Predictive modeling and decision management are moving beyond traditional uses in the credit risk management world and into the revenue enhancement arena. A particularly interesting application is in the area of tailoring offerings or responses to customers as the transaction is happening in real time. Conceptually, the responses received from a customer during a transaction can be used to generate additional offerings of complementary products or even to set prices in a real-time context.

The idea of being able to predict what a customer might additionally buy is exciting to business managers. As with any new approach, many ideas bubble up that sound great in the conference room but fall flat when the realities of implementation are considered. IT and business managers need to evaluate the use of data mining, modeling and decisioning approaches to determine whether an idea in front of them makes business sense and can be implemented technically.

There are three major areas of analytics to be considered: data mining seeks to discover relationships in the data about your business that can drive improved performance. This supports predictive modeling, where mathematical or information theory-based techniques are used to analyze a set of data and create a model that can predict the occurrence of customer behavior in the future. This enables decisioning, where a set of decision rules are developed from data analysis. The resulting decision tree or matrix is used to recommend a course of action based on the customer's characteristics.

These technologies are related but separate. Data mining and predictive modeling measure past relationships and use that information to create probability functions that predict future events. These are best thought of as suggestive approaches. They can provide hints to actions that can be taken. These hints become high value inputs into decisioning or the use of decision trees to direct actions. Decisioning is concrete and yields a direction for the activity based on the inputs. Together, these approaches can be used to steer an interaction with a customer in real time.

An example might be as follows: Data mining of an electronics company's sales records yielded information that purchasers of a Yamaha home theater receiver are likely to purchase upscale cabling equipment, a high margin product, while purchasers of a Magnavox receiver are not.

A predictive model is used to create a purchase-likelihood score for a range of products, based on the current contents of a shopping basket and pages visited in the site. If the customer has already visited the page of the high-end cables and not placed them in the basket, this suggests that they won't respond to an offer of these cables at full price, but might respond to a lower-cost offering.

A decision score, based on likelihood to respond along with the profit potential of the item, is calculated for a range of products, which results in a suggestive sell during the checkout process.

By using all their available data, this electronics company is able to drive additional revenue through existing interactions. They could further enhance their customer interactions through price modeling, which might suggest a discount that would be appealing to a difficult customer.

Usefulness of Real-Time Analytics

The electronics company example can be generalized to describe situations where the use of analytics in real time is likely to be helpful:

  • Data is collected in the process that is unique to the transaction.
  • The process is flexible and offers the opportunity to improve an outcome through manipulation of the process.
  • A decision structure can be planned and used as a way to help direct personnel or process. A good example of this is financial planning interactions, where the information about a customer can suggest appropriate products as well as support a model for a propensity to buy each of the products.

In contrast, there are also situations where real-time analytics does not make sense, for instance when the interaction with the customer doesn't provide data that is sufficient or relevant to model. Many call center interactions are not conducive to trying to steer the transaction. Customers don't want to be interrogated, and call center personnel have a finite capacity for responding to interactive direction.
The automated environment is not sufficiently modifiable to be able to make use of a prediction. An ATM, for example, doesn't have the screen real estate or ease of programming that would make it productive as a suggestive sell tool.

There is no desire to change the actions taken for the customer. If you are always going to suggest an extended warranty, there is no need to try to predict whether the customer will respond.

Barriers to Real-Time Analytics

Prediction-based transaction direction has been in use for some time at online retailers such as Likewise, large financial institutions have been using decisioning analytics to prompt and direct customer interactions. These initiatives have been the province of these large companies because of three factors:

  • Cost and inertia of obtaining modeling data to perform the research and develop the models. This data must be captured from enterprise systems, which rarely store data in a form useful for modeling. The growth of data warehousing reduces this barrier.
  • Technical challenges of integrating these models into the systems infrastructure that is being used to manage the customer transactions. Historically, models must be implemented in programming code, which can be costly and risky. The emergence of runtime environments for model execution reduces deployment effort and risk.
  • Esoteric tools and techniques to be used to develop and implement the models. These have historically been the domain of statistical math packages, which require specialized and experienced analysts to be used effectively. The evolution of modeling workbenches and decision tool editors has reduced the use barriers of these types of products.

