by John Edwards

7 secrets to predictive analytics success

Sep 11, 2018
AnalyticsData Science

Forecast the future with accuracy and confidence. Here are the tools and tactics you need to know to translate data into foresight.

01 intro prediction
Credit: Thinkstock

Predicting the future has always been a dicey proposition. Fortunately, the arrival of predictive analytics technology, which allows users to project future outcomes based on historical data and analytics techniques, such as statistical modeling and machine learning, has made forecasting results and trends far more reliable than in past years.

Still, as with any emerging technology, predictive analytics can be difficult to use to its full potential. Compounding the challenge is the fact that inaccurate or misleading results caused by poorly developed strategies or the misuse of predictive analytics tools may not become apparent for weeks, months or even years.

Predictive analytics has the potential to revolutionize a wide range of industries and operations, including retail, manufacturing, supply chains, network management, financial services and healthcare. “Deep learning and predictive AI analytics are going to transform all segments of our society on par with this decade’s transformation of the internet and cellular technology,” predicts Bob Friday, CTO and co-founder of Mist Systems, an AI network technology company.

Here are seven tips designed to help your organization get the most out of its predictive analytics initiative.

1. Have access to high-quality, well-understood data

Predictive analytics applications are data-hungry, relying on information fed through a feedback loop to continuously improve. “Data and predictive analytics feed from each other,” observes Soumendra Mohanty, chief data and analytics officer at L&T Infotech, a global IT solutions and services provider.

It’s important to understand the type or types of data that’s flowing into a predictive analytics model. “What kind of data does one have?” asks Eric Feigl-Ding, an epidemiologist, nutritionist and health economist who’s currently a visiting scientist at the Harvard Chan School of Public Health. “Is it live data that’s collected daily, like on Facebook and Google, or difficult to access healthcare data that’s needed for medical records?” To make accurate predictions, the model needs to be designed to work with the specific types of data it’s ingesting.

Predictive modeling efforts that simply throw gobs of data at computing resources are generally doomed to fail. “Since there’s an abundance of data, most of it may not be relevant to a given problem but may appear to be relevant in a given sample,” explains Henri Waelbroeck, vice president and director of research for portfolio management and trading solutions for FactSet, a financial data and software company. “Without an understanding of the process that originates the data, a model trained on biased data could be completely wrong.”

2. Pay attention to patterns

Everyone obsesses about algorithms, but algorithms are only as good as the data that’s fed into them, observes Richard Mooney, lead advanced analytics product manager at SAP. “If there’s no pattern to find, then they won’t find one,” he notes. “Most datasets have hidden patterns.”

Patterns are generally hidden in two ways:

  • The pattern is found in the relationships between two columns. A pattern, for instance, could be discovered by comparing an impending deal’s closing date information with associated email open rate data. “The email open rate should increase massively if a deal is going to close, because there will be a lot of people on the buyer side reading the contracts and reviewing them,” Mooney says.
  • The pattern is revealed in how a variable changes over time. “In the example above, knowing a customer opened an email 200 times is not half as useful as knowing they opened in 175 times in the last week,” Mooney

3. Focus on manageable tasks that are likely to deliver positive ROI

“There’s a temptation these days to just apply machine learning algorithms to mountains of data in the hopes of gaining insights,” claims Michael Urmeneta, director of analytics and business intelligence for the New York Institute of Technology (NYIT). The problem with this approach, he says, is that it’s like trying to cure all forms of cancer all at once. “The problem is too big, the data is too messy — there‘s not enough funding, there’s not enough support,” Urmeneta explains. “It’s impossible to get a win.”

There’s a much greater probability of success when a task is focused. “It’s likely we will have access to subject matter experts who understand the intricacies if there are questions,” Urmeneta notes. “It’s likely we will have cleaner, or better understood, data to work with.”

4. Use the right approach for the job

The good news is that there are an almost unlimited number of methods and approaches that can be used to generate accurate predictive analytics. Yet that also happens to be the bad news. “There’ s a new, hot analytic approach every day, and it’s easy to get excited about using a new methodology,” states Angela Fontes, director of the behavioral, economic analysis and decision-making practice for NORC (formerly known as the National Opinion Research Center) at the University of Chicago. “However, in my experience, the most successful projects are those that really think deeply about the desired outcome of the analytics and let that guide their choice of methodology — even when the most appropriate method is not the sexiest, newest approach.”

“Users must be cautious in choosing the right approach for their needs,” advises Shanchieh Jay Yang, an associate professor and head of the computer engineering department at the Rochester Institute of Technology. “[Have] an efficient and explainable technique that leverages the statistical properties of sequential, temporal data and extrapolates it into the most likely future,” Yang says.

5. Build models with precisely-defined goals

It seems obvious, but many predictive analytics projects start with the goal of building a magnificent model without a clear plan for how it will eventually be used. “There are tons of great … models that never went anywhere because nobody knew how to use the information to achieve or deliver value,” observes Jason Verlen, senior vice president of product management for CCC Information Services, a SaaS provider to the automotive, insurance, and collision repair industries.

Fontes agrees. “Using the right tool certainly makes sure we’re getting the desired result from the analytics … because it forces us to be really clear about our goals,” she explains. “If we’re not clear about the goals of the analysis, we can throw the kitchen sink at a problem and never really get what we are looking for.”

6. Form a close partnership between IT and relevant business units

It’s essential to establish a solid partnership between business and technical organizations. “You should be able to understand how the new technology addresses a business challenge or improves the existing business landscape,” says Paul Lasserre, vice president of product management for artificial intelligence at Genesys, a customer experience technology provider. Then, once a target has been set, test the model in a limited-scope application to determine if the solution will actually provide value.

7. Don’t get misled by a poorly designed model

Models are designed by people, so they often contain lurking imperfections. A faulty model, or one built using incorrect or poorly selected data, is prone to delivering misleading or, in extreme cases, totally wrong predictions.

Selection bias, for instance, in which proper randomization is not achieved, can muddle predictions. In a hypothetical weight loss study, for example, perhaps 50 percent of participants opt to drop out of follow-up weight measurements. Yet the individuals who dropped out had different weight trajectories than the subjects who stayed in. This complicates the analysis, since people who stay with the program in such a study are usually those who actually lose weight. Quitters, on the other hand, are typically individuals who experienced little or no weight reduction. Therefore, while weight loss could be causal and predictive across a full world, in a limited database with a 50 percent dropout, actual total results could be hidden, Feigl-Ding reports.

The takeaway

“Enterprises are going through growing pains and learning that predictive analytics isn’t something you can dabble in,” says Arvin Hsu, senior director of data science for GoodData, a business intelligence and analytics software developer. “However, the impact that strong predictive analytics can have on business efficiency, revenue and product performance is well worth the time, energy, and resources that are needed to ensure success.”