Neural networks are popular techniques for modeling the probability of an event. Most Neural Network implementations however do not deal with missing values of special cases -- e.g. that there has not been an event to measure. In this paper we demonstrate how we can incorporate such special values into the neural network paradigm where they can also be trained via the backpropagation algorithm.

We demonstrate how the approach increases the K-S statistic by 2 percentage points, and the monetary impact it could have on typical credit limit management strategy.

Increasingly in today's world we are faced with situations where multiple actions are possible and the optimal sequence of events are dictated by business policy and strategy. For example, a financial institution might want to focus on reducing short-term defaults but be more flexible for other timeframes. Traditional optimal control theoretic approaches do not work well under changing objective functions.

In this paper we develop the application of reinforcement learning paradigm, in its modern form, to discovery of optimal policies for risk management, specifically in the sequence of actions to take when a customer misses a payment for the first time. We show how we can use reinforcement learning to obtain the best sequence of actions for individual customers. This has very direct application to credit collections and recovery.

Given the vastness of transactional data and the unpredictability of its volume at granular levels, behavior modeling has to balance information density with practical issues such as computational complexity and fast response times under uncertain loads.

Recursive profiling is one way to address those conflicting goals. Unlike standard behavioral model framework, where multiple pieces of information, often highly condensed, are evaluated at the same time, recursive profiling simply updates the entity profile using some rules of thumb or empirical guidance. However, with that advantage comes a handicap as new models are installed and learn the entity behavioral preferences.

In this study, we review various ways of recursively updating the customer or entity profiles and show how long it takes for them to mature, or reach such a stage that the movement in terms of trends is much lower than stochastic movements.

We conclude with remarks on how quickly different aspects of the customer behavior can be ascertained with a high degree of accuracy, and the aspects that need a lot of exploration and observation before becoming predictive. These findings clearly have implications for behavioral modeling and the predictors used for various behaviors and behavioral approaches.

Event-triggered marketing is a strategy that triggers an appropriate response to a customer through an appropriate channel at the opportune time based on his or her recent characteristics. This approach can increase the marketing RoI by helping design much more specific and effective marketing campaigns arising from a better understanding of a customer and a sophisticated way of interpreting the information available on a customer in real time. Understanding the deposits to an individual's account allows us to understand the components of his or her income and make intelligent marketing initiatives based on that. For example, a large deposit to a customer's checking account which is either unprecedented or does not fit a pattern indicate that he probably has additional or new sources of income. If this event is detected in a timely fashion and acted upon, the bank would contact the customers at a time when he would be open to consider investment opportunities such as money markets, CDs or treasury bills. The above information can be utilized to cross-sell relevant products to the customer depending on the categories he or she shops or invests in. This approach can be easily extended to cover online bill payments-- for example, if the wireless bill of a customer increases steadily, he could be offered a more suitable plan from a competitor. In this paper we develop the methodology to ascertain which deposits to a customer's account do not fit a pattern and indicate disposable income. The technique used is a hybrid of information-theoretic and statistical pattern recognition techniques. This scheme can be easily extended to cover a whole gamut of services, enabling a sophisticated and smart marketing paradigm that can be applied on a daily basis to house file customers.

Loyalty-based marketing has become a key strategy for most companies in today's competitive marketplace. The practice is based on a very simple premise: as you develop stronger relationships with your best customers they will stay with you longer and become more profitable. With that in mind, we develop a framework for creating a loyalty metric which can be used to identify important customer attributes such as customer inertia, price sensitivity and switching behavior. We show that customer loyalty is not a static or marginal metric but rather a dynamic one. It means that we need to evaluate the history of a customer in order to assign a loyalty score to him or her. We argue that the multidimensional vectors of observed probabilities formed with a moving average window have the potential to correctly quantify and categorize customer loyalty. We develop and use probabilistic and information-theoretic measures to derive expressions for customer loyalty scores. Clearly, having a representative and customer-centric loyalty metric provides competitive advantage for companies that can capture and act on this information. Marketing campaigns can and should be designed to cater to individuals based on the core customer values derived from their loyalty scores. For illustration, we use transactions with supermarkets and home-improvement or do-it-yourself stores to generate customer loyalty scores for two categories and discuss how the loyalty scores vary across these categories. In this study, we use transactional data which does not suffer from self-reporting bias and considers the transaction history of a customer.