Applying AI to Finance. The Symbolic and Sub-Symbolic approaches.

University Degree Mathematical and Computer Sciences

Introduction

Recent years have seen a broadening of the array of computer technologies applied to finance, showing that the diversity and richness of the applications of AI addressed to solve problems in finance is very successful and developing.

Solving complex problems and dealing with uncertainty, such as financial investment planning, foreign exchange trading, and knowledge discovery from large/multiple databases, involves many different components or sub-tasks, each of which requires different types of processing. To solve such complex problems, a great diversity of intelligent techniques are required, which can be divided into two approaches: symbolic, which includes traditional hard computing techniques such as expert systems and sub-symbolic, which includes soft computing techniques such as fuzzy logic, neural networks, and genetic algorithms

Each technique has particular strength and limitations, and cannot be successfully applied to every type of problem. Moreover, some of the techniques are complementary in many aspects, so they can mutually compensate weaknesses and alleviate inherent problems. These results in systems called hybrid intelligent systems, which have recently begun to gain prominence as a potential tool in solving a wide variety of complex tasks. According to Zahedi (1993), expert systems and ANN offer qualitative methods for business and economic systems that traditional quantitative tools in statics and econometrics cannot quantify due to complexity in translating the systems into precise mathematical functions.

Expert systems Advantages and disadvantages

To begin with, the most used Symbolic approach to AI methods in financial field have been expert systems which deal best in the field of financial analysis. Expert Systems possess knowledge acquired in practice and which cannot be found in literature or acquired in any other way, but which is invaluable to a business success of a firm or a financial institution. Therefore, the first and the most important advantage and purpose of creating ES is to make the domain knowledge of an expert accessible to a wider circle of people. This would ensure business existence and survival when the expert and his knowledge of doing the work are no longer available to the company. This leads to some other advantages such as reproductively, as many copies of an expert system can be made, while training new human experts is time-consuming and expensive.

Important advantages of using an expert system are the uniformity of knowledge and possibility of its improvement over time. The knowledge of multiple human experts can be combined to give a system more breadth that a single person is likely to achieve. For example, if an expert system is used in giving help while evaluating risks of investment in a firm, then every relevant parameter is treated with special attention, without the fear that some situations could be differently estimated by various experts, or that some important parameters would not be taken into consideration. If any new parameter is important for the company, then it can be continuously augmented as necessary with accumulating experience.

The ability to provide users with explanations of the reasoning process is important for complex decision making. Explanation facilities are required, both for user acceptance of the decisions made by a system, and for the purpose of understanding whether the reasoning procedure is sound. Good examples of this requirement can be found in loan granting, legal reasoning. There have been fairly successful solutions to the explanation problem in expert systems; symbolic machine learning and case based reasoning learning. In expert systems, explanations are typically provided by tracing the chain of inference during the reasoning process.

This is a preview of the whole essay

On the other hand, ES also have their disadvantages. One of the biggest limitations of ES is that they require full information about outcomes and therefore deal poorly with uncertainty, which is essential for making financial decisions. When no answer exists or when the problem is outside their area of expertise, the knowledge of humans is observed to gain acquisition. Video tapes, interviews, protocol, and other techniques are used to try to capture the thought process of experts. This is a crucial stage in the development of intelligent systems. As a process, it involves eliciting, interpreting and representing the knowledge from a given domain. Knowledge acquisition for expert systems is time consuming, expensive and potentially unreliable. Furthermore, expert systems do not have mechanisms to deal with any changes in their operating environment and therefore require continuous updating. Thus the maintenance of knowledge in expert systems which is very important in the changing business environment is also time consuming and expensive.

Another problem with expert systems is writing the rules themselves. Thought processes that are highly rule oriented such as credit evaluation where the nature of the task is repetitive and unstructured like loan granting, ES are easier to write and the company benefits of using ES through the speed and accuracy, both which far exceed human capacity. However, the problems that rely more on creativity or intuition are very hard to write and in some cases is even impossible, as argued by Huber Dreyfus , human expertise depends on unconscious instinct rather than conscious symbol manipulation and on having ‘feel’ for situation rather than explicit symbolic knowledge.

