How Predictive Analytics Work

We've been busy at Inuvo building all sorts of new models. I'll talk about those in the next few posts. I also had a new baby girl, so its been busy around here!

So how do predictive analytics work? The basic idea is to build a statistical model that looks at a variety of input data and produces a numerical score as the output. The input data could be the age of the person, their gender, their income and their occupation. The output could be the probability that they buy a Toyota Prius. The input data could be the time of a click on an advertisement, the number of previous clicks from the IP address and the click through rate on that ad. The output could be the probability that the click is fraudulent.

Most simple predictive models are a mathematical equation that combines the various input data to compute the output value. Numerical data, such as age, might be input simply as a number into the model. Categorical data such as occupation might be grouped into groups, and then a numerical weight assigned to each group. There are dozens of techniques for managing different types of input data.

But how do these equations get built? The simplest way to build predictive models is to take a large set of past data where we know the outcome. At Inuvo, we recently built a very powerful fraud model by examining many months of previous transaction data. Because time had passed on that data, we knew exactly which transactions turned out good, and which turned out bad. We used this good/bad data to build the equation inside the model, and now it can monitor a publisher's transaction traffic, and predict which transactions are fraudulent.

How accurate is it? Well, like any statistical system, it only prescribes a probability of a transaction being fraudulent. But every day we review the output of the fraud model, checking through the most suspicious activity. We've already shutdown a handful of bad publishers, including some that were caught after just four transactions! That's pretty powerful.

Building predictive models has all sorts of pitfalls and techniques, smart rules of thumb and clever approaches to use under different circumstances. Over the next few posts, I'll talk about some of the ones we use at Inuvo, why and what the benefit is to our customers.