Alma Lilia Garcia-Almanza

Simplifying Decision Trees Learned by Genetic Programming

This work is motivated by financial forecasting using genetic programming. This paper presents a method to post-process decision trees. The processing procedure is based on the analysis and evaluation of the components of each tree, followed by pruning. The idea behind this approach is to identify and eliminate rules that cause misclassification. As a result we expect to keep and generate rules that enhance the classification. This method was tested on decision trees generated by a genetic program whose aim was to discover classification rules in financial stock markets. From experimental results we can conclude that our method is able to improve the accuracy and precision of the classification.

Detection of stock price movements using chance discovery and genetic programming

The aim of this work is to detect important movements in financial stock prices that may indicate future opportunities or risks. The occurrence of such movements is scarce, thus this problem falls into the domain of Chance Discovery, a new research area whose objective is to identify rare events that may represent potential opportunities and risks. In this work we propose to capture patterns of the rare instances in different ways in order to increase the probability of identifying similar cases in the future. To generate more variety of solutions we evolve a genetic program, which is an evolutionary technique that is able to create multiple solutions for a single problem. The idea is to mine the knowledge acquired by the evolutionary process to extract and collect different rules that model the positive cases in several and novel ways. Once an important movement in financial markets has been discovered, human interaction is needed to analyze the markets conditions and determine if that movement could be a good opportunity to invest or could be the principle of a bubble or another critical event that represents a risk. Standard decision trees methods capture patterns from training data sets. However, when the chances are scare, some of the patters captured by the best rules may not repeat themselves in unseen cases. In this work we propose Repository Method which comprises multiple rules to form a more reliable classifier in rare cases. To illustrate our approach, it was applied to discover important movements in stock prices. From experimental results we showed that our approach can consistently detect rare cases in extreme imbalanced data sets.

The repository method for chance discovery in financial forecasting

The aim of this work is to forecast future opportunities in financial stock markets, in particular, we focus our attention on situations where positive instances are rare, which falls into the domain of Chance Discovery. Machine learning classifiers extend the past experiences into the future. However the imbalance between positive and negative cases poses a serious challenge to machine learning techniques. Because it favours negative classifications, which has a high chance of being correct due to the nature of the data. Genetic Algorithms have the ability to create multiple solutions for a single problem. To exploit this feature we propose to analyse the decision trees created by Genetic Programming. The objective is to extract and collect different rules that classify the positive cases. It lets model the rare instances in different ways, increasing the possibility of identifying similar cases in the future. To illustrate our approach, it was applied to predict investment opportunities with very high returns. From experiment results we showed that the Repository Method can consistently improve both the recall and the precision.

Evolving Decision Rules to Discover Patterns in Financial Data Sets

A novel approached, called Evolving Comprehensible Rules (ECR), is presented to discover patterns in financial data sets to detect investment opportunities. ECR is designed to classify in extreme imbalanced environments. This is particularly useful in financial forecasting given that very often the number of profitable chances is scarce. The proposed approach offers a range of solutions to suit the investor’s risk guidelines and so, the user could choose the best trade-off between miss-classification and false alarm costs according to the investor’s requirements. The Receiver Operating Characteristics (ROC) curve and the Area Under the ROC (AUC) have been used to measure the performance of ECR. Following from this analysis, the results obtained by our approach have been compared with those one found by standard Genetic Programming (GP), EDDIE-ARB and C.5, which show that our approach can be effectively used in data sets with rare positive instances.

Repository method to suit different investment strategies

This work is motivated by the interest in finding significant movements in financial stock prices. The detection of such movements is important because these could represent good opportunities for invest. However, when the number of profitable opportunities is very small the prediction of these cases is very difficult. In previous works, we have introduced the repository method (RM). The aim of this approach is to classify financial data sets in extreme imbalanced environments. When opportunities are extremely rare, the investor needs a sharper balance between not making mistakes and not missing opportunities. RM offers a range of solutions to suit the risk guidelines of the investor. The aims of this paper are 1) to show that RM can produce a range of solutions to suit the investors preferences and 2) to analyze the impact of the evolutionary process to RMs performance. Three series of experiments were performed, RM was tested using two artificial data sets whose solutions have different level of complexity. Finally RM was tested in a data set from the London stock market. Experimental results show that: 1) RM offers a range of solutions to fit the risk guidelines of the investor and 2) the contribution of the evolutionary process is very valuable to the performance of RM and 3) RM is able to extract predictive rules even from earliest stages of the evolutionary process.

Simulation in Computational Finance and Economics Tools and Emerging Applications

The works presented in this book can be used as an inspiration for economic researchers interested in creating their own computational models in their respective fields.

Bankruptcy Prediction for Banks: An Artificial Intelligence Approach to Improve Understandability

Artificial Intelligence (AI) is a prominent field within Computer Science whose main goal is automatic problem solving. Some of the foundations of this area were established by Alan M. Turing in his two seminal papers about machine intelligence [39] and [40]. Machine Learning (ML) is an important branch within the AI field which currently is on an intensive stage of development due to its wide range of applications. In particular, ML techniques have recently gained recognition in finance, since they are capable to produce useful models. However, the difficulty, and even the impossibility, to interpret these models, has limited the use of ML techniques in some problems where the interpretability is an important issue. Bankruptcy prediction for banks is a task which demands understandability of the solution. Furthermore, the analysis of the features (input variables), to create prediction models, provides better knowledge about the conditions which may trigger bank defaults. The selection of meaningful features before executing the learning process is beneficial since it reduces the dimensionality of the data by decreasing the size of the hypothesis space. As a result, a compact representation is obtained which is easier to interpret. The main contributions of this work are: first, the use of the evolutionary technique called Multi-Population Evolving Decision Rules MP-EDR to determine the relevance of some features from Federal Deposit Insurance Corporation (FDIC) data to predict bank bankruptcy. The second contribution is the representation of the features’ relevance by means of a network which has been built by using the rules and conditions produced by MP-EDR. Such representation is useful to disentangle the relationships between features in the model, this representation is aided by metrics which are used to measure the relevance of such features.