The costs of these three factors have dropped dramatically in recent years. Implementing a data warehouse used to require expensive storage network systems and specialized data tools. Now large storage platforms can be built on inexpensive Linux servers for less than $1,000 per terabyte, and the analysis tools are packaged economically with the database software. Several modeling software vendors now offer execution engines that can be integrated easily into system environments, reducing implementation expense and risk as well as offering easy addition and updating of models. Finally, simpler modeling platforms have been developed that focus explicitly on the workflow of development and validation of models.
All of these factors point to the likely rise of analytic use in business, which means that businesses need to plan on how to develop their own capability if they wish to maintain their competitive position. A prudent manager would do well to assess the level of opportunity in his or her company's space. But how do you do that?

Uncovering Opportunities for Real-Time Analytics

Prime opportunities where analytics technology is likely to provide advantage have these attributes:

  • Customers vary recognizably in some trait/traits of interest. The simplest example is willingness to buy a specific product or service. The next level is detecting complementary goods. The ultimate goal is complementary cross-sells or up-sells that are not apparent from the choice of the good. Most audio equipment customers need cables, so this is deducible without statistical mining techniques. A more subtle example would be an online purchaser of economical bed sheets in the month of August might also be a customer for a desk lamp because he is buying supplies for his college dorm room.
  • The criteria for the decision are unique to the instant in time. If you always intend to cross-sell a TV stand to purchasers of a TV, you don't need to invest in a predictive engine to make that link, when a table of suggested sell items can be implemented more economically.
  • There is a choice that the business can make that would improve the profitability of the transaction. This can be through an additional sale, maximizing the price charged or avoiding attrition.
  • The business has the option of changing its behavior if a reasonable prediction could be made.

Developing Analytic Capability

Once you've uncovered trends in your business that suggests the need for real-time analytics, you will need to focus on the following areas to bring that capability to fruition:

Data. The most important step to take is to start hoarding your data. Analytics requires data, and the data that you need is likely not being captured in your enterprise systems. The core capability you need is a data warehouse that captures the states of your customer data repeatedly, over time, along with relevant transaction fact data so that you can determine outcome data. This is now a well-understood activity, and most organizations have some capability in this area, used most often for reporting and analysis.

Staff. You need to have someone that understands model development. This has historically been the domain of statisticians, but advanced applications have embedded the sophisticated science into the technology so that computer science graduates and even some line-of-business managers can develop models. However, the skills required should not be minimized - an inappropriately constructed model can be worse than useless. One fledgling catalog company inverted a score and sent their Christmas catalog to customers least likely to purchase. By the time they discovered the error, the season was upon them, and having no time to react, they lost a year's worth of business and subsequently failed.

Partners. If an in-house modeling staff is not feasible, you should begin exploring outsourced providers. Simply engaging in discussions with vendors will guide you to possibilities that you hadn't considered and will provide input into your data warehouse development to energize future projects.

Risk versus Reward: Why You Should Take This On

At this point, you might be asking, why bother with the effort and risk associated with real-time analytics and model-driven processing? If the models can fail and require additional staff to develop and maintain, why add this difficulty to the work day? The answer: value lies hidden in the volumes of data that can be analyzed. Statistical mining and modeling techniques can discover relationships in data that humans simply cannot. Often, these relationships are unknown to the owners of the data but are very useful in developing future interactions with customers. This technology improves the quality of response, while supporting a tailored approach to the customer. Real-time analytics drives consistent and managed interactions, while maximizing metrics important to your business.

Register or login for access to this item and much more

All Information Management content is archived after seven days.

Community members receive:
  • All recent and archived articles
  • Conference offers and updates
  • A full menu of enewsletter options
  • Web seminars, white papers, ebooks

Don't have an account? Register for Free Unlimited Access