A multi-agent architecture for an integrated system that considers data identification, asset valuation, and risk management has been proposed by researchers at Carnegie Mellon University. The system is called WARREN which refers to the first name of the famous investor Warren Buffet (Sycara, Decker, Pannu, Williamson, & Zeng, 1996). The WARREN system design includes components for collecting large amounts of realtime data, both numeric and textual. The data would be pre-processed and then fed to various asset valuation agents that would, in turn, feed their assessments to a portfolio management agent. The portfolio management agent would interact with clients of WARREN. Systems with various features of WARREN are available from commercial vendors and are developed in-house by large investing companies, but more research is needed on how to develop integrated, AI systems that support investing. Natural language processing systems may include large bodies of domain knowledge and parse free text so as to make inferences about the content of the text. However, such natural language processing systems do not seem as popular in investing applications as much simpler natural language processing techniques. The natural language processing work that has been applied to investing seems to be largely of the sort in which the distribution of word frequencies in a A Category: Accounting and Finance 3 document is used to characterize the document. In this word-frequency way, Thomas (2003) has shown a potential value to processing news stories to help anticipate stock price changes. As one can see cycles in the value of financial assets, one can also see cycles in the frequency of publication of articles on certain topics. In the field of artificial intelligence, one might identify, roughly speaking, three phases as follows (Rada, 2008): 

1. Machine learning, in what was then called perceptron and self-organizing systems research, was popular from 1955 to 1975, 

2. Knowledge-based, multi-agent, or expert systems work was popular from 1975 to 1995, and 

3. Machine learning research, now called neural networks or genetic algorithms research, returned to dominate the AI research scene from 1995 to 2013. 

When AI research has been applied to investing, the AI technique used has tended to be the technique popular at the time. This leaves unaddressed the question of whether investing is more appropriately addressed with one AI technique or another. The recent literature is rich with neural network applications to investing, but a new trend is the combining of knowledge-based techniques with neural network and genetic algorithm techniques. For instance, Tsakonas, Dounias, Doumpos, and Zopounidis (2006) use ‘logic’ neural nets that can be directly understood by people (traditional neural nets are a ‘black box’ to humans). Genetic programming modifies the architecture of the logic neural net by adding or deleting nodes of the network in a way that preserves the meaning of the neural net to people and to the net itself. Bhattacharyya, Pictet, and Zumbach (2002) have added knowledge-rich constraints to the genetic operators in their application for investing in foreign exchange markets.

A promising research direction is to combine the earlier knowledge-based work on financial accounting with the more recent work on machine learning for stock valuation. For instance, neural logic nets could represent some of the causeeffect knowledge from a bankruptcy system and become part of a learning system for predicting stock prices. Some of the bankruptcy variables are readily available online, such as a company’s debt, cash flow, and capital assets. The financial markets are human markets that evolve over time as opportunities to make profits in this zero-sum game depend on the changing strategies of the opponent. Thus, among other things, what is important in the input may change over time. An AI system should be able to evolve its data selection, asset valuation, and portfolio management components. The future direction for AI in investing is to integrate the three major tools of AI (knowledge-based systems, machine learning, and natural language processing) into a system that simultaneously handles the three stages of investing (data collection, asset valuation, and portfolio management). Such systems will interact with humans so that humans can specify their preferences and make difficult decisions, but in some arenas, such as program trading, these sophisticated AI systems could compete with one another.