Methodology

This report is based on primary research consisting of interviews with end users and briefings with vendors as well as secondary research consisting of articles, reports and websites that I’ve perused. However, the bulk of the report consists of a survey of 3,474 business and information technology (IT) professionals conducted in March and April, 2014. Not all respondents completed all questions, which is why the number of responses for some questions is lower than 3,474. The results in this report are based on 1,857 respondents who completed the entire survey and 1,617 who partially completed it.

Among the respondents, 61% are IT professionals and the remainder (39%) are business professionals. Almost one-third (28%) work in large companies with more than $1B in revenue, while one-quarter (25%) work in medium-size companies with between $100M and $1B in revenues. The remainder, 47%, work in small companies with revenues under $100M. Almost one-third (29%) of respondents are from the United States, followed by 22% from India, 13% from the United Kingdom, 7% from Australia and the rest from other countries.

Introduction

Definition

Eckerson Group defines predictive analytics as the use of statistical or machine-learning models to discovers patterns and relationships in data that can help business people predict future behavior or activity. Also known as advanced analytics, data mining or knowledge discovery, predictive analytics involves:

• Creating models using data that represents past business activity.
• Applying these models to new business data to make predictions and create rules.
• Embedding models into applications to optimize business processes, improve business decisions and automate responses.

Statistics, with its ability to define the shape and nature of data, is the heart of predictive analytics. But machine learning is now a coequal partner in creating predictive models. That’s because machine learning leverages compute power to run complex algorithms against data that manual statistical techniques could never accomplish. Machine-learning models can build more accurate models against more data to help address a wider variety of applications.

To be honest, predictive analytics is a misnomer. Not all analytics defined above are predictive. Most of the time, the output of an analytical model is simply descriptive: it describes a pattern or relationship that can help business people better understand what’s happening and what to do about it. It’s not until the model (or mathematical equation) is applied to new records through a process called scoring that a model becomes predictive. The scores, usually a decimal between 0 and 1, define the propensity of the record to exhibit a specific behavior.

For example, a model may score a customer’s likelihood to respond to a particular marketing offer. A marketer might use that score to determine whether to send the customer a direct mail offer. Or a bank teller might use the score to determine whether to offer a premium rate for a certificate of deposit to a customer who just made a deposit. An e-commerce application might look up the score of a customer who just purchased a product, feed it into a dynamic rules engine and display a cross-sell offer to the customer in real time.