Author: Yip, Kwong-hung.
Title: A neuro-fuzzy approach for stock forecast
Degree: M.Sc.
Year: 2006
Subject: Hong Kong Polytechnic University -- Dissertations
Stock exchanges -- Forecasting
Fuzzy logic
Department: Department of Computing
Pages: 111 leaves : ill. ; 30 cm. + 1 computer optical disc
Language: English
Abstract: The conventional stock forecasting research emphases on predicting the stock price or determining the price trend in near future, the mainstream of this topic in academic concerns in the prediction capability of the mathematical model rather than the practical usage, there are only a few studies to investigate the productivity when applying the model on daily security trading, and this feature should be interested by investor in their own perspective. This research studied and implemented an agent prototype, which can advise the neutral or long position of a security that should be held at next day by analyzing the related information available at parent. The buy and sell transactions are executed whenever two consecutive positions have been switched, and profit can be obtained if positions switch at the right time and the right price. Instead of the predication accuracy, the wealth growth objective function was defined for evaluating the agent performance. The objective function measures the reward obtained by the agent, it also considers the transaction fee that is an important and has a significant impact on the investment return. In the experiment, the fuzzy-classifier proposed by Shigeo and the back-propagation neural network (BNN) were employed and compared their performance on forecasting by running a paper trade simulation, fuzzy-classifier is a hybrid learning approach that combines clustering, fuzzy logic and feedforward network that able to train automatically. Since the agent output is the advised position of security at next day, which implies that the training data had also to be labeled with positions, and base on these positions, the agent should be able to carry out a profitable reward at least in the training period. With the given security price series, the positions for training were generated automatically that in order to maximize the objective function, and a customized genetic algorithm (GA) was implemented for solving the optimization problem for generating the training data. The experiment sample was the security price series of HSBC Holding Limited, the corresponding technical indicator and other data in different categories were collected for supervised learning. In order to reduce the number of input dimensions, the RELIEF and contextual merit, and correlation analysis were taken for feature selection and discarded the duplication inputs.
The empirical result of this study indicated that the agent employed fuzzy-classifier was able to produce positive return in the paper trade simulation, and fuzzy-classifier was likely to perform better than BNN in terms of generality. The findings indicated that it employed the unsupervised learning approach for grouping data into clusters according to their similarity, this process was iterated until each cluster only consisted data with unique class label, finally each cluster was assigned with a fuzzy rule for classify data consists in it. The consideration of similarity and the top-down approach for resolving overlapping region between data let fuzzy-classifier took in advance when dealing with complex problem. On the other hand, the BNN advanced in terms of training efficiency and algorithm complexity, however, the random initialization of the connection weights degraded its stability, with the identical training data, different network had been built because, of the difference of initial values. For the data preprocessing, the feature selection and correlation analysis aimed to reduce the number of input dimensions from 408 to 20, the selected relevance variables were able to improve the efficiency and accuracy. Finally, the GA illustrated the capability to label the training data with the optimized security positions, which was an important contribution toward to a fully automated solution. This study was an exercise to apply the complete knowledge discovery cycle to the live problem, the techniques applied to each process in the cycle were primitive having room for improvement. In the aspect of classification problem, the top-down approach of fuzzy-classifier can be further enhanced by assigning different classification algorithms into each cluster instead of using fuzzy rule, the ideal behind is to depend on the data distribution of each cluster, and then selects the most suitable algorithm to classify the data within this region.
Rights: All rights reserved
Access: restricted access

Files in This Item:
File Description SizeFormat 
b20602972.pdfFor All Users (off-campus access for PolyU Staff & Students only)7.63 MBAdobe PDFView/Open


Copyright Undertaking

As a bona fide Library user, I declare that:

  1. I will abide by the rules and legal ordinances governing copyright regarding the use of the Database.
  2. I will use the Database for the purpose of my research or private study only and not for circulation or further reproduction or any other purpose.
  3. I agree to indemnify and hold the University harmless from and against any loss, damage, cost, liability or expenses arising from copyright infringement or unauthorized usage.

By downloading any item(s) listed above, you acknowledge that you have read and understood the copyright undertaking as stated above, and agree to be bound by all of its terms.

Show full item record

Please use this identifier to cite or link to this item: https://theses.lib.polyu.edu.hk/handle/200/5390