I have been dabbling in predictive analytics for six years. In the first couple of years, I focused on the output of the model. Everyone else in my organization did too. We wondered, was it accurate? Could we trust it?
It was as if we expected some major gift miracles to fall out of the sky as reward for having a model. We hadn’t formulated any ideas about what we could do after getting the model predictions. We weren’t approaching our business much differently after the model was built than we were before it existed.
This article by Michael Fitzgerald of the MIT Sloan Management Review outlines four common pitfalls that tend to accompany organizations approaching predictive analytics.
Here’s a summary of the 4 common analytics pitfalls:
- Magical thinking – it is apparently common for organizations to believe that if they invest in the time and effort to develop analytics, profit windfalls will magically appear across the organization. In truth, a model is merely a way of asking a question of your data. Nothing incredible is going to happen unless you choose to act in a different way after you see the answers.
- Starting at the top – predictive models do a great job of identifying opportunity (such as candidates for major giving) or risk (such as credit applicants who have a high chance of defaulting on a loan). Model results are best applied to inform transactional business processes rather than executive strategic decisions.
- Models that don’t scale – creating a model requires some rigor and while it is true that models are designed with a single purpose, reusable data sets and data transformation processes can go a long way in helping to apply analytics informed decision making to more than one process in the organization.
- Seeking purified data – a common barrier to embarking on analytics is an organization’s concern about the state of their existing data. There are ways to fill in missing values, transform variables or select substitute variables such that the input data is reasonably sound for the purposes of computing a model.
Click on the mousetrap if you’d like to read the original article.