TDWI | Training & Research | Business Intelligence, Analytics, Big Data, Data Warehousing

Taking Advantage of Predictive Models (Last in a Series)

Predictive analytics is most powerful when it’s incorporated into your business processes but that may be hard to enable. Here’s how to start incrementally.

• By David Loshin
• March 23, 2016

The value proposition of predictive modeling is achieved when a model is fully integrated into the business process it was intended to improve. That essentially means that a model’s predictive power is used to augment the logical flow of the process to influence desired results. This, however, may be more easily said than done for a few key reasons:

Also in This Series: Embracing Predictive Analytics How Machine-Learning Techniques Use Methods Supervised vs. Unsupervised Learning

Lack of context: Attempting to “use analytics” without truly understanding a specific business problem that needs to be solved is like looking for a needle in a haystack. The benefit of prediction is that the presumption of foreknowledge of prospective behavior can enhance the way your process influences the desired result. For example, a predictive model for credit approval addresses the business problem of avoiding giving credit to a person likely to default on the payments.

Lack of plan: There is a need to devise a “plan of attack” to determine what kinds of techniques will be used, which specific machine-learning methods will be used, what order they will be tried, what type of models will be created, along with a plan for testing, integration, and implementation.

Integration touch points: When the business process is opaque, it may be difficult for the analyst team to figure out how to incorporate the predictive model into the application. This means you need to investigate where the integration touch points are. To use the credit approval example, obviously the model must be invoked before the decision is made to approve the credit application, meaning that the application must be modified to embed the invocation of the model at the appropriate point.

Need for testing: Because the ways predictive models are created involve heuristics that may embed latent biases, it is possible that the models devised using the supervised approach may be great for the training data set but not useful for a more general data set. That means that a successful predictive modeling program must have well-defined means for testing the models on a variety of data sets, along with clearly defined measures to assess how well the model performs. Not understanding how to test the model’s validity adds a measure of risk of failure.

A predictive analytics program will benefit from incremental design, development, and implementation, with a particular focus on identifying the business problem to be solved. Here are four practical steps to take before you purchase and install an analytics package:

• Review the business landscape to determine where there are opportunities to exploit a predictive capability. Examples include customer-facing applications that drive sales, marketing activities looking to fill a sales pipeline, examining manufacturing processes to reduce unexpected downtime, assessing field part performance for predictive maintenance, and evaluating regional sales patterns for predictive warehousing.


• Determine what data is available for analysis and what data is not yet available for analysis. Identify sources of the data not yet available and consider methods of acquiring that data.


• Evaluate the platform requirements for ingesting and managing the data to be used for analysis and ensure that your systems can support these requirements.


• Train your team in machine-learning methods and techniques and how they are applied to address business challenges.

Finally, consider engaging expert help in launching the program. Benefitting from others’ experiences may short-circuit potential land mines along the way and it may shorten time-to-value for your predictive analytics program.