Non-Conformity Detection in High-Dimensional Time Series of Stock Market Data
It is desired to extract important information on-line from high-dimensional time series because of difficulty in reflecting the data fully in decision making. In econometric techniques, previous works primarily focus on prediction of price. To make decis
- PDF / 711,847 Bytes
- 12 Pages / 439.37 x 666.142 pts Page_size
- 67 Downloads / 142 Views
2
Department of System Innovation, The University of Tokyo, Hongo 7-3-1, Bunkyo-ku, Tokyo, Japan [email protected], [email protected] Daiwa Securities Co. Ltd., GranTokyo North Tower, 9-1, Marunouchi 1-chome, Chiyoda-ku, Tokyo, Japan {takaaki.yoshino,shunichi.ashida}@daiwa.co.jp
Abstract. It is desired to extract important information on-line from high-dimensional time series because of difficulty in reflecting the data fully in decision making. In econometric techniques, previous works primarily focus on prediction of price. To make decision in business practice, it is important to focus on human-machine interaction based on chance discovery. Here we propose Non-Conformity Detection as a method for aiding to humans to discover chances. Non-Conformity Detection is designed to detect a noteworthy point that behaves exceptionally compared to surrounding points in time series. In the experiment, the method of Non-Conformity Detection is applied to the time series of 29 stocks return in the electrical machine industry. As the result, four dates among the detected top five non-conformity points coincide with the important dates that professional analysts judged for making investment decision. These results suggest Non-Conformity Detection support the discovery of chances for decision making. Keywords: Non-Conformity Detection · Chance discovery · Highdimension · Time series · Stock return · Decision making · On-line
1
Introduction and Motivation of Research
In recent years, it has been enabled to obtain a large quantity and highdimensional data, which here means data including several dozens or hundreds variables for each time, called data-rich environment. It is desired to extract important information from high-dimensional time series because it is difficult to reflect the data fully in decision making. Dynamic factor model has been proposed for predicting future price from high-dimensional data on economic time series. In the recent trends on machine learning, deep learning is applied to prediction of stock price. However, the prediction accuracy hardly exceeds 60 %. For the time, if a stock investor seeks to obtain much confidence in investment, one cannot completely rely on automated reasoning by these algorithms in business practice. Therefore, the interaction between data mining system and human c Springer International Publishing Switzerland 2016 H. Fujita et al. (Eds.): IEA/AIE 2016, LNAI 9799, pp. 701–712, 2016. DOI: 10.1007/978-3-319-42007-3 60
702
A. Kasuga et al.
intelligence is important than relying on only automated quantitative analysis or machine learning in order to make decision. In this paper, we propose Non-Conformity Detection in high-dimensional time series as a method for aiding to humans to discovery chances through human-machine interaction. We define a non-conformity point as a data point that behaves exceptionally compared to surrounding data points in time series. The reason we focus on non-conformity is based on chance discovery. An event as a non-conformity point tends
Data Loading...
