Computers are making more and more decisions for us, and increasingly so in areas that require human judgment. There is a palpable increase in machine intelligence across the touch points of our lives, driven by the proliferation of data feeding into intelligent algorithms capable of learning useful patterns and acting on them. A natural question to ask is how we should be thinking about the role of computers in managing our money. Should we trust our money to a robot? In an era of big data and machines to make sense of it all, do machines have an inherent advantage over humans? There is a surge of interest in Artificial Intelligence for financial prediction. Should we pay attention? Or is this an area where human judgment and input is always essential?

Predictive models in regression and classification problems typically have a single model that covers most, if not all, cases in the data. At the opposite end of the spectrum is a collection of models, each of which covers a very small subset of the decision space. These are referred to as “small disjuncts.” The trade-offs between the two types of models have been well documented. Single models, especially linear ones, are easy to interpret and explain. In contrast, small disjuncts do not provides as clean or as simple an interpretation of the data, and have been shown by several researchers to be responsible for a disproportionately large number of errors when applied to out-of-sample data. This research provides a counterpoint, demonstrating that a portfolio of “simple” small disjuncts provides a credible model for financial market prediction, a problem with a high degree of noise. A related novel contribution of this article is a simple method for measuring the “yield” of a learning system, which is the percentage of in-sample performance that the learned model can be expected to realize on out-of-sample data. Curiously, such a measure is missing from the literature on regression learning algorithms. Pragmatically, the results suggest that for problems characterized by a high degree of noise and lack of a stable knowledge base it makes sense to reconstruct the portfolio of small rules periodically.

In many applications, large volumes of time-sensitive textual information require triage: rapid, approximate prioritization for subsequent action. In this paper, we explore the use of prospective indications of the importance of a time-sensitive document, for the purpose of producing better document filtering or ranking. By prospective, we mean importance that could be assessed by actions that occur in the future. For example, a news story may be assessed (retrospectively) as being important, based on events that occurred after the story appeared, such as a stock price plummeting or the issuance of many follow-up stories. If a system could anticipate (prospectively) such occurrences, it could provide a timely indication of importance. Clearly, perfect prescience is impossible. However, sometimes there is sufficient correlation between the content of an information item and the events that occur subsequently. We describe a process for creating and evaluating approximate information-triage procedures that are based on prospective indications. Unlike many information-retrieval applications for which document labeling is a laborious, manual process, for many prospective criteria it is possible to build very large, labeled, training corpora automatically. Such corpora can be used to train text classification procedures that will predict the (prospective) importance of each document. This paper illustrates the process with two case studies, demonstrating the ability to predict whether a news story will be followed by many, very similar news stories, and also whether the stock price of one or more companies associated with a news story will move significantly following the appearance of that story. We conclude by discussing how the comprehensibility of the learned classifiers can be critical to success.}

We compare four nonlinear methods on their ability to learn models from data. The problem requires predicting whether a company will deliver an earnings surprise a specific number of days prior to announcement. This problem has been well studied in the literature using linear models. A basic question is whether machine learning-based nonlinear models such as tree induction algorithms, neural networks, naive Bayesian learning, and genetic algorithms perform better in terms of predictive accuracy and in uncovering interesting relationships among problem variables. Equally importantly, if these alternative approaches perform better, why? And how do they stack up relative to each other? The answers to these questions are significant for predictive modeling in the financial arena, and in general for problem domains characterized by significant nonlinearities. In this paper, we compare the four above-mentioned nonlinear methods along a number of criteria. The genetic algorithm turns out to have some advantages in finding multiple "small disjunct" patterns that can be accurate and collectively capable of making predictions more often than its competitors. We use some of the nonlinearities we discovered about the problem domain to explain these results.

A common view about data mining is that it is an exercise of “clustering” customers, markets, products, and other objects of interest in useful ways from large amounts of data. In this paper, I demonstrate that the real value of data mining, particularly in the financial arena, lies more in revealing actions that lead to interesting distributions of outcomes, distributions that are not directly observable in the data. Simulating actions or events that did not occur is often more useful than trying to cluster the data, which represents only those states of nature that occurred and were recorded. I show how the use of counterfactuals, which are hypothetical events, coupled with certain types of machine learning methods produce models that promote human dialog and exploration that does not otherwise occur in routine organizational activity. I demonstrate with real-world examples, how the combination of database systems, counterfactuals and machine learning methods combine to provide a powerful bottom-up theory building mechanism that is useful in enabling organizations to use databases for learning about things that are useful to them.

