The different levels of analytics can be summarized into three categories. These are descriptive, predictive and prescriptive analytics.
The first category, descriptive analytics, is necessary to be able to perform analysis. The tools and processes that belong in this category are not necessarily simple, but are sometimes only simple statistics describing data. Traditional business intelligence (BI) falls under this category. The products of data exploration during data science does as well. Interestingly enough, even “alarming” falls under this category. Many times, alarms are considered advanced tools. They are certainly valuable, but really only tell a business what has already happened based on hard data and rules. The reason analytics in this category are necessary for being able to perform analytics in either of the other two categories is that achieving descriptive analytics prepares the organization’s data for consumption. Even if analytics implemented in systems are not part of an organizations regular business, personnel take in the descriptive analytic products and perform additional processing themselves to determine decisions. Another area that falls under descriptive analytics is data exploration. Data exploration is a manual, iterative process that generates statistics, which lead to insight on how to build models useful for implementation in a process. This is the data scientist’s first step.
Predictive analytics are the more common of proactive analytics. This category is vast. In statistics, prediction is also known as estimation and not necessarily forecasting. Predictive analytics are concerned with unknown quantities and factors. Factors can be flags, levels, classification or clustering. Predictive analytics are performed with any of a multitude of tools. A simple example is ordinary least squares (OLS) estimation. The machine learning wave mostly lives in this category. Hierarchical & k-means clustering, random forest regression & classification, support vector machines and ridge regression are some of the tools that exist in this category. These tools each have their strengths and weaknesses and it takes a learned person (or automated evaluation of each one) to figure out which is the best to apply to a specific problem. These problems include finding production excursions, detecting fraud, estimating power output, demand forecasting, risk estimation and claims prediction. This list is barely a scratch on the surface of the use cases for predictive analytics.
The final category, prescriptive analytics, is the other proactive category of analytics shared by two general classes of tools. Machine learning is one of these. Operations research is the other. Each of these classes makes extensive use of mathematical optimization techniques. Artificial neural networks and deep learning methods are machine learning techniques which are capable of recommending decisions in certain applications. The problem domain for deep learning is continually expanding. This technology is what makes autonomous vehicles possible. There are other techniques that can be applied to recommend decisions through mathematical optimization, such as stochastic gradient descent and other unconstrained minimization techniques. When the rules of the decisions are known ahead of time, a machine learning method can be applied to recommend a decision. Many times, decisions must be made within more general constraints instead of rules. For instance, a supplier might need to deliver product from different plants to a variety of depots. This scenario has demands for product at each depot and maximum supply at each plant. The transportation costs between plants and depots can be taken into account and a minimum-cost solution found using linear programming. This use case is one that has had an efficient mathematical solution since World War II. The applications of linear programming have been explored and expanded throughout many industries. It can be applied to manufacturing (blending, for instance) and finance (using portfolio balancing) for two examples.
These categories of analytics can be regarded as a progression, but they need not be, with the exception of having a handle on descriptive analytics before beginning any proactive analytics. The two proactive categories can overlap some as far as techniques, but one distinguishing characteristic is the level of additional insight that must be obtained to make a decision. For instance, if a rule is applied to results of a machine learning method, then the method is not prescriptive. Whereas, if the results of a technique are decision points by themselves, then it is prescriptive.