While most enterprises claim a degree of science behind corporate decisions, the truth remains that in what are still essentially product-centric silos, most production is decided on the basis of the intuition of the management team. When that fails and business results do not match expectation, the result is usually to sack the management and find others who the shareholders believe will guess better.
The better-organized enterprises will employ historic reporting to give them more detailed pictures of what has happened in the past, and that factual data will be used to frame the decisions. This tends to be fine until business change becomes more rapid than the ability of historic analysis to reflect, because the vast number of variables interacting and changing overwhelm the ability of mangers to analyze and synthesize in a meaningful fashion. At this stage, it becomes imperative to manage by predictive modeling, where analysis of the data enables the variables to shape the future with far more certainty than an educated guess.
A predictive model is a mathematical description to predict the future, in terms of a behavior or outcome, based on an existing data pattern. The model is built using known outcomes to train the model to identify the conditions that typically contribute to the result. Using mathematical models rather than intuition means that the explanation of the model is far more transparent. In a fast-moving world, it then becomes far more obvious what is changing and how that change is impacting outcomes. The models will not only identify what is important to an outcome, but will also assign value to their impact on the outcome. Therefore, instead of having hundreds of individual drivers, attention can be focused on those with the largest impact.
The New CRM Paradigm
It is now increasingly recognized that the aim of customer relationship management (CRM) must be to understand a customer's impact on business performance - to determine how best to communicate to customers in order to cost efficiently and effectively enhance the level of their current performance, be that by staying longer, buying more or buying things with a higher net margin. All of this is undertaken within a framework of managing customer lifetime value, so parochial end-of-quarter sales decisions should not endanger the overall value of an account. The analysis of all of the facts within the CRM systems captured and fed to the data warehouse needs to be readily converted into knowledge and returned to the CRM systems to influence and coordinate behavior.
In the new paradigm, advanced analytics become essential. However, to date the means of undertaking that analysis has often been at odds with the business reality. There are few organizations with the required levels of trained statisticians and data analysts to support the business. Few of those trained as statisticians or data analysts actually understand the business in sufficient detail to make best use of their skills, and the business domain experts are too uncomfortable with the arcane and opaque nature of analysis to be able to employ the techniques themselves or readily understand some of the more obscure outputs when presented by the analytics professionals.
While the current popular predictive analytics platforms have made strides in presenting information through mediums such as Internet portals, the techniques themselves still create many problems. Traditional methods are still largely reliant on an installed base of trained and skilled advocates who, while important in advancing the case for analytics, also act as barriers to any fundamental changes to their elite status as the masters of the arcane dark arts of modeling and analysis.
New tools are now available that are designed for use by both the trained and the untrained statistical user. These tools automate the process converting data into knowledge and provide results that are essentially far more intuitive for the non-statistician to understand. For the trained modeler, it is a productivity aid, whether used in its own right or in support of other traditional techniques (such as neural networks) where it can be used to identify the variables to be used in a model quickly and reliably rather than needing to use trial and error.
The Demand for Change
Traditional statistical techniques were developed to enable small samples to be analyzed. From the findings within the sample, it was then possible to generalize about a larger population. The key to the process of generalization was to make certain assumptions about the shape and the distribution of the data - and if the data was skewed away from the desired shape, to adjust it mathematically so that technique would work correctly and the generalization would be robust.
However, in business we often do not want to deal with a small sample; we actually want to analyze the full population. Much skill is then required to manipulate the data to ensure that it is distributed correctly for the technique to be applied soundly, and these are skills that are clearly not common among business experts.
The technique that is employed when building a predictive model is to split the data in two. One set is the training set, and the other is the validation set. We know the outcomes for both sets. The training set data is used to train the model to identify the relevant outcome. The validation set is then used to check that the model is able to identify the correct outcomes in another set of data. This ensures that the model is not so overly adapted to the training set that it cannot handle other data. This is a process of iteration with the process passing through loops of modify, model, test, modify, model and test to avoid what is termed an over-fitted model.
Be the first to comment on this post using the section below.