Assembling financial instruments such as equities, bonds, options, and other derivatives into a portfolio requires a thorough understanding of how the portfolio will behave in response to changes of specific economic variables and parameters of the instruments. With more information about a more diverse set of instruments becoming available to traders, it is becoming important to limit the complexity of the analysis involved. We show how this complexity can be limited by using qualitative analysis, where the objective is to construct a few good vehicles which can then be analyzed quantitatively. We illustrate how two qualitative reasoning techniques — qualitative simulation and qualitative synthesis — are used to design investment vehicles for risk management purposes. These techniques are currently employed by a prototype expert system that aims at assisting traders solving a risk management problem called hedging.

While other work in model management has looked at models as static objects to be administered in database-like fashion, we emphasize the central role of process knowledge in model construction, usage, and maintenance. We claim that tools are needed to collaboratively elicit knowledge, to represent it conceptually, and to maintain its consistency over time by constraint propagation and reason maintenance. We describe examples of such tools and propose a general approach for integrating them.

The problem of discovering interesting patterns in large volumes of data is studied. Patterns can be expressed not only in terms of the database schema but also in user-defined terms, such as relational views and classification hierarchies. The user-defined terminology is stored in a data dictionary that maps it into the language of the database schema. A pattern is defined as a deductive rule expressed in user-defined terms that has a degree of uncertainty associated with it. Methods are presented for discovering interesting patterns based on abstracts which are summaries of the data expressed in the language of the user.>

We present a general representation for problems that can be reduced to constraint satisfaction problems (CSP) and a model for reasoning about their solution. The novel part of the model is a constraint-driven reasoner that manages a set of constraints specified in terms of arbitrarily complex Boolean expressions and represented in the form of a dependency network. This dependency network incorporates control information (derived from the syntax of the constraints) that is used for constraint propagation, contains dependency information that can be used for explanation and for dependency-directed backtracking, and is incremental in the sense that if the problem specification is modified, a new solution can be derived by modifying the existing solution. The constraint-driven reasoner is coupled to a problem solver which contains information about the problem variables and preference orderings.>

Support for various stakeholders involved in software projects (designers, maintenance personnel, project managers and executives, end users) can be provided by capturing the history about design decisions in the early stages of the system's development life cycle in a structured manner. Much of this knowledge, which is called the process knowledge, involving the deliberation on alternative requirements and design decisions, is lost in the course of designing and changing such systems. Using an empirical study of problem-solving behavior of individual and groups of information systems professionals, a conceptual model called REMAP (representation and maintenance of process knowledge) that relates process knowledge to the objects that are created during the requirements engineering process has been developed. A prototype environment that provides assistance to the various stakeholders involved in the design and management of large systems has been implemented.>

This article reports findings of empirical research that investigated information searchers' online query refinement process. Prior studies have recognized the information specialists' role in helping searchers articulate and refine queries. Using a semantic network and a Problem Behavior Graph to represent the online search process, our study revealed that searchers also refined their own queries in an online task environment. The information retrieval system played a passive role in assisting online query refinement, which was, however, or that confirmed Taylor's four-level query formulation model. Based on our empirical findings, we proposed using a process model to facilitate and improve query refinement in an online environment. We believe incorporating this model into retrieval systems can result in the design of more “intelligent” and useful information retrieval systems.

The maintenance of large information systems involves continuous modifications in response to evolving business conditions or changing user requirements. Based on evidence from a case study, it is shown that the system maintenance activity would benefit greatly if the process knowledge reflecting the teleology of a design could be captured and used in order to reason about he consequences of changing conditions or requirements, A formalism called REMAP (representation and maintenance of process knowledge) that accumulates design process knowledge to manage systems evolution is described. To accomplish this, REMAP acquires and maintains dependencies among the design decisions made during a prototyping process, and is able to learn general domain-specific design rules on which such dependencies are based. This knowledge cannot only be applied to prototype refinement and systems maintenance, but can also support the reuse of existing design or software fragments to construct similar ones using analogical reasoning techniques.>

Problem solvers that use heuristics to guide choices often run into untenable situations that can be characterized as overconstrained. When this happens, the problem must be able to identify the right culprit from among its heuristic choices in order to avoid a potentially explosive search. In this paper, we present a solution to this for a certain class of problems where the justifications associated with choice points involve an explicit assessment of the pros and cons of choosing each alternative relative to its competitors. We have designed a problem solver that accumulates such knowledge about the pros and cons of alternative selections at choice points during heuristic search, which it updates in light of an evolving problem situation. Whenever untenable situations arise, this preserved knowledge is used to determine the most appropriate backtracking point. By endowing the backtracker with access to this domain‐specific knowledge, a highly contextual approach to reasoning in backtracking situations can be achieved.