Logo

Multivariate Inputs on a MIMO Neuro-Fuzzy structure with LMA training. A studycase: Indonesian Banking Stock Market

Pasila, Felix and Santoso, Murtiyanto and Lim, Resmana (2014) Multivariate Inputs on a MIMO Neuro-Fuzzy structure with LMA training. A studycase: Indonesian Banking Stock Market. Australian Journal of Basic and Applied Sciences (4). pp. 476-481. ISSN 1991-8178

[img]
Preview
PDF - Published Version
Download (314Kb) | Preview

    Abstract

    The paper describes the design and implementation of the multivariate inputs of multi-input-multi-output neuro-fuzzy with Levenberg-Marquardt algorithm training (MIMO neuro-fuzzy with accelerated LMA) to forecast stock market of Indonesian Banking. The accelerated LMA is efficient in the sense that it can bring the performance index of the network, such as the root mean squared error (RMSE), down to the desired error goal, more efficiently than the standard Levenberg-Marquardt algorithm. The MIMO neuro-fuzzy method is a hybrid intelligent system which combines the human-like reasoning style of fuzzy systems with the learning ability of neural nets. The main advantages of a MIMO neuro-fuzzy system are: it interprets IF-THEN rules from input-output relations and focuses on accuracy of the output network and offers efficient time consumption for on-line computation.The proposed architectures of this paper are a MIMO-neuro-fuzzy structure with multivariate input such as fundamental quantities as inputs network (High, Low, Open and Close) and a MIMO-neuro-fuzzy structure with other multivariate inputs, which is a combination inputs between two fundamental quantities (High and Low) and two inputs from technical indicator Exponential Moving Average (EMA High and EMA Low). Both proposed learning procedures, which are using accelerated LMA with optimal training parameters with at least one million iterations with different 16 membership functions, employ 12% of the input-output correspondences from the known input-output dataset. For experimental database, both structures are trained using the seven-year period (training data from 2 Oct 2006 to 28 Sept 2012) and tested using two-weeks period of the stock price index (prediction data from 1 Oct 2012 to 16 Oct 2012) and the proposed models are evaluated with a performance indicator, root mean squared error (RMSE) for mid-term forecasting application. The simulation results show that the MIMO-neuro-fuzzy structure with combination of fundamental quantities and technical indicators has better performance (RMSE) for two-weeks forecast.

    Item Type: Article
    Subjects: T Technology > TK Electrical engineering. Electronics Nuclear engineering
    Divisions: Faculty of Industrial Technology > Electrical Engineering Department
    Depositing User: Felix Pasila
    Date Deposited: 05 Aug 2014 09:51
    Last Modified: 05 Aug 2014 09:51
    URI: https://repository.petra.ac.id/id/eprint/15998

    Actions (login required)

    View Item