Looking for something else?
At the oil and gas drilling company Halliburton, traditional BI is still important, but there is a growing emphasis on predictive analytics models. One company official said this trend is going to be the key to differentiating the Houston-based firm from its competitors and making it more successful.
"You can do as much business intelligence as you want but it's not going to help you win against your competitors in the long run," said Satyam Priyadarshy, chief data scientist at Halliburton, in a presentation at the Predictive Analytics World conference in Boston. He added that predictive modeling is going to be a "game changer."
But simply doing predictive analytics modeling isn't enough. For Priyadarshy and other conference presenters, predictive initiatives are only successful when they are business-oriented and narrowly tailored to address specific problems.
Predictive modeling is a stat-heavy, technically intensive exercise. But when implementing a predictive modeling program within a company, it's important to not get bogged down in these areas in order to push projects to deliver true business value.
Satyam Priyadarshychief data scientist, Halliburton
For Priyadarshy, this approach means breaking down some of the data silos that inevitably spring up. During the process of exploring and drilling a new gas or oil well, tremendous volumes of data are generated. But they come from several different departments. For example, data from seismic surveys of sites have traditionally not been shared with the drilling operations teams, Priyadarshy said. But there's an obvious need for the crews manning the drills to know what kind of material they're likely to hit at certain depths.
Priyadarshy said he and his team are working on a homegrown data platform that would make this data more accessible. The platform is a combination of Hadoop, SQL, and in-memory database tools. It also includes a data virtualization tool that allows different teams to access data wherever it is stored. Doing so allows drilling teams to build predictive analytics models based on data coming off of drilling sensors and from seismic surveys. These models allow the drilling teams to predict in real time how fast they should run the drill bit and how much pressure to apply.
Having such knowledge separates predictive modeling from traditional BI, Priyadarshy said. Rather than producing a static BI report that retrospectively explains certain events during the drilling process, the predictive models allow teams to make adjustments in real time and address specific problems.
"With predictive models, you want to build actionable things rather than just dashboards," Priyadarshy said.
Keep predictive modeling projects business-focused
Predictive modeling is most effective when it's used to tackle known business problems, rather than looking to predict correlations that don't necessarily have specific business value.
"You want to be clear about what types of problems you're trying to solve," said Alfred Essa, vice president of analytics at McGraw-Hill Education in Columbus, Ohio, during a presentation at Predictive Analytics World. "This helps you ask deeper questions."
McGraw-Hill works with clients -- primarily local school districts and colleges -- to look at their data to predict student performance. McGraw-Hill and the schools have been able to reliably predict how students are likely to perform in classes, including which students could fail or drop out, Essa said. But simply giving this information to schools isn't necessarily helpful. He talks to clients to make sure they have a plan for how they intend to use the information. Just telling students they're likely to fail and they need to work harder might actually backfire, causing them to give up. Schools need to develop curriculums to help failing students before they do anything with the predictions, he said.
For Essa, the answer to this kind of question often comes during exploratory data analysis. This early stage of modeling typically involves just looking at the data, graphing various elements and trying to get a feel for what's in the data. This stage can help modelers see variables that may point to trends, Essa said. In the case of predicting student failure, they may be able to see factors that lead students to fail, enabling schools to address these worries. This action goes beyond just a predictive model.
"Before you start to do modeling, it's really helpful to pose questions and interactively get answers back," Essa said.
Simplify outputs of predictive analytics models
There's always statistics underpinning any predictive model, which are useful to the modelers. But for the lines of business that interact with the results of predictive models, these stats are nothing but distraction.
Instead, predictions need to be clear and concise, said Patrick Surry, chief data scientist at airfare prediction mobile app Hopper, based in Cambridge, Mass. He talked about how one of Hopper's competitors gives customers purchasing recommendations as a confidence interval. The problem is that few people understand what the site means when it says, for example, it's 70% confident a given price is the lowest that can be expected. Similarly, when Hopper was testing its service it used the word "forecast" to talk about changes customers should expect in prices. Surry said this just made people think Hopper was talking about the weather.
"When you watch people try to interact with predictions, there are things you don't even think about," he said. "As soon as you put the word 'confidence' in there you've lost 90% of the audience."
Today, the Hopper app simply tells users to buy now because prices are as low as they're likely to get or to wait because a better deal is likely to pop up. There are some complicated predictive models running behind the scenes analyzing things like historic price data, prices for given days of the week and month, destinations and past sales. But Surry said customers don't need to know all these calculations; they just need to know if they should buy an airline ticket or wait.
Speed up the predictive analytics modeling process for better results
Get the team involved when building predictive models
Predictive analytics is much more than just a math problem