|
|
Research Projects
Current Projects
Collateralization of Intellectual Property
The availability of competitively priced debt and equity financing for small start-up technology businesses continues
to constrain the growth of technology nationwide. Historically, the most formidable obstacles that confront a start-up company
are a lack of access to capital and a limited cash flow. Commercial lenders, often avoid the start-up technology market
entirely, or at a minimum demand an accelerated repayment schedule that, given the financial constraints and management
difficulties associated with new businesses, typically result in under-capitalization and a high risk of
failure.
M-CAM is a new corporation formed with the goal of providing lending institutions access to a markets that, to date, have been inaccessible to traditional lenders and investors. M-CAM offers lenders a guaranteed purchase price for intellectual property owned by a loan applicant. This effectively allows the lender to treat intellectual property as collateral. The range of assets considered by M-CAM includes patents and associated
licenses; marketing and distribution rights; trademarks and copyrights; trade secrets and marketing plans; and hardware,
equipment and other unique assets.
Our research with M-CAM focuses on improving the accuracy and objectivity of the valuation process. The central task in this
work is the development of mathematical models to predict the likely value of intellectual property, relative to comparable
assets for which market valuations are known. The automated prediction capabilities of the models allow M-CAM to limit the
scope and number of judgments required of evaluation officers, resulting in a faster and more objective evaluation
process.
Combination of Credit Scores and Classifier Outputs
The typical lender has access to several different credit scores on each individual applicant for credit. A lender might, for example, have both a credit bureau and an application score on each applicant.
It has been observed that, in such situations, one can construct a superior score by combining the available scores. The goal of our study is to investigate Bayesian methods for score combination. We show that a Bayesian method leads to improved expected profit, improved divergence and ROC shape, and an improved efficient frontier defined in terms of expected profits and expected market share. We are currently
generalizing these methods and results so that they hold in a general classification framework, with both discrete and continuous outputs.
Dominant Score Cutoff Strategies
The purpose of this research is to develop new results for (1) the equivalence of statistical, business and economic dominance in risk
scoring, (2) dominant risk scoring strategies in the presence of non-dominant scores, and (3) the effect of Bayesian score
combination on dominant risk scoring strategies. One can show that there is ROC dominance if and only if there is dominance of
expected profits or efficient frontiers that involve different business measures such as profit/volume tradeoffs. If there is
no such dominance, an intersection of the ROC curves for two different scores nevertheless yields a dominant strategy for use
of the different scorecards and the cutoffs. Finally, we show that a Bayesian combination of the two scores leads to a
dominant ROC curve with a single dominant strategy.
New Risk Concepts in Portfolio Optimization
An optimal portfolio selection model should take into account the investor's attitude towards risk. We argue that variance as a generally accepted risk measure is not capable of accurately describing investment risk. Risk, as a measure of both likelihood and consequence of adverse effects resulting from configuration of a portfolio, should
encompass both the mean and the variance of the portfolio return. We are proposing a new portfolio optimization model based on Markowitz's
mean-variance theory in an attempt to accommodate the risk attitude of different investors. We are studying a number of test scenarios to show how the optimal portfolio varies according to the investor's risk attitude.
Pricing of Exotic Options
The past decade has seen financial option contracts become popular methods of risk mitigation. Major users of these contracts
include portfolio managers and large corporations. Since risk exposure differs across users, commercial banks have begun
customizing option contracts for clients. In order to ensure a profit, the banks must be able to estimate the value of the
contract accurately. A range of option valuation models have existed since the early seventies, but these methods are generally ill-equipped to
handle the "exotic” options that are created today. Simulation has proven to be the most flexible method of valuing these
options, but large underlying state spaces invoke the curse of dimensionality, rendering simulation infeasible in many situations.
Neuro-dynamic programming (NDP) is a form of approximate simulation that deals with large state spaces by approximating the
value at each state, rather than calculating it exactly.
The purpose of the research is twofold. First, we aim to show that NDP is a valid pricing method for financial
option contracts. This will be done by valuing simple options and showing that the estimated prices are consistent with
prices generated by simpler models. Secondly, we will create a NDP model development process that will serve as
a guideline for an analyst who wishes to estimate the value of any exotic financial contract. To illustrate the process,
we will develop valuation models for various exotic options.
Cliques in Bayesian Networks
Bayesian Networks (BNs) are directed acyclic graphs, in which each node represents a random variable of interest and each
directed edge represents statistical or causal dependencies between variables. The notion of a clique is important in classification applications of BNs. A clique in a BN is a group of variables that are conditionally independent of each other given the state of the class variable. If the cliques in the network are known, the computational complexity of the task of learning network structure from data can be considerably reduced. Additionally, identification of cliques provides insight about the relationship between variables, which can be valuable in designing enhancements and updates to the classification model. We are developing new algorithms for clique identification. Related work includes the development of algorithms for learning the structure of BNs with continuous, non-Gaussian variables.
Recent Projects
Palmtop Loan Processing Systems
A loan processing application was designed and implemented on palmtop computers for mobile bankers. Two portable handheld devices (3Com’s PalmPilot and Microsoft’s Windows CE based computer) were evaluated and feasible development methods for the lending application were investigated. The lending application is an automated system that is accessible through a graphical user interface on the palmtop computers. Loan officers can enter data through the palmtops, store the loan information locally, submit this information to be processed at a server for credit scoring analysis, and then download the analysis
and lending recomendation.
Diversifying Credit Union Membership
Traditionally, membership in a credit union has been limited to a group that shares a common employment or occupation, such as teachers or employees of a particular firm. The narrowness of membership criteria has caused credit unions to bear significant risks associated with downturns in sectors of the economy and with strikes or labor actions in industries. Recent legislation has considerably relaxed restrictions on credit union membership, opening up the possibility that these risks can be mitigated though diversification of the membership bases across occupations and employers. Our research centers on identifying optimal diversification strategies within the framework of risk-return portfolio theory.
Credit Collections Systems
Collection of delinquent accounts is one of the most important functions in managing credit accounts. The goal of this project was to develop a forecasting and prioritization system for debt collection in the context of call center operations. Related work includes the development of predictive and simulation models that help lenders gaue the probable effects of various collections strategies.
|
|
|
