• Kernel Optimization Techniques for Price Prediction

    Hayato Kijima1, Hideyuki Takada2

    1. Graduate School of Information Science, Toho University, Miyama 2-2-1, Funabashi, 274-8510, Chiba, Japan.
    2. Department of Information Science, Toho University, Miyama 2-2-1, Funabashi, 274-8510, Chiba, Japan.

    Abstract: In this paper, we employ Support Vector Machine to predict future directions of the Nikkei 225 futures by learning from the dynamics of Limit Order Book. In order to improve its accuracy, as our previous paper Kijima and Takada (2017) reported, we apply the method of conformal transform of the kernel function pioneered by Amari and Wu (1999). For comparison we also apply Fisher Criteria based data-dependent kernel optimization method proposed by Xiong, Samy and Ahmad (2005) to evaluate their performance. In this sense the paper is a companion to Kijima and Takada (2017) and we conclude, by comparing empirical results, that the conformal transform of Amari and Wu with ex-ante calibrated model parameters improved the precision more than 3.5% in average compared to the standard Gaussian kernel, while the method of Xiong, Samy and Ahmad improved only 1.5% in average.

    Keywords: Limit Order Book, Support Vector Machine, Conformal transformation, Empirical feature space.

    Pages: 319 – 333 | Full PDF Paper