Predictive customer relationship management (CRM) ­ the discipline of getting to know your customers by performing complex analysis on data about them ­ is rapidly changing the way companies make operational and strategic decisions about procurement, production, marketing and sales of products and services. Several factors are driving the development of predictive CRM technologies and applications:

Computing power ­ Application of Moore's Law in the new economy. In 1963, a typical CRM application processed data for 600 customers on a room-sized, multimillion- dollar computer with 0.2MB of RAM and processor speed of 0.5 MHz. Today, the bottlenecks are ideas, not computers.

Data sources and volume ­ Dealing with the data deluge is the dilemma. It comes fast and furious from electronic points of sale, stock trades, phone calls, direct mail catalog orders, bank and credit card transactions and, increasingly, online transactions.

The Internet ­ Access to customer relationship information on a new level of quantity, detail and timeliness. The previous technologies and sources of real-time information about customer purchasing patterns and preferences included supermarket scanners and in-home television rating systems. The Internet is revolutionizing the meaning, analyses and application of real-time customer data, enabling better access and more sophisticated predictive methodologies.

Product and service customization ­ Ability of technology to allow custom offerings is driving requirements for more targeted sales and marketing. Mass production gave way to mass customization which, in turn, is giving way to individual customization. The new business need is to better understand customer requirements and preferences, and to translate that information into the right products and services.

This article will discuss how new technologies and applications are changing the CRM landscape and the resultant implications and ramifications for enterprise customer relationship management in the electronic economy.

Predictive CRM: Techniques and Applications

The application of predictive CRM involves three problems:

  1. Classification of customers into groups based on similar behavior toward a given set of marketing and sales actions, usually referred to as customer segmentation.
  2. Describing customer behavior by building a behavior model and estimating its parameters.
  3. Deciding which marketing actions to take for each segment and then allocating scarce resources to segments in order to meet specific business objectives.

Because these three problems are interconnected, integrated approaches that address all three problems simultaneously prove much more powerful than those that address them individually.
In attacking these problems, predictive CRM draws on a variety of disciplines in computer science (in particular, data mining, database management and data visualization), economics, operations research and mathematics. This article describes some of these techniques and how they are applied to contemporary business problems.

Dynamic Segmentation and Product Bundling

A broad area for applying predictive CRM is in dynamic customer segmentation, the continuous reclassification of customers into categories based on observed customers' behavior. In an e- commerce environment where feedback is almost immediate, changes in customer behavior patterns can be observed as different products, services, information and marketing incentives are offered to customer segments. This is the basis for good estimates about the market effectiveness of various campaigns and allows dynamic customer segmentation based on selected business criteria and objectives.

The online channel provides rapid availability of new information about customer choices; as such, it is an excellent environment in which to continuously learn and adaptively improve the segmentation. For example, robust and scalable incremental clustering algorithms are now emerging which can be used to discover new behavioral and attitudinal differences among customers by allowing newly observed behavioral data to immediately participate in the existing data pool. New customer segments may be discovered, and rules for assigning customers to segments may evolve. Dynamic segmentation can be used to quickly adapt channel and campaign strategies effectively targeted at these segments.

Rapid availability of information about customer choices can be used to dynamically determine affinity between products and services in a fine-grain manner and help suggest new product and service bundles. Advances in affinity analysis algorithms, including faster procedures to analyze large quantities of data in near-real time, to analyze sequence affinities (e.g., buying a photo scanner followed by buying a digital camera) and to analyze indirect affinities, can now be used to dynamically personalize offerings to customers (i.e., "segment of one"). Emerging mobile devices coupled with pervasive computing infrastructures have further opened the possibility of collecting "location-aware" behavioral information, the analysis of which will enable the offering of products and services to a customer at the right time and place.

Figure 1 shows the results of an ex-perimental system developed at Hewlett-Packard Laboratories for analyzing product online sales history and constructing product affinity groups.

Figure 1: Visualizing product affinity groups. Each product is represented by a sphere. As products are sold together over time, their individual spheres become adjacent, forming affinity groups in the data visualization.

Adaptive Estimation of Conversion Rates

In a stable, high- volume, low forecast-variability marketing situation where products are mature and markets are stable, the technical requirement of estimating customer behavior is often reduced to determining the smallest sample size needed so that resulting conversion rate estimates are statistically accurate. A static point estimate of conversion rate and its associated confidence interval is quite adequate and will be valid for an extended period of time. However, in today's electronic marketplace, many products have short life cycles and do not have much historical data. Additionally, customer behavior toward these products changes rapidly. At Hewlett-Packard, our lines of mobile computing devices, such as laptops and PDAs, often have life cycles of only three to six months.

In this dynamic and uncertain environment, a more adaptive approach for estimating future customer behavior is required. Adaptive methods rely on dynamic sample sizes that are updated immediately when new events occur, resulting in a quicker, more reliable estimates. In addition, Bayesian approaches allow the meaningful combination of information to be gathered from different sources and at different points in time.

