Your Source for Data Science

The Current State of Predictive Analytics and Machine Learning | TDWI Report

To understand the current state of predictive analytics, we asked respondents if they are using predictive analytics or planning to do so. For those who are not using the technologies and have no plans to do so we asked, “Why not?”


In this report, the group currently using the technology is referred to as the active group. Those who are planning to use the technology are referred to as the exploring group. In this survey, half (50%) already use the technology; more than half of this active group has been using it for more than three years. Forty percent are exploring the technology now. Ten percent are not using the technology. This should not be viewed as an adoption rate because respondents tend to gravitate to surveys they can relate to.

The majority of active group enterprises have models in production. Models provide the most value when used in production to make decisions and take action. The majority of active group respondents (73%, not shown) already have models in production. Another 17% plan to have models in production in the next few months. About 10% either don’t have models in production or don’t know when they might. The fact that the vast majority have put models into production is good news and illustrates that this group of respondents realizes the value of taking action on analytics. In fact, operationalizing models (to make them part of a business process) is one of the top areas of interest for moving programs forward. It is also important for making predictive analytics more pervasive.

Predictive analytics is used across a range of use cases. We asked both the active group and the exploring group what kinds of use cases they are using or plan to use in predictive analytics. As illustrated in Figure 1, there are many popular use cases for predictive analytics for both groups.

What is Predictive Analytics? Predictive analytics consists of statistical and machine learning algorithms used to determine the probability of future outcomes using historical data. Some people think of machine learning as being completely different from predictive analytics. Machine learning techniques, however, are often used in predictive analytics; they just use a different approach.

• Marketing applications often lead the way. Over half (52%) of the active group is using predictive analytics for retention analysis or direct marketing. Cross-sell and up-sell is also popular in marketing. These are popular use cases for the exploring group as well with a third of respondents planning to deploy these use cases in the short term. Previous TDWI research indicates that marketing is often one of the first departments to adopt more advanced analytics.

• Default prediction is also popular. In addition to marketing use cases, respondents are also utilizing or interested in other kinds of use cases. For example, default prediction ranked high for both groups with 46% of the active group already doing this kind of analysis and 34% of the exploring group planning to do so in the short term. Default analysis is important for a number of use cases including loans, credit cards, premium payments, and tuition payments.

• Newer use cases such as predictive maintenance are gaining steam. Thirty-four percent of the active group is already using predictive maintenance and 22% of the exploratory group plans to use it. In predictive maintenance, organizations calculate the probability of an operational asset requiring servicing or even failing. Some organizations make use of sensor data from the Internet of Things (IoT). For instance, a fleet operator might use sensors to collect data from their various trucks. Such data might include the temperature or number of vibrations per second of a particular part or parts. This data can be analyzed using machine learning to determine what precipitates a part failure or when undue wear and tear is occurring. The system “learns” the patterns that constitute the need for repair. That information might be encoded into a set of rules or a model and used to score new data from trucks in order to improve fleet maintenance and operational efficiency. Other use cases include cybersecurity, where 25% of the active group is using predictive analytics (not shown).

What is Machine Learning? Machine learning methods originated in the field of computational science a few decades ago. In machine learning, systems learn from data to identify patterns with minimal human intervention. The computer learns from examples, typically using either supervised or unsupervised approaches. In supervised learning, the system is given a target (also known as an output or label) of interest. The system is trained on these outcomes using various attributes (also called features). In unsupervised learning, there are no outcomes specified. 

• In applications. An up-and-coming area of interest is embedding predictive analytics and machine learning models in applications that require intelligence. We did not ask specifically about these applications but they are worth mentioning here. These include voice recognition, traffic apps, and chatbots. Although only 12% (not shown) cite building apps as a driver for predictive analytics and machine learning, we expect that percentage
to grow.

Source Download Site