As a Data and Analytics company, we at CompassRed have seen the full cycle of leveraging data for insight with each phase of development being just as important as any other. Deployment is the part of the predictive analytics process that puts the predictions to work.  Predictive analytics requires a very time-consuming process of data preparation and a very complex process of finding the best algorithms (also known as “modeling”) for producing predictions.  However, all of this effort is for naught if the predictions are not used by an organization to meet their objectives, which typically means increasing revenue and decreasing costs.

Considering the End User

The people affected by deployment are usually not the same people who care about data structure or complicated algorithms.  They are the folks who will use predictions to work more efficiently and effectively.  It’s incredibly important to take into consideration the audience that uses the data to make decisions. Here are some examples of effective deployment:

Journal

Predictive Analytics Output

Example Uses and Users

Sales leads sorted in descending order by their likelihood to become a sale

Sales people increase their number of sales with fewer calls

Forecasts of demand

Production planners order supplies that are closer to demand, thereby reducing waste as a result of overproduction, and reducing opportunity costs as a result of underproduction

Specific conditions in which factory equipment is likely to fail

Factory staff reduce the number of equipment failures by identifying conditions in which failure becomes likely

The likelihood of a loan applicant to repay based on their profile 

Loan officers increase repayment rates by identifying applicants who are the most likely to repay 

Estimated risk of hospital readmission

Hospital staff identify patients who are the most likely to be readmitted, and thereby take actions that minimize the risk of readmission

Predicted amount of time an applicant will remain employed with a company

Human resources staff hire applicants who are the most likely to remain with the company more than a year

Time and place of neighborhood with a higher than average probability of a crime (“hot spots”)

Police officers schedule their patrols to ensure they are in hot spots when the risk of crime is highest

Predicted outcome of sales events

Marketers can identify the most important drivers of a successful ongoing sales events, and anticipate revenue generated by different types of sales events


Scoring

The deployment process requires new cases to be “scored”.  Each model represents an algorithm, or equation, that captures the relationships between a set of predictors and a target field found in an organization’s historical data.  Each predictor is part of the unique profile of a case, such as the characteristics of a sale, a piece of equipment, a loan applicant, a hospital patient, an employee, a neighborhood, or a customer.  The algorithm can produce scores for new cases based on those characteristics.  The scores may be the predicted probability of an event occurring (known as “propensity scores”) or the predicted level of an outcome, such as revenue or time.

Delivering the Predictions

Effective deployment requires the presentation of the predictive analysis results in a format that is easy for staff to use.  This often involves business intelligence software that enables staff to access the specific information that is relevant to them.  For instance, sales people will need to access only the leads assigned to them.  A dashboard may also provide a graphical interface that facilitates access to the predictions, and can also be used to graphically explain the rationale for the predictions, such as time plot that show how the forecasts are consistent with historic trends.

Potential Resistance

People generally prefer to act on intuition rather than the results of data analysis.  In his bestseller book entitled “Thinking Fast, Thinking Slow”, Daniel Kahnamen cited research that confirmed this “irrational” bias.  For instance, if a behavior produces the desired results, most people will continue this behavior even if an analysis reveals that it only works once out of ten times, and ignore another new behavior that has been found to produce the same results more than once every ten times.  Successful deployment will require the full participation of the persons who will be acting on predictions, with a thorough explanation of the predictive analytics process provided to them.  

Commitment to Data Quality

A successful predictive analytics process requires that the historical data used for the analysis is consistently and accurately updated.  This typically means that the staff who are using the prediction enter the results of their work based on the predictions.  For instance, sales staff should consistently record whether a sales call was successful or not.  Without a process of predictive analytics and deployment, this recording task may be seen as burdensome and a low priority.  However, once staff begin seeing a connection between the quality of their data entry and the contribution of deployment to their success, their commitment to the quality of the data they enter can be expected to increase.  Ideally, their careful entry of data will contribute to a virtuous cycle of improved data quality and more effective deployment.

At CompassRed, we believe that Predictive Analytics is a process that must deliver business value.  Based on our experience with hundreds of organizations that were interested in acquiring predictive intelligence, we have observed how deployment is an integral part of delivering business value, and we have noticed that this step in the Predictive Analytics process is often overlooked. This experience has let us to uncover the most effective ways to deliver predictive intelligence to the right audience in the right medium.  Contact us (info@compassRed.com) for more information on integrating deployment into your analysis.

Posted
AuthorPatrick Callahan