Expert systems are improving as technology advances. In the past, expert systems have received criticism and some negative publicity because of the failures that were highly publicized. Unfortunately, the successes are less publicized, because companies want to maintain their competitive edge. In my view ES could create barriers of entry for potential competitors and in some cases an expert system can differentiate a product or can be related to the focus of the firm thus reducing the risk of doing business.

Expert systems are a great tool for companies especially, as depicted here, companies in finance. However, it is important for companies to remember that humans should make the final decision, and not the computer. Humans still have the insight and intuition that computers are unable to possess now. The arguments for and against of AI ability to make computers think and feel is still taken place. Therefore, ES should be viewed as a complementary to humans’ decisions in order to get the most benefits for the business and reduce risk minimizing uncertainty, which are the greatest challenge facing managers and researches in field of finance.

Preparation and analysis of reports

As mentioned above ES are very good at financial analysis field. CoverStory is an expert system developed by IRI to tackle the problem of too much data. Ocean Spray has been a development partner and first client. CoverStory automates the creation of summary memoranda for reports extracted from large scanner databases. The goal is to provide a cover memo, like the one a marketing analyst would write, to describe key events that are reflected in the database - especially in its newest numbers. The project began as a teaching exercise in marketing science - "How would you summarize what is important in this data?" (Stoyiannidis, 1987; Little, 1988) - and has developed into a practical tool. CoverStory is a particularly desirable development because, with very little effort, it provides users with top line summaries and analyses across a wide variety of situations. Previously this required time-consuming intervention by a skilled analyst. Furthermore the technology is an extensible platform on which to build increasingly sophisticated decentralized analysis for the user community.

Appraising loan applications

ALEES – an agricultural loan evaluation expert system that incorporates qualitative factors such as a loan officer’s intuition, experience and judgement as well as quantitative factors (Bryant 2001) In this case the major advantage of the ES is time saved -- a decision can be obtained from it in 30 seconds that

ANN Strengths and Weaknesses

The most used Sub-Symbolic approach to AI is Artificial Neural Networks (ANN), which deals best with uncertainty, as they could be seen as information processing systems which use learning and generalization capabilities and are very adaptive. In contrast to ES, artificial neural networks are easy to construct and deal very well with large amounts of noisy data. They are especially suited to solving nonlinear problems. They work well for problems where domain experts may be unavailable or where there are no known rules, for tasks that require constant adaptation and learning from the operating environment. Therefore, the main advantage of ANN is their adaptively in nature, which makes them particularly useful in fields such as finance where the environment is potentially volatile and dynamic.

ANN robustness in storing and processing data, as they are very tolerant of noisy and incomplete data sets, earned them some applications, where fault tolerant types of equipment are required. This flexibility derives from the fact that information is duplicated many times over in the many complex and intricate network connections in ANNs, just like in the human brain.

The training process of an ANN itself is relatively simple. The pre-processing of the data, however, including the data selection and representation to the ANN and the post processing of the outputs, required for interpretation of the output and performance evaluation, require a significant amount of work.

There are many statisticians who argue that ANNs are nothing more than special cases of statistical models, and thus the rigid restrictions that apply to those models must also be applied to ANNs as well. However, constructing a problem with ANNs is still perceived to be easier than modelling with conventional statistical methods. There are probably more successful novel applications using ANNs than conventional statistical tools. The prolific number of Ann’s applications in a relatively short time could be explained by the universal appeal of the relatively easy methodology in setting up an ANN to solve a problem. The restrictions imposed by many equivalent statistical models are probably less appealing to many researchers without a strong statistical background. ANN software packages are also relatively easier to use than the typical statistical packages. Researchers can successfully use Ann’s software packages without requiring full understanding of the learning algorithms. This makes them more accessible to a wider variety of researchers. ANN researchers are more likely to learn from experience rather than be guided by statistical rules in constructing a model and thus they may be implicitly aware of the statistical restrictions of their ANN models.

In contrast to ES, reasoning in neural networks involves the numeric aggregation of representation over the whole network. This distributed representation and reasoning allows the systems to gracefully degrade. That is, even if some parts of neural network are made non-operational, the rest of the neural network will function and attempt to give an answer. This type of inherent fault tolerance contrasts strongly with expert systems, which usually fail to function even if one single processing part is non-operational.