Optimization of Resources

An important goal of predictive CRM is the efficient use of scarce resources such as marketing budgets, product availability and advertising space. Knowing how customers react to a certain marketing action is only one side of the equation; making good decisions about how to allocate resources in such a way that business objectives are optimized across all segments and available marketing actions is equally important. Often, there will be several business objectives ­ number of conversions, profit, revenue, customer acquisition, etc. ­ and a number of physical and business constraints. Analyzing tradeoffs between objectives and determining the impact of a given constraint on the business goal are integral parts of predictive CRM. Techniques used to solve these problems include linear and non- linear programming, dynamic programming, meta-heuristics such as genetic algorithms and simulated annealing, and use of sophisticated heuristics developed specifically for the problem on hand.

Dynamic Pricing

In today's environment, successful enterprises are starting to appreciate the value of selling their products and services using both online and traditional channels, leveraging the strengths of each to complement the others. One feature of the Internet channel is unique: because of the immediate availability of results, products which are sold online using an auction format can quickly yield valuable and useful pricing information for other channels. The characteristic information of an auction ­ bid history, closing price, reserve price, number of bidders, etc. ­ can be used to predict price sensitivity of demand. This information is processed with statistical techniques and converted into price-elasticity relationships which, in turn, can be used to predict the impact on demand of changes in price. Thus, information derived from Internet sales is directly applicable to other channels such as direct mail and brick and mortar.

Figure 2 shows the results of deriving a price-demand relationship from online auction data.

Figure 2: Price-demand relationships from auction data. Statistical methods are applied to historical data from online auctions (left) to derive the price versus relative demand curve (right).

The applications are immediate and valuable. Manufacturers and retailers can get rid of excess inventory by offering the optimal discount or, when appropriate, increasing the price of products to optimally extract economic surplus from the system while ensuring a high level of customer satisfaction.

For consumable products, such as pens and cartridges for printers, dynamic pricing is emerging as a new tool for increasing revenues and/or profits while securing customer loyalty. Using information and communication technologies, manufacturers and retailers can monitor customers' usage of products and offer online advice when the consumables need to be replaced. The business enterprise can predict future usage more accurately and plan production and inventory levels accordingly, and customers can avoid the problem of running out of a critical product or component.

Additionally, retailers can capitalize on varying price-elasticity relationships of different usage patterns and design usage-driven product discount structures that allow increased revenue generation from selected customers. With known product usage patterns, manufacturers and retailers can design incentive structures and policies that benefit both the enterprise and their customers. The customer benefits with a price that is appropriate for his or her specific usage pattern, an incentive structure (such as points redeemable for products and services) and proactive guidance to avoid discontinuity in service. Manufacturers and retailers benefit from increased revenue by offering the right discount on the most price- sensitive usage and all the advantages that accrue from loyalty programs.

Two Sides of the Same Coin

In the past, predictive CRM decision-support processes have prescribed directions and recommendations without consideration of the effect these directions and recommendations may have on other operations within the enterprise. In particular, the results of CRM decisions often have disruptive effects on supply chain management (SCM) operations and planning. In many companies, however, CRM and SCM are isolated systems with limited interoperability. This lack of communication can cause costly and wide-ranging problems.

For example, a computer printer manufacturer might have limited supplies of a printer model for sale over the Internet. Overselling of the designated printer by exposing potential Internet buyers to e-mails and banner ads can deplete the printer supply and lead to discontented customers. Sharing of supply chain printer inventory with the CRM customer contact process can, for example, limit printer advertisements over time as customer orders deplete inventory.

The traditional view of business offers a clear distinction between CRM and SCM. Typically, CRM and SCM refer to where an enterprise is in the value chain: relationships that are "upstream" and coming from vendors are on the supply side and those "downstream" are on the customer relationship side. While these processes may be distinct in the abstract, they are not so distinct in the realm of real-world business where executives and managers must often consider them collectively in many decisions.

The Internet and other advances in information technology have now made it possible to allow fine-grain, real-time interchange of event information between the upstream and downstream processes and between supply chain players. New models are being developed which can simulate and predict the global impact of decisions or events occurring in one process on the other process and ensure that such decisions are stochastically optimized in a global context. For example, effective use of predictive CRM technology is directly linked to the other predictive process faced by most enterprises: forecasting customer demand and planning supply chain operations accordingly. Real-time information about customer buying patterns ­ from scanners in stores and e-commerce ­ is having a significant impact on the quality and usefulness of supply chain demand forecasts. When other supply chain initiatives such as vendor-managed inventory are supported by real-time customer purchase information, customers, retailers and manufacturers benefit. In the future, the real payoff from predictive CRM analysis and decision making will be this positive connection to supply chain management planning. The CRM supply chain combination gives reduced forecast error, resulting in better quality of demand prediction, reduced inventory stocking requirements and better order fulfillment. Customers are satisfied because they can find the goods they are looking for, retailers can reduce inventory costs and increase profits, and manufacturers' margins improve through inventory and price-protection cost reductions. These process enhancements are especially critical for products with short life cycles.

