superior decisions that take less time to make and drive better business results. You can see a wide range of examples of how Gurobi customers have used mathematical modeling to make smarter decisions on our case studies page.
To illustrate the difference between predictive analytics and prescriptive analytics, consider the following situation. If we use machine learning techniques to detect potential cases of fraud and as soon as we detect a potential case of fraud we investigate, then prescriptive analytics is not needed. This is because each decision is independent since once fraud is detected we take action. However, real-world decisions are rarely so simple as there is often a dependency between decisions. For example, in the case of fraud, given:
Given this interdependency between the various decisions (which to investigate, who to assign each case to and what the relative priority of each case is), this is best handled as a prescriptive analytics math programming problem.
While there is a natural tendency to think the three parts of analytics are sequential stages a company must go through (from Descriptive to Predictive to Prescriptive), there is a strong argument to be made that the best approach is actually to move in the opposite direction. That is, companies should start with the types of decisions they want to make more effectively and then work backward towards the data they will need to help best make those decisions.
Starting with the decisions to be made rather than all the available data helps avoid a common analytics “trap”. This trap happens when organizations aren’t clear on what they are solving for, and analysts end up trying to “boil an ocean” of data to see what comes to light. The end result of this trap is that the analysts are forced to spend precious time and effort trying to analyze and clean data that aren't actually needed to inform the decisions their business wants to make.
Beginning with the end in mind, the decisions that need to be made, and building models to help better make those decisions, can dramatically increase the success rates for analytics efforts while helping set priorities for the company’s descriptive and (if needed) predictive efforts.
Another analytics trap that people often fall into is to assume that their business problem can be considered as strictly a predictive or a prescriptive analytics problem. While there are certainly many common and important problems that benefit from just prescriptive analytics, there are many more where good predictions are essential for making good decisions. While it is tempting for prescriptive analytics practitioners to use simple rules of thumb to create their predictions, there are many examples where an integrated analytics effort that combines sophisticated predictive analytics techniques like machine learning with sophisticated prescriptive analytics techniques like math programming produces much better results.
Regardless of whether building predictive models is valuable, building prescriptive math programming models that take advantage of solvers such as the Gurobi Optimizer can allow businesses to make better decisions in less time with more confidence.