CIO — It can create new business. The Navy Federal Credit Union has applied predictive analytics technology from SPSS to the design of new products.Analyzing how ATM withdrawals spiked just before and after a deployment led to the introduction of a checking account with ATM fee rebates for members on active duty. CIO Jerry Hermes says that other business units have since invited the analytics unit into their planning process.

You can't really know the future. Predictive analytics forecasts about your business are useful only as long as you understand that they describe probabilities. "The weatherman gets it wrong some times, even though we've spent hundreds of years collecting data and looking at correlations," says Royce Bell, CEO of Accenture Information Management Services.

Results can mislead. You need to apply business acumen to make sure you draw the right conclusions, Hermes says. Alan Payne, who manages an R&D group at the Navy credit union, remembers when the model seemed to show that more members were deployed than they expected. It turned out that the survey used for the analysis needed to better distinguish between households and individuals; the spouses of deployed members didn't know which box to check.

Watch your gut. People tend to be quickest to accept predictions that match their expectations. These predictions can be valuable when they provide insight into the variables that drive them, Bell says. But lately, C-level executives get most excited "by the nonintuitive ah-ha," Bell adds. Results that prove the limits of intuition are a "tough but valuable sell," because employees often resist conclusions that go against their experience and instincts.

Garbage in, garbage out. "A good number of analytic programs fail on questions about the veracity of data," Bell says, so getting serious about data quality is one of the prerequisites for success. That may mean you have to be selective about the data you feed into your model, he adds. Less is more when you focus on the most accurate information and leave out questionable numbers.