This paper takes two approaches to prediction of financial markets using text data downloaded from web bulletin boards. The first uses maximum entropy text classification to predict based on the whole body of text; the second uses a genetic algorithm to learn simple rules based solely on numerical data of trading volume, number of messages posted per day and total number of words posted per day. While both approaches produce positive excess returns in some cases, it is found that integrating the two predictors together produces far superior results. Furthermore, aggregating multiple GA trials to build single predictors increases performance even more. 1 Introduction There has long been a strong interest in applying computational intelligence to financial data. Traditionally such attempts have been concerned with forecasting the future based on past price data. However, recently a new source of data has become available: text. Not only have large amounts of financial tex...

This paper presents a number of experiments in which GP-evolved technical trading rules outperform a buy-and-hold strategy on the S&P500, even taking into account transaction costs. Several methodology changes from previous work are discussed and tested. These include a complexity-penalizing factor, a fitness function that considers consistency of performance, and coevolution of a separate buy and sell rule.

This paper presents two methods for increasing comprehensibility in technical trading rules produced by Genetic Programming. For this application domain adding a complexity penalizing factor to the objective fitness function also avoids overfitting the training data. Using pre-computed derived technical indicators, although it biases the search, can express complexity while retaining comprehensibility.

This paper describes how cooperative coevolufion can be used for GP of technical trading rules. A number of different methods of choosing collaborators for fitness evaluation are investigated. Several of the methods outperformed, at a statistically significant level, a buy-and-hold strategy for the S&P500 on the testing period from 1990-2002, even taking into account transaction costs.

In stock market or other financial market systems, the technical trading rules are used widely to generate buy and sell alert signals. In each rule, there are many parameters. The users often want to get the best signal serious from the in-sample sets, (Here, the best means they can get the most profit, return or Sharpe Ratio, etc), but the best one will not be the best in the out-of-sample sets. Sometimes, it does not work any more. In this paper, the authors set the parameters a sub-range value instead of a single value. In the sub-range, every value will give a better prediction in the out-of-sample sets. The improved result is robust and has a better performance in experience.

AI has long been applied to the problem of predicting financial markets. Recently, developments in both AI and financial economics have opened up the possibility for close collaboration between the two fields. First, a line of economics research has emerged that uses AI market forecasting as a form of applied econometrics. Just as importantly, an entirely new source of financially relevant data has become available and amenable to computational analysis: text. Access to text data -- and the associated AI techniques for analyzing it -- not only hold out the hope of improved prediction on the AI side, but also enable financial economics to ask new kinds of questions about how markets react to events. I propose a line of research that develops a set of AI tools, specifically adapted to financial markets, to exploit this convergence for both economics and AI. This research has three main elements. The first thread takes representations from technical analysis -- a semi-rigorou...

The use of machine learning techniques to automatically analyse data for information is becoming increasingly widespread. In this paper we examine the use of Genetic Programming and a Genetic Algorithm to pre-process data before it is classified by an external classifier. Genetic Programming is combined with a Genetic Algorithm to construct and select new features from those available in the data, a potentially significant process for data mining since it gives consideration to hidden relationships between features. We then examine techniques to improve the human readability of these new features and extract more information about the domain. Categories and Subject Descriptors I.2.2 [Artificial Intelligence]: Automatic Programming –

Multiple Objective Genetic Programming (MOGP) is a promising stock-picking technique for fund managers, because the Pareto front approximates the risk/reward Efficient Frontier and simplifies the choice of investment model for a given client’s attitude to risk. Unfortunately GP solutions don’t work well if used in an environment that is different from the training environment, and the financial markets are notoriously unstable, often lurching from one market context to another (e.g. “bull” to “bear”). This turns out to be a hard problem — simple dynamic adaptation methods are insufficient and robust behaviour of solutions becomes extremely important. In this paper we provide the first known empirical results on the robustness of MOGP solutions in an unseen environment consisting of real-world financial data. We focus on two well-known mechanisms to determine which leads to the more robust solutions: Mating Restriction, and Diversity Preservation. We introduce novel metrics for Pareto front robustness, and a novel variation on Mating Restriction, both based on phenotypic cluster analysis.