Strategic Relationship Management

Apart from global optimization of operational or tactical CRM and SCM decisions, the traditional view that the distinction between CRM and SCM derives from a flow model of the value chain is limited by today's world of complex partnerships where firms may have reciprocal relationships: they can simultaneously be vendors, customers and competitors. When the distinctions between SCM and CRM are transaction-based instead of based on the enterprise organization chart, then many SCM aspects of the relationship between two firms may be relevant to the customer side of the equation. For example, a computer system company A can be both a supplier of integration services to a chip manufacturer B, as well as a customer who buys chips from B. This becomes even more complex when considering a trading community where any participant may be a vendor or a buyer of any other participant in the community. The distinction between CRM and SCM is a matter of where you sit in the value chain for some aspect of a given transaction. The reciprocal roles among enterprises will require a higher- level integration of relationship management with peer enterprises.

This new way of looking at operational processes has ramifications for how enterprises organize and how executives and managers assign duties and tasks. It requires fresh thinking regarding CRM and SCM, especially in light of the new capabilities of data mining and decision support technologies. Because of the increased complexity and uncertainty associated with these trans- enterprise business processes, a greater value will be placed on predictive technologies such as data mining and decision support to provide busy executives the ability to make holistic sense out of their new environment. The effectiveness and value of a CRM system can be significantly enhanced by the information considered in its supplier relationship management (SRM) system. Emerging strategic relationship management technologies are aiming to enable prediction and decision making that combine CRM and SRM at a strategic level.

Buy or Rent

The traditional software application paradigm of buy-install-maintain-use is quickly giving way to the xSP model of broker- and-use-on-demand. While this service provider model will revolutionize our way of thinking about and using information technology by lowering entry barriers and reducing risks, it is especially beneficial for predictive CRM for several reasons:

Predictive CRM is a quickly evolving discipline ­ Being tied to a specific CRM system or approach that is difficult to maintain and update can result in the use of inferior or outmoded technology. The service provider model allows rapid update of CRM systems and applications and, if warranted, the relatively easy adoption of new solutions and approaches.

The ad hoc nature of predictive CRM ­ There is no "right way" to perform predictive CRM and determine the best relationships a company should have with its customers. The application of CRM is a continually evolving process. The service provider model, with predictive CRM functionality being supplied on demand, is ideal for testing new approaches and adopting the ones that work.

Diversity of CRM information sources and analytic techniques ­ The service provider model provides a wider range of CRM functionality, as well as a diverse choice of information sources about customers. The quality and profitability of customer profiles can benefit from multiple information sources supplied as services.

Predictive CRM as an E- Service

Offering predictive CRM as an e-service is a form of business intelligence service provisioning (BISP). For example, targeting rules may be computed by a BISP using a variety of analytic tools, techniques and sources of information, and supplied to an enterprise's store-front engine for deployment. While primitive BISPs are emerging, the real ramp up of their adoption will rely on e-service technologies which enable secure, trusted, dynamic interconnection among enterprises to be rapidly developed and deployed. These technologies provide frameworks and tools to allow service providers' services to be universally described, be discovered by potential customers and become accessible to customers' operational systems (e.g., a store-front system). They also allow the components of a customer's operational systems to be properly "wrapped" into units that can automatically engage in a protocol- driven interchange with the components at the service provider's side.

Predictive CRM is useful to enterprises that recognize the real-time realities brought about by a combination of the Internet as a vehicle for commerce and powerful advances in computational technology. The application of disciplines such as data mining, operations research and economics enables near real-time analysis of customer data and lets an enterprise more effectively forecast, price and promote products and services into the market. Moreover, with the increasing reciprocal nature of relationships between many firms, especially in electronic marketplaces, CRM and SCM need to be treated as two aspects of a unified process to optimize a firm's plans and actions. With the ability to move predictive CRM to an e-service, hosted model, more enterprises will have opportunities to effectively target their customers. As a result, predictive CRM is a fast-moving discipline that will rapidly evolve in the future.

Most electronic economy businesses never see their customers; they only see data about those customers. The advent and adoption of predictive CRM technologies is allowing businesses to increase the personalization in a previously impersonal process.

Acknowledgments: We want to thank our colleagues at Hewlett-Packard Laboratories for critical reading and contributions to this article: Dirk Beyer, Ming Hao and Alex Zhang.

Shailendra Jain ( is manager of the decisions technology department of Hewlett-Packard Laboratories.

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