Skip to main content
Different approaches to adding insight to your decisions

I got an interesting comment the other day asking how to effectively conduct data mining in an EDM context - outsource or not? This made me think it would be worth posting a general description of the different ways to add analytic insight to a decisioning environment. Assuming we are talking about adding predictive analytic models to operational system (see this article for a discussion on predictive analytics), there are several ways to bring analytic insight to bear on a decision.

  1. Find a score
    Clearly if there is an externally available score with a good reputation and relevance then use that. This is by far the cheapest and most effective - you don't need data or data mining/analytic expertise, just an understanding of your business and an ability to match it with a score. If, for instance, you are trying to manage credit risk then the FICO score or its global counterpart Global FICO is a great tool. These are typically available from bureaus (in the US) and sometimes have different names (BEACON and EMPIRICA are both FICO scores, for instance).
  2. Find a score that seems to be a good substitute
    If, as in most decisioning problems, you cannot find a score that works exactly consider if any of the scores that are available might be a good surrogate. Larry Rosenberger (Fair Isaac's head of R&D) gave a great keynote on this at InterACT recently where he talked about the fact that many scores can predict unexpected things - for instance, your FICO score turns out to be a reasonably good predictor of whether you will stay fit or not. No-one is suggesting that there is a causal link here - having a good FICO Score does not MAKE you fit - but that some underlying characteristics (responsibility or integrity perhaps) mean you tend to act in ways both that drive up your FICO score and that increase your likelihood of fitness. You can thus analyze your customers by comparing behavior to available scores to see if any are a good (non-causal) predictor. A word of warning - don't ever rely on a single score, especially a single score for which you don't have causal linkage, to make a decision on its own. This will not work well and will get you in trouble with regulators (and quite rightly too).
  3. Work with a third party DSP
    There is growing interest in something called a Decision Service Provider or DSP. This involves basically outsourcing the decision-making piece of your process. You make the issue of finding the right data, working out where and how to get it and then decisioning with it someone else's problem. The DSP hosts the common and regulatory rules (helping you be compliant) and allows you to customize for your business while handling the complexities of assembling a rich data set and analytics that use that data set. By and large this works best when there is lots of third party data that might make sense for you to include, e.g. about consumers, but where you don't have relationships with all the data providers.
  4. Have someone analyze your data and send you models (not reports)
    This is the classic way to get started. Work with someone like Fair Isaac who has the analytic/data mining skills to turn data into insight. You make your data available, though hopefully they have expertise too that will be brought to bear, and they help you by cleaning and arranging the data to make it usable to build predictions and then by building a predictive model from it. You want someone to hand over an actual model, or an analytically derived set of business rules, rather than a report that just tells you what you should do as it is important to keep the time delay between getting data and using that data to make better decisions to a minimum. Ideally you will use a package or build into your EDM framework a place to put the model so as to enhance your operational decisioing. You will likely also want to have the model refreshed on a timely basis (what is timely varies depending on the model).
  5. Have someone analyze the data AND teach you in the process
    You can take the approach in 4 and adjust it by having it done on-site with tools you can buy yourself. If you combine this with training and having your staff act as second-seat behind a qualified modeler you can build your own team and skills in parallel.This works great if you have a desire to staff up and do this yourself long term but want to get short term results.
  6. Buy tools and training and do it yourself
    There are good tools out there for doing the analysis and building models from it and even a few, like Fair Isaac's Model Builder, that will deploy the model for you into production systems. You can hire staff, train them and develop a unique brand of analytics for your decisions based on your customers.

related posts