Long-term forecasting in financial stock market using accelerated LMA on neuro-fuzzy structure and additional fuzzy C-Means clustering for optimizing the GMFs

Felix Pasila, Sautma Ronni, Thiang, Lie Handra Wijaya
2008 2008 IEEE International Joint Conference on Neural Networks (IEEE World Congress on Computational Intelligence)  
The paper describes the combination of two modeling strategies between the accelerated Levenberg-Marquardt algorithm (accelerated LMA) on neuro-fuzzy approach and fuzzy clustering algorithm C-Means that can be used to forecast financial stock market such as Jakarta Stock Indices (JCI) using the Takagi-Sugeno (TS) type multi-input single-output (MISO) neuro-fuzzy network efficiently. The accelerated LMA algorithm is efficient in the common sense that it can bring the performance index of the
more » ... ork, such as the root mean squared error (RMSE), down to the desired error goal much faster than the simple Levenberg-Marquardt algorithm (LMA). The C-Means fuzzy clustering algorithm allows the selection of initial parameters of fuzzy membership functions, e.g. mean and variance parameters of Gaussian membership functions of neuro-fuzzy networks, which are otherwise selected randomly. The initial parameters of fuzzy membership functions, which result in low Sum Squared Error (SSE) value with given training data of neuro-fuzzy network, are further fine tuned during the network training. As a final point, the above training algorithm is tested on TS type MISO neuro-fuzzy structure for long-term forecasting application of Stock Market in Indonesia. Keywords: TS model MISO-NF network, fuzzy C-Means clustering algorithm, accelerated Levenberg-Marquardt algorithm, financial stock market. Neuro-fuzzy Systems Selection for modeling and forecasting A neuro-fuzzy network with an improved training algorithm for MIMO case was developed by Palit and Popovic [3, 4] and used for forecasting of electrical load data. However, the same network for MISO case was developed by Palit and Babuška [2] . Compared to ANFIS by Jang [1], the neuro-fuzzy models of [2, 3, 4] have achieved better model accuracy and faster training performance. In addition, the neuro-fuzzy model of [3, 4, 5] is an upgraded version of Takagi-Sugeno type multiple input single output NF network. Feedforward type MIMO-NF network is proposed by Palit and Popovic [4, 3] , as shown in Fig. 1 . For our stock market forecasting application we will also use the same model but with number of outputs equals to one. Neuro-fuzzy model as shown in Fig.
doi:10.1109/ijcnn.2008.4634367 dblp:conf/ijcnn/PasilaRTW08 fatcat:z76x4bavvne2xlsz2wqhvgbvxa