The major weakness of ANNs is their lack of explanation for the models that they create. Research is currently being conducted to unravel the complex network structures that are created by ANN. Even though ANNs are easy to construct, finding a good ANN structure, as well as the pre-processing and post processing of the data, is a very time consuming processes.

The lack of explanatory capabilities is considered as the main shortcoming of the application of neural networks. The adduced incapacity to identify the relevance of independent variables and to generate a set of rules to express the operation of the model makes neural networks be usually deemed as “black boxes”. Primarily, ANNs have a good ability to represent "empirical knowledge", like the one contained in a set of examples, but the information is expressed in a "sub-symbolic" form -- i.e., in the structure, weights and biases of a trained ANN, not directly readable for the human user. Thus, an ANN behaves almost like a "black box", providing no explanation to justify the decisions it takes in various instances. This forbids the usage of ANNs in "safety-critical" domains, which include the economic and financial applications, and makes it difficult to verify and debug software that includes ANN components.

Bankruptcy predictions.

(From the study of Dorsey et al., University of Missouri Business School, 1995, as cited in Siegel and Shim, 2003.)A neural network was used to predict which companies would become bankrupt during the year 1995, based on publicly available information (from America Online’s Financial Reports database) from the preceding year 1994. Information about 20 companies was used to train the net, 10 of which did become bankrupt during 1995, 10 of which did not but were considered to be high-risk companies which had been showing some signs of financial distress. The input data took the form of various ratios between financial quantities which might have a bearing on the financial health of the companies. These financial ratios were chosen because they had been used since the 1960s to predict company insolvency (although within the context of statistical analyses, not neural network models).

Predicting cash machine usage

Banks want to keep cash machines filled, to keep their customers happy, but not to overfill them. Different cash machines will get different amounts of use, and a neural network can look at patterns of past withdrawals in order to decide how often, and with how much cash, a given machine should be refilled. Siemens developed a very successful neural network system for this task; in benchmark tests in 2004 it easily outperformed all its rival (including non-neural) predictor systems, and, as reported below, could gain a bank very significant additional profit from funds that would otherwise be tied up in cash machines.

Advantages of hybrid systems

According to Arash Bahrammirzaee HIS is an efficient and robust learning system which combines the complementary features and overcomes the weaknesses of the representation and processing capabilities of symbolic and non-symbolic learning paradigms. HIS is a system that that integrates intelligent techniques to problem solving. Therefore, the combination of the symbolic and sub-symbolic systems could provide superior knowledge and performance. The interest of using them together is of rising interest in artificial intelligence. The current view is that sub symbolic and symbolic approaches are complementary, not competing. Intelligent hybrid systems are a very powerful class of computational methods that can provide solutions to problems that are not solvable by an individual intelligent technique alone.

The opportunity of employing neural techniques in expert systems is often suggested on the ground that the learning, generalization, fault, and noise tolerance capacities of neural networks can alleviate well-known shortcomings of symbolic problem solvers, such as brittleness in front of incomplete or noisy data, no increase in performance with experience, and time-consuming knowledge acquisition.

A hybrid approach can facilitate the development of more reliable and effective intelligent systems to model expert thinking and to support the decision making processes. It has diverse potential and is a promising direction for both research and application development in finance in the years to come.

The motives of interest for combining rule-based and neural computations in expert systems are:

Technique enhancement. Automatic data acquisition for expert systems may require kinds of sensory processing that are effectively dealt with by neural nets. In turn, hypothetical reasoning is often called for in interpreting and classifying sensory inputs. Thus, adaptive neural nets detecting perceptual clues and rule-based computations performing interpretative reasoning can fruitfully cooperate in this area.
Realising multifunctioning. By ‘reversing’ the neural computations applied in classification tasks, one may obtain instances of the classes that neural nets were trained to classify. In this way the information can be exploited to provide nonsymbolic forms of explanation in expert systems.
Multiplicity of application tasks. The hardware implementation of parallel processing models of propositional rule systems can make the difference when expert systems are required to make very fast decisions.

Hybrid system examples

http://dspace.mit.edu/bitstream/handle/1721.1/47088/coverstoryautoma00schm.pdf?sequence=1

Expert Systems. The technology of knowledge management and decision making for 21st century. Volume 5. Cornelius T. Leondes