descriptive analytics predictive modeling

6 challenges of building predictive analytics models

The use of predictive analytics in marketing can bring benefits companywide. But building a good predictive analytics model is not trivial. Here are six challenges.

The availability of low-cost business analytics tools has motivated many organizations to roll out customer profiling and predictive analytics applications in the hopes of revitalizing their marketing and sales initiatives. However, while there is no doubt that the right predictive analytics models can add significant value to customer outreach efforts when properly designed and deployed, there are some situations in which they might not meet user expectations.

First, let's consider the typical analytics framework, which includes customer profiles and collections of historical transactions. Organizations collect customer data from among their different internal systems and consolidate that historical data into a unified customer database. Those customer records may be augmented with demographic and psychographic data from multiple sources -- such as third-party data providers or online resources -- to create customer profiles.

The second part of the analytics framework requires accumulating different sets of customer transactions. These will include, but are not limited to, records from marketing, sales, finance, credit, fulfillment and distribution, customer support and legal.

The goal is to collect all the interactions with customers so they can be subject to data analysis that looks for sentinel patterns that precede desired outcomes. A good example is identifying sequences of actions that a customer performs that result in a product sale.

Advanced analytics technology has become sophisticated enough to analyze transactions from the different business functions to find complex combinations of event sequences that are presumed to predict desired outcomes.

However, it's useful to maintain some healthy skepticism about the precision and accuracy of predictive analytics models, as there may be situations in which they inaccurately predict situations or attempt to influence behaviors that aren't characteristic of the individuals involved. The result is reliance on models that may have limited predictive power, or even pose ethical challenges for businesses.

Six predictive analytics challenges

Here are six challenges of using predictive analytics models to consider.

1. Incompleteness

The accuracy of predictive analytics models is limited by the completeness and accuracy of the data being used. Because the analytical algorithms attempt to build models based on the available data, deficiencies in the data may lead to deficiencies in the model.

Correspondingly, the developed model might not encompass enough information to be able to recognize enough sentinel predictive patterns to be of any value. For example, a customer retention model might be built using customer service event histories and transactions, but the most accurate prediction models might require sales and returns transactions to provide patterns that yield value.

2. Data myopia

Customer profiles are engineered using guidelines based on people's expectations, which come with certain biases. Limitations in the range of different demographic variables in the model may force customers to be classified in ways that are too limited. As an example, individuals might be classified using salary averages calculated within the boundaries of defined census tracts. However, certain urban areas may have census tract regions in which there are multiple discrete micro-communities with significantly different salary demographics. In this case, refining the size of the area of focus for average salary will improve the precision of the customer classification model.

3. Narrow-ization

This refers to what happens when the reliance on predictive analytics models to shape the business processes that influence customer behavior creates artificial boundaries that narrow the range of a customer's anticipated behaviors. In this case, there may be business opportunities -- such as product bundling or up-selling -- that are not even considered because the analytics-driven business process does not expect those opportunities to arise.

4. Spookiness

For a long time, automated systems have been capable of simple ad tracking in which sites drop cookies that provide information that can be accessed by partners within an ad network. Systems are becoming increasingly capable of scanning customer actions within a hierarchical semantic context to provide increased information about customer interests. A person's search terms coupled with product page visits may provide enough information to make inferences about what the customer is really looking for. However, as these bits of information are being employed to present advertisements and product placements, customers are becoming unnerved by automated systems attempting to anticipate their intent and influence their activities.

Some of the aforementioned technical challenges related to building the right predictive model for the best results can be resolved by addressing the following people challenges:

5. Skills

Predictive analytics is a team sport, and the number of players required for success is expanding. Rapidly changing market conditions and customer expectations means that having people with domain knowledge on the team is more critical than ever. Meanwhile, fast-evolving tools that incorporate machine learning and other artificial intelligence technologies have expanded the repertoire of technical skills and expertise required to build predictive models. In addition to the statisticians that work on model accuracy, successful predictive analytics projects increasingly involve data scientists and data engineers schooled in model selection and evaluation. In time, advanced analytics will become more democratized, experts predict, as analytics tools become embedded in applications and "one-click" functions akin to the self-service BI offerings that made business intelligence more accessible mature.

6. Adoption

Even as analytics platforms become more business user friendly, the perennial challenge of getting people to adopt these tools remains. According to Forrester Research analyst and longtime analytics expert Boris Evelson, "not more than 20% of all decision-makers who could be using -- and should be using -- these tools are using them today." Distrust of the metrics used to build the model, and an abiding attachment to Excel are just two reasons for low adoption rates. "It's a process." He is hopeful that the emerging user-friendly tools for advanced analytics, which include virtual and augmented reality and AI technologies such as natural language generation, will break that barrier.

Last thoughts

On the technical side, organizations must strike a balance between three different facets of exploiting predictive analytics models: accumulating the right data to build accurate models, ensuring that the models are complete and accurate, and using the models at the right time and place.

It is important to review how business analytics applications are configured, utilized and put into production to determine the best way to overcome the challenges that may impede optimal use.

Editor's note: This article was updated in December 2021 with the addition of two people-related challenges.

Next Steps

7 top predictive analytics use cases: Enterprise examples

Predictive analytics vs. machine learning

Dig Deeper on Data science and analytics

Data Management
Content Management