What's the difference between data mining, forecasting and optimization? When should you use each technique, and how do they interact? Jeremy TerBush, Vice President of Global Analytics, Wyndham Exchange and Rentals explains that he uses all techniques together in an overall predictive process.
"Data mining is first step in the system," says Terbush. At Wyndham he uses data mining to look for patterns and break data down to find nuggets of information that are really important. Then, he uses the results of his data mining within his forecasting system, feeding important variables to help predict demand for rental properties. "Next, his demand model feeds into the optimization process, which provides a specific decision variable, giving me an optimal decision I can make today about the business."
When Rey talks about the difference between techniques, he starts by discussing time. "In the case of prediction, you may or may not have time involved," says Rey. You may be predicting the best location, for example, or the next likely customer action, neither of which involve time. If you want to make predictions over a 12-month, 24-month or 36-month period time, however, that is a forecasting problem. "Forecasting is a special case of prediction," says Rey.
Rey, who was also a keynote speaker earlier this week at Analytics 2012, applies data mining techniques to time series data from the Web and from third party data providers to feed into his forecasting process.
"The interdisciplinary notion of data mining and forecasting is important because there is a plethora of time series data available to plug into forecasting," says Rey. "You can buy 30 million variables from companies like Experion but you have to use data mining to sift through that data to find out what the important variables are."
Rey's team has a backbone of 7,000 models that provides forecasting that feeds the strategy process from executive sales and planning to marketing and financials, as well as specialty forecasting for freight logistics and marine costs.
The analytic process at Dow Chemical includes four steps, says Rey:
- Define the problem.
- Gather data.
- Build, test and refine the models.
- Deploy the models.
"Fundamentally, it’s the scientific method," says Rey. "First you have to understand the problem, then you develop the hypotheses, which drives how you’re going to build the model. It's really about understanding the problem, knowing business, knowing what data is available, and then the hard work is finding, extracting, cleaning and harmonizing the data. Building the model is fun stuff, and then